22 Apr 2026 · 30m 19s
Ai In Architecture & Engineering: Ai Leadership Forum | Part 1
The video explores the integration of AI in architecture and engineering, highlighting the roles of companies like Chaos and SWECCO in utilizing AI technologies for visualization, internal processes, and software development. It emphasizes the importance of trustworthy AI application, transparency, client engagement, and the challenges faced by companies in adapting to AI's rapid advancements while improving productivity and transforming industry standards.
AI in architecture and engineeringAI integration in workflowsTrust and transparency in AI implementationAI integration into software developmentImpact of AI on industry processesHuman-AI collaboration in engineering and architecture
ChaosSWECCOSweco
APSRevitFusion
Claude CodGPTChatGPTCopilotAI agentsNeural AI
Key Frames (6)
A speaker is presenting at the Autodesk DevCon AI Leadership Forum, with a slide displaying the event name and collaborators.
Autodesk, Arcadis, AWS, Bosch, Choo, Microsoft, Sweco
The slide introduces Jeremy Kahn as the host, with a title mentioning his role as AI Editor at Fortune Magazine, as part of the Autodesk DevCon event.
AUTODESK
A slide from Autodesk DevCon about the acceleration of AI capabilities, showing a chart detailing software task durations relative to LLM release dates.
AUTODESK
A slide shows a chart comparing theoretical AI coverage of job tasks across various occupational categories, indicating AI exposure is highest in knowledge work.
AUTODESK, FORTUNE
This slide features a panel titled 'From automation to AI agency' with images and titles of three speakers, discussing agentic AI.
AUTODESK
A presentation slide at Autodesk DevCon featuring speakers discussing AI agency.
AUTODESK DevCon
►
Full Transcript (274 segments)
[0:00:00]Okay, awesome. You guys can hear.
[0:00:04]Other than my initial sort of soldering and assembly cord,
[0:00:09]we have all been dabbling with AI.
[0:00:12]How many people actually spend all your weekend hours dabbling with AI?
[0:00:17]Of some sort.
[0:00:19]As sad, isn't it?
[0:00:20]I thought AI is going to fix all the problems and I'm going to have more weekend time
[0:00:23]but it's not looking like that.
[0:00:26]A lot of us are trying to figure out this technology.
[0:00:30]And, you know, we haven't changed as humans and leaders.
[0:00:35]I think one of the key things that I look at my days is really obsessing as to how
[0:00:41]this technology is going to evolve our lives and what makes sense and what doesn't make sense.
[0:00:46]I think that is the fundamental shift that we're going to go through and it's in a way
[0:00:51]liberating because we are kind of like the pioneers who can figure this out.
[0:00:57]If it is not us, who else?
[0:00:59]Especially the next generation is looking at us saying, you know what?
[0:01:01]Figure this out, people.
[0:01:03]So we know what to do.
[0:01:05]So as we're going through one of the key things at Autodesk for us,
[0:01:10]what's most important as we figure this out is one thing I'll tell you is for sure
[0:01:15]is we're going to make mistakes.
[0:01:17]That is for sure.
[0:01:19]Because we are all learning and we are moving and we want to do this together.
[0:01:23]So you can be guaranteed if you're trying hard, we're going to make mistakes.
[0:01:28]And when we do that, I know this community is going to tell us exactly when we're doing
[0:01:32]something wrong or when we're not doing right.
[0:01:35]And I'm going to rely on that and we're all going to rely on that and why we are being
[0:01:39]transparent and talking about including you in our workflows and all of that stuff.
[0:01:45]I'm sure something's going to go wrong.
[0:01:47]But guess what?
[0:01:48]If you don't try, we're never going to learn.
[0:01:51]That is the most important sort of grounding principle for us.
[0:01:55]Another thing for Autodesk for us as technologists, there's a lot of great technologists that work
[0:02:00]at Autodesk.
[0:02:01]You know, they're sitting right in front of here, but many who are at home, right, in other
[0:02:06]areas.
[0:02:07]One of the key things we worry about, we wake up and say, where do we apply this technology?
[0:02:13]We don't start off as, what is this cool technology?
[0:02:15]Let's go play with it.
[0:02:16]We don't start there.
[0:02:18]We go, where is this most important and impactful in our industries?
[0:02:23]Let's start there.
[0:02:25]The second thing we look at is, what is real here and what's not?
[0:02:30]I don't know.
[0:02:31]You know, there was A to A and MCPs and all of that stuff that went through and we went
[0:02:35]through that whole sort of journey for the past year.
[0:02:39]There's still some things that are evolving, right?
[0:02:42]We were actually among the very few people working with the MCP Linux Foundation to ensure
[0:02:48]that it's actually securely done.
[0:02:50]That protocol is really securely.
[0:02:52]We care about these things.
[0:02:53]Our industry cares, our customers care about security, permissioning, all of these things
[0:02:58]that cannot be sacrificed in this technology.
[0:03:03]And the other bigger thing for us is not messing up.
[0:03:06]Who heard Shelley talk about from, you know, probabilistic answers to deterministic answer?
[0:03:13]You saw BCA here.
[0:03:16]There's no way in Singapore he's going to be okay if he give him a different answer
[0:03:20]the next morning.
[0:03:22]That's not going to work, really.
[0:03:24]Even for this 4,000 deterministic answers, rules that he had,
[0:03:28]there's 2,000 that still are not deterministic and we cannot screw that out.
[0:03:32]So these are the most important grounding principles for us is to ensure we take risk
[0:03:37]but we need to do that in a very sort of calculated and make sure that it actually works.
[0:03:43]Some of you have asked me this question, you guys are not making enough noise about AI.
[0:03:48]You're not really talking about all the stuff that you need to put out there.
[0:03:52]It's because we want to get this right.
[0:03:54]One of the key things, even when we rolled out assistant in the past few weeks,
[0:03:58]we wanted to make sure we had running all our e-vow frame work,
[0:04:00]making sure that it actually comes through right.
[0:04:03]We're monitoring all of our social media and all of that channel just to ensure we didn't mess this up.
[0:04:09]So that is important and I'm looking to this community to make sure you hold us to the bar
[0:04:16]that we hold ourselves.
[0:04:20]The second thing for us, most important is making sure that AI actually shows up in the workflows.
[0:04:28]We have realized it's not just about the task automation.
[0:04:31]It is that workflow automation that I talked about.
[0:04:33]That is the most important thing because we might live in our products.
[0:04:39]You live in workflows.
[0:04:41]And when you live in workflows, we need to show up in those workflows and make sure that workflows automated.
[0:04:46]So that is a key focus for us and you'll see that in many of the examples that we will give here
[0:04:51]or even in any other places.
[0:04:53]You'll see us play out that kind of workflow automation.
[0:04:57]So there's quite a bit.
[0:04:59]I want you to assure you that as technologists, we want to make mistakes.
[0:05:03]We want to try. We want to not stop trying.
[0:05:05]But at the same time, one of the key things we want is ensure that it is applicable for our industries
[0:05:11]and applicable for our customers. That's most important to us.
[0:05:15]That, I'm going to turn this over to Jeremy Khan from our AI editor from Fortune,
[0:05:22]right here with us and he's got some amazing interesting content around this.
[0:05:27]Can you, Jeremy?
[0:05:29]Thank you.
[0:05:30]Awesome. Thank you so much.
[0:05:33]Thank you, Rajee.
[0:05:35]I'm Jeremy Khan. I'm the AI editor at Fortune.
[0:05:37]I'm also the author of a book, Mastering AI, a survival guide to our superpowered future
[0:05:42]that came out about a year ago if you want to pick up a copy of that.
[0:05:47]I'm going to give a quick overview of where we are in the AI landscape before we get to our first panel.
[0:05:53]This sort of organizes the what, the why and the how of AI.
[0:05:59]So where are we right now with AI and AI development?
[0:06:02]AI capabilities are coming along at a very rapid clip.
[0:06:07]And you may have be familiar with this particular chart behind me.
[0:06:12]This is from Meter, METR, which is one of the leading AI kind of evaluation,
[0:06:18]independent evaluators of frontier AI models.
[0:06:21]And what this shows is they test how long AI can work autonomously to certain standards.
[0:06:29]This is on software development tasks.
[0:06:32]This is how long can an AI model work autonomously with a 50% success rate
[0:06:37]and we're currently up to over 10 hours that the leading models like Claude Cod from Anthropic
[0:06:47]using Opus 4.6, their latest model, that can work for now.
[0:06:51]Over 10 hours continuously with a 50% chance of success on these software development tasks.
[0:06:56]And what this chart has shown is that these AI capabilities have been doubling about every seven months since 2019.
[0:07:04]And there's some indication that actually in the last two years they may be moving at an even faster pace
[0:07:09]that we're at a kind of an inflection point where capabilities may be doubling every three months.
[0:07:16]And I should say also on that last chart, if you, meter also looks at how long it takes to develop for software
[0:07:24]to AI models to complete these same software development tasks with an 80% success rate,
[0:07:30]there we're at just over an hour of autonomous work that the leading models can do with an 80% success rate.
[0:07:37]Despite the model capabilities coming along very rapidly, we still live in a world of what Google CEO,
[0:07:44]Sundar Pachai, has dubbed Jagged Intelligence.
[0:07:47]And I think an interesting way to think about this Jaggedness is that we have models that excel in some areas,
[0:07:52]coding and software development being one of them.
[0:07:55]But in other areas they really struggle.
[0:07:57]And actually in an area that we're going to talk a lot about here today,
[0:08:01]and which is very relevant for Autodesk and all of its customers, is visual reasoning and visual understanding.
[0:08:07]Is an area where AI models are still not as capable as people.
[0:08:12]The Jaggedness is actually reflected in how you see AI's impact across different professional areas.
[0:08:19]The chart behind me now is also from Anthropic, and they tried to look at what tasks in what professional areas could AI today theoretically perform,
[0:08:30]and then also looking at based on their own usage statistics, where do they actually see people using AI to perform tasks.
[0:08:38]And what really comes across in this is that sort of jaggedness of that profile.
[0:08:42]So again, in certain tasks, management, business and finance, legal, you see the blue area, which is the theoretical coverage,
[0:08:49]is quite deep, close to 100% theoretically AI models today could do a lot of those tasks.
[0:08:56]And then the red area is actually what's actually being done.
[0:08:59]And you see even in the areas where AI is potentially very capable, we're still just sort of scratching the surface of what people are actually using the models for.
[0:09:08]And there's often good reason for that.
[0:09:09]This is another way of looking at the same chart.
[0:09:12]A lot of people like that radar chart from Anthropic, but a bunch of people said it's really hard to see what's going on, and doing it as a bar chart is better, so this is as a bar chart.
[0:09:20]And you'll see that when it comes to architecture and engineering, for instance, the theoretical coverage by AI that Anthropic calculated was 82%.
[0:09:30]But if you look at the red here, the actual observed activity in architecture and engineering, it's just about 18% of tasks right now that they're seeing people actually using AI to do.
[0:09:42]But again, these capabilities are coming along very quickly.
[0:09:46]And to talk about this more, moving from automation to AI agency, I want to introduce to my panelists.
[0:09:51]We've got Vladimir Nikolov, the head of innovation at Chaos, Henry Vethel, who's the Chief Digital Officer from SWECCO in the Netherlands, and Ben Cochran, the VP of Developer Enablement at Autodesk.
[0:10:06]So let's welcome them onto the stage, and they're going to join me in conversation for our first panel.
[0:10:21]Yeah, take a take a seat at the end, Ben.
[0:10:28]Great. Why don't we start with you? Why don't you talk a little bit about where you're using AI right now, sort of within within what Chaos is doing, explain a little bit to people about what Chaos does generally for those who may not be familiar.
[0:10:40]And then talk a little bit about where you're using AI right now.
[0:10:44]So Chaos is, as a company, has been around for more than 25 years.
[0:10:49]We are creating software for visualization, simulation, design companion pieces.
[0:10:56]We're mostly famous for three pieces of software, NSKP, which is our design companion software for architects.
[0:11:04]We also have VRA, which is an offline render engine that is used for architecture visualization, but also a lot of movie effects, basically, like all the 10 nominations for Oscars.
[0:11:17]For visual effects this year, used VRA in some form or shape.
[0:11:21]And then we have Corona, which is a render that is specifically focused on architectural visualization.
[0:11:28]For us, visualization at the moment is the biggest chunk of what we do, and AI has a big impact on how people approach visualization.
[0:11:40]And up until this point, you needed a render to create an image.
[0:11:45]So you needed a 3D model. You needed to apply materials, you needed to apply lights and all that until you get an image that you can show to your clients.
[0:11:53]Today you can just type a text prompt or you can draw a sketch and AI will do the work for you.
[0:11:58]So it's another way, completely another way to create images. You can also generate videos out of that.
[0:12:04]So for us, that's an interesting opportunity. You still need a physically based engine like VRA and Corona and NSKP.
[0:12:14]If you want to stay true to the original 3D model, and you want to show exactly what you designed.
[0:12:20]But in the early stages, when you are just trying to figure out what the project might look like, if you just want some inspiration or some mood boards to show to a potential customer, this is where AI can be really helpful.
[0:12:31]Because you can just from a few sketches, you can create something that you can tell you if you are going in the right direction or not.
[0:12:38]So in this stage, AI is very, very helpful. The other area is the end stage where you already have a final image, but you want to do some final touches, add some details, make it a little bit more realistic.
[0:12:50]Especially people and vegetation are really, really difficult to do realistically with 3D rendering, but AI can help a lot in that regard.
[0:12:58]So at the beginning of the process, when you are doing initial creative visualization, it's very good for that.
[0:13:04]And maybe at the end of the process, to add some touch ups to an image, but the middle bit where you actually have to get physical reality to match up with the image and where you need some engineering knowledge.
[0:13:19]That's where you are still not really seeing some models capable enough.
[0:13:24]Exactly, that's a little bit more difficult. And what we see is the current models, LLM, LLM's in VLM's visual language models, don't really have a very good understanding of spatial relations between objects.
[0:13:36]So if you say I want this object on top of another object, or I want you to move this so far, like 5 meters to the left, to the right, they're not really very good at this yet.
[0:13:47]Alright, Henry, so I swekko, why don't you talk a little bit about people maybe very familiar with you, but just say a little bit about what you guys do and then where you're using AI in your process.
[0:13:57]So I work for swekko, swekko is the largest engineering company in Europe located in the northwest part of Europe.
[0:14:05]And we use it almost in every part of our company. It was in 2022 that Chatshipiti became big.
[0:14:13]What I'm really proud of was that we as a company created our own swekko GPD just a few months after this launch of Chatshipiti.
[0:14:20]And then our journey started. And now it's being used in every part of our company. We have for instance an HR chatbot.
[0:14:28]So people want to in the past ask questions to the HR manager. Now they have their own chatbot where they can ask questions.
[0:14:35]So what I see at the moment, it's really spreading through our whole company to give you some statistics.
[0:14:40]We find out that 95% of our colleagues uses swekko GPD and 60% uses this on a daily basis.
[0:14:50]So that's really impressive numbers. And it's also been our strategy from the start.
[0:14:57]If everybody uses it in our company and makes a very small step every day as a company we make huge steps.
[0:15:03]So that's our philosophy.
[0:15:05]Yeah. So lots of small productivity gains add up to big productivity gain. That's interesting.
[0:15:09]And I know when we were talking in preparation for this session, you said though that swekko ultimately your engineering important objects, bridges, roads, key buildings,
[0:15:19]they have to stand up, they have to work. And a big problem, you know people thought with AI is can you trust it?
[0:15:25]And ultimately, you know your business is all about trust. All about things that have to work and be trustworthy in the real world.
[0:15:31]How have you managed to sort of marry AI into a business which is so dependent on trust?
[0:15:37]First of all, I think it's very important to engage your clients. So that you engage your clients in collaborating on AI.
[0:15:45]And also to make it traceable that if you create an answer that you always can find the source of truth of this answer.
[0:15:52]So what we have in swekko GPD is that if you have an answer from swekko GPD, you can always find the source of your information.
[0:16:01]Right. Yeah, so that grounding with truth. Ben, I want to bring you in here. Where are you seeing AI being deployed within Autodesk itself?
[0:16:10]And how are you thinking about moving AI enabled features into what Autodesk can offer?
[0:16:18]Yeah, so I have a outside of APS in the developer community. I also look after developer Ambulent inside of Autodesk.
[0:16:25]And so I'm actually responsible for how we deployed AI tools in our development process as we build software.
[0:16:31]And I've noticed there is the usage, I sort of break the usage into three chapters or maybe three acts of a play.
[0:16:39]The first one really is using the tools themselves. So you get a cursor or a quadcopilot.
[0:16:44]And what you're seeing here when we build software is you can see like a maybe five to 30% improvement on productivity gains.
[0:16:52]And it's all aspects of building software. It's just the agents then help you or the system helps you build software.
[0:16:57]But then there's what we see next is in the second act of the second part is what you can do if you actually transform or change the way you work.
[0:17:06]Where you actually build teams of agents and you get, you know, some people will say spectrum and design.
[0:17:12]Or you actually have this entire agentic system that changes the way we build software.
[0:17:17]You challenge the process, you challenge all the roles essentially.
[0:17:21]And that's where you start to see things like you see 5X, 20X, 30X improvements where you see things that once took months happening in an afternoon.
[0:17:30]And that is a transformational way of approaching building software where you have agents that are looking out for not just the production of the features but the quality itself.
[0:17:40]So you're building teams that understand how to do the right the validation.
[0:17:45]What's important, it's not just getting the code out there, it's the right code.
[0:17:48]And then the third act of what happens is the impact of that.
[0:17:51]It challenges the processes for how we build software.
[0:17:54]And I think the whole industry is faced with this right now.
[0:17:57]We don't have the answers yet. But I don't think, you know, I don't think agile in our current form holds up.
[0:18:03]I don't think, you know, some of the scrum ceremonies and the processes and the two-week experience.
[0:18:07]I don't, I think they're going to need to be revisited.
[0:18:10]And we're doing it, we're, right now we're looking at that.
[0:18:12]We're looking at, what does it mean for teams, you know, disruption teams or catalyst teams as we call them.
[0:18:18]What does it mean when they work in this way and how do we experiment to learn what the new processes will be?
[0:18:24]And then we're also staying connected to, in the industry, to help answer that question.
[0:18:28]Whatever this post agile way of working is as we both thought.
[0:18:31]Are you changing the way you kind of configure your teams because of the use of AI?
[0:18:35]And are you looking for sort of different qualities and capabilities from the people that are, you know, are going to be assisted by AI in this new process?
[0:18:44]We don't know what, we don't actually know what those are, but we know that we need to.
[0:18:49]And like a great example would be, you know, look, let's say you're going to add a new feature.
[0:18:55]In the past, you did a usability study and you did it with Mox and you, you, you, you, you engage with, you know, a customer to get data and information back.
[0:19:02]Well, that takes a lot of time to build up what that possibility is.
[0:19:06]Why not just build it in the product? If it takes you in afternoon, go out with, you know, a beta test and actually do it in product itself and get that validation.
[0:19:14]It, you no longer took months to build the thing. And so now you can actually put that into an afternoon, get three outcomes, do the experiments, do the learning before you actually put that into the main product.
[0:19:25]And once you put it in the product, you actually have the things already been done. It's already being created.
[0:19:29]And so it just, it, it's not that the things that we did are wrong, but the order of them change.
[0:19:35]And you need people who understand that way of working. And through this transition, understanding the thing that we don't, you know, you can't go to a textbook.
[0:19:43]Colleges aren't teaching this new way of working because the industry hasn't even solved for it yet.
[0:19:47]And so it's an interesting transition, which I think actually creates opportunity for people who may be more entrepreneurial in their approach.
[0:19:54]But I think it's also something we have to help people through because they, there's, there's fear there as well.
[0:19:59]Henry and, and Vlad, I want to bring you in here too. How are you dealing with this issue of, you know, you have these productivity savings.
[0:20:06]Are you potentially reconfiguring the way teams work within like a thin swekko? Are you seeing a reconfiguration of roles?
[0:20:12]Thanks to AI and thanks to some of maybe the productivity savings.
[0:20:16]Exactly what you said before. I see in our software development teams, we have software development teams within swekko.
[0:20:22]I really see the old fashioned way of stream teams disappear and vibe coming really be the new thing happening.
[0:20:31]But what I really think is important in this world of agentic AI and fast decision making is that there's always a human between us and AI.
[0:20:41]And it's also one of the pillars of our AI strategy never put any AI generated contact directly to a client always have the human checks and balances.
[0:20:52]Vlad, how about you? Since you are able to use AI potentially in that first stage and maybe for some of those later touch ups that you used to take humans time in front of a computer to do that.
[0:21:02]What do you redeploying people to do now?
[0:21:05]So first I wanted to continue something that Ben mentioned for us internally so in terms of our products.
[0:21:13]AI is there but we also use the internet to develop our products and it's also another interesting topic and it completely changed the way that we approach our roadmap and how we develop software.
[0:21:26]AI allows us to basically any developer can go into any code base and start doing work there.
[0:21:32]They don't need to be exactly familiar or they don't need to be an expert in a specific language. They don't need to be an expert in a specific code base.
[0:21:41]This allows us to basically break the silos of our teams so instead of having teams working on one product we can take people from multiple.
[0:21:48]Other teams and give them a task to work on for the next two or three months.
[0:21:53]Even if this is something that is not in their usual role and we found that this actually works pretty well because with AI you don't need like the language that you use in the programming language becomes a lot less important than it was originally.
[0:22:09]Besides we already have a huge code base so the AI can figure out based on the context what it should be doing most of the time.
[0:22:16]So if you want to add a new feature chances are that the AI will be able to figure out how to do this and then the human just needs to give it a direction and then review the result.
[0:22:26]Ben I wanted to talk a bit about something that Vlad mentioned that AI right now is not so good at actually doing the very precision engineering steps.
[0:22:37]Of course the vision here is maybe to go directly from artistic rendering to blueprint.
[0:22:43]How close are we to sort of realizing that vision and I know Autodesk is doing some work on trying to train models that might take us towards that.
[0:22:50]Where are we in that process?
[0:22:52]Yeah well it's a journey right and when you know you if you watch some of the talks at Tina you got to see some of the investments we're making and we call it neural AI and some investments we're making and some models.
[0:23:06]Right now when you look and you see what the models are doing they're produced the end result the pixels that you're looking at is the end result.
[0:23:14]The system in the model doesn't understand what's behind that so if you create something even the rendering sometimes you know when you're doing these statistical renderings it doesn't know that that model behind it is a 3D object or it doesn't know that when it creates a symbolic 2D represent what that actually means that that wall actually is a wall that can accept a door that you know can accept a window sitting on it.
[0:23:35]The window is sitting on a foundation with the roof and that's really what we're trying to use those models but also train models as well and even partner with some hyper scolars to build things and understand what you're actually looking at.
[0:23:50]And it's it's it's almost you know if you know I think a number of us have been around for a while we took this journey from 2D into 3D and when we talked about 3D it wasn't just a 3 dimensional thing it was an understanding of what that actual object was not just what you see on the screen and the models don't understand that yet and so that's where it breaks down when you look at those you know what what's what you're looking at visually and we're trying to knock over to conquer some of those those obstacles to further the model so they do.
[0:24:19]Understand that that context and don't lose the understanding that we're trying to the thing we're trying to build and you need that for the quality if you don't have that the you know that's where where the doesn't understand really what the intent is right.
[0:24:33]One of the other trends we're seeing a lot of is people starting to experiment with a gentick AI systems systems that can use other software what's what do you think about an auto desk about the use of agents maybe buy your customers and the idea that maybe it's not going to be a human using the object.
[0:24:48]Not going to be a human using the software in the in the near future it's starting ready to be maybe an AI agent that's performing certain functions on auto desk using the software and how is how you viewing that and trying to sort of accommodate that reality but also ensure that there's the trust and the safety that you need.
[0:25:06]Yeah well and and some of you hopefully you took away from this morning right as you start to work with agents what happens is that those agents have access to tools and those tools are the same tools that we have trust in today and we can provide
[0:25:17]validation and so if the agent is you know driving rev it or driving fusion you know all of a sudden now you you have this thing that you can validate you can still run the same simulation analysis you can still do the same
[0:25:30]VIM validation and and that data is there it's very much true that if you're just looking you know at the end result of a picture like what that's not you know what we're building is far more than that and so what you'll see was as these agents are there it's it's you know auto desk and and this community that builds those capabilities and and bridges us between where we are now to to where we're where we're going to go which is you know when you speculate in the future there's always questions but but we can't lose the ability to do that.
[0:25:59]But we can't lose the core tools that that provided that validation right.
[0:26:04]Henry I wanted to talk a little bit to you about you said one of the issues that may happen as we start to get more of these productivity enhancements is more and more people get used to the idea of AI agents AI models doing things is client expectations change and when we were speaking earlier you said one of your concerns is that people start to expect everything will be done at AI speed even things that really cannot be done at AI speed and should not be done.
[0:26:28]AI speed how are those conversations that are going with customers and convincing them that actually know this process does require human does require expertise and will take several days to or weeks to complete.
[0:26:41]To be honest this morning when I was in this conference I was really impressed about how fast this industry is going forward.
[0:26:50]I see the two biggest challenge is to keep our colleagues into the same pace but also our clients so we have a lot of discussions with clients about the adoption of AI of course what I believe in is make it in small understandable steps and therefore they can adapt to AI right.
[0:27:12]Yeah absolutely how about you what are you what are you using in terms of expectations from customers.
[0:27:18]For us it depends on the other job but when customers can go and generate an image in the matter of seconds based on a prompt that's kind of their expectation now so obviously real time solution is especially if you can produce results as quickly as possible that's what customers want but also the same ease of use that you get with with an image.
[0:27:47]With an image generator they kind of started to expect this from 3d software so basically type of prompt and things magically happen this is kind of the expectation we're not there yet but we're trying to get as close as possible to that place where the tools that we have today so we're building building all kinds of AI assistants.
[0:28:07]So we're trying to use the tools that we're trying to get in the way for customers to use a chatbot to control the software super interesting things that we couldn't do even just a few months ago.
[0:28:17]So you also have a point about training and what AI does to the whole training pipeline for the industry traditionally you would bring in junior engineers they would know certain things from university but they'd really learn in first several years on the job from the tasks that you would give them but some of those tasks are now the things that AI is good at and can do what are you thinking about a sweat go in terms of the training pipeline and also you need to train these people to use the AI tools and where to use them appropriately and where not to how are you addressing that issue.
[0:28:45]I think that's one of the biggest challenges of introducing AI in a company like this back up but I think in all companies and I had a discussion with somebody who just graduated from university and he told me something which really made me think was because what is the truth.
[0:29:02]If you're experienced if you're working in this industry for 20 30 years you know what the truth is if you see a design you can directly see if something something makes sense or calculation but if you're just graduated from university.
[0:29:15]What's the truth even studying it's already becoming more complex in my time in the 90s the truth was written down in the book in the age of the internet it already became a little bit fake but in the time of the AI generated content it becomes even more fake so I think for the younger generation what's the truth.
[0:29:35]Coming more experience I think that's a big challenge and we don't have an answer yet but we know that we have to think about it yeah and universities are behind that that I think that's a core problem is that universities need to catch up in terms of what is required from from the next workforce yeah no absolutely that's great we're just out of time I want to thank Ben and Vlad and Henry for being with us give them all round of applause and we'll bring on the next panel.
[0:30:06]Thank you so much.
[0:30:08]Thanks so much.
[0:30:10]Thank you.
22 Apr 2026 · 33m 0s
Ai Agents In The Enterprise: The Business Case For Going Beyond The Poc
The video discusses the integration of AI agents in enterprise settings, emphasizing the business case for moving beyond the proof of concept stage and focusing on data strategies, ethical AI practices, and trusted infrastructures. It highlights examples from companies like Bosch, Novo Nordus, and Autodesk, showcasing uses of AI for automation, design, operational efficiencies, and time-saving in clinical trials, while encouraging exploration despite potential risks to achieve reliable business outcomes.
AI agents in enterpriseAutomation of business processesPartnerships in AI implementationAI integration in business processesEnterprise AI strategies and challengesData strategy and architecture in AIIntegration of AI agents in enterprise systemsCommand and control in building automationAdvancements in software development practices with AI
Novo NordusBoschMicrosoft
Autodesk agentsFusion
AI agentsCopilotIoT
Demo Segments
- 0:00:00 Bosch's process for automating fire alarm system proposals using agentic systems.
- 0:00:00 Transform 2D plans into 3D models using Autodesk agents for system design.
- 0:00:00 Design a table in Fusion using AI tools
Key Frames (5)
The slide presents data comparing AI performance against industrial and mechanical engineers, highlighting win rates and ties in specialized tasks.
AUTODESK
A presentation slide discussing the 'Rise of Agentic AI' and introducing 'OpenClaw'.
Autodesk, OpenClaw
A slide presentation discussing the Agentic Economy and AI with reference to a Fortune article.
AUTODESK
A slide showing 'The AI Opportunity' session at Autodesk DevCon, featuring speakers from Bosch, Microsoft, and Autodesk.
►
Full Transcript (340 segments)
[0:00:00]Welcome back. I hope you were able to get a coffee during that break. I'm Jeremy Khan,
[0:00:11]Fortune's AI editor again, and I'm going to be moderating this next session. This is kind
[0:00:16]of in three parts. So the last one was on sort of the what of AI. This is on the why and the why
[0:00:22]now, kind of the business case for AI agents. But I'm just going to give a quick overview of where we
[0:00:27]are. I'm going to try to give some different data than I did in the last little intro. This is
[0:00:32]some data from OpenAI's evaluation called GDP Val. And if you haven't looked at it, it's really
[0:00:38]worth looking at because they had professional experts in a whole range of knowledge work,
[0:00:45]develop realistic kind of scenarios and tests for to try AI models, frontier AI models against.
[0:00:52]And it ranges from sort of things in banking, things in tax, things in audit, government services,
[0:01:00]and some in engineering tasks as well. And this is some of the data from industrial engineering and
[0:01:07]mechanical engineering from the last assessment they did. And this showed and they do a whole bunch of
[0:01:12]they don't just test their own models, they test other frontier models as well. Currently it looks
[0:01:17]like most of the best models are just over sort of 50% wins and ties against human performance.
[0:01:25]Again, this is judged by human experts. And when it comes to mechanical engineering, the models
[0:01:31]actually tie human performance or exceed human performance 60% of the time. So they're coming along
[0:01:38]really quite quite significantly. And it's worth keeping an eye on that benchmark. Of course,
[0:01:43]one of the other things we've seen in the last few months has been the real rapid adoption and
[0:01:48]thusiasm for AI agents, in particularly these what are called agentic harnesses, this kind of software
[0:01:55]code that sits around an AI model and tells it kind of what tools it's allowed to use and what
[0:02:00]scenarios provide some guard railing. How many here are familiar with OpenClaw or have played around
[0:02:06]with OpenClaw? Quite a few, that's good. Yeah, there's been huge enthusiasm for OpenClaw as a
[0:02:12]agentic harness. And a lot of people now are trying, but the problem with OpenClaw is actually
[0:02:17]it was fairly unsafe. You kind of had to give it control over everything on a laptop and there were
[0:02:22]lots of nightmare scenarios where OpenClaw had gone out and done things that the users hadn't really
[0:02:26]wanted, including in some cases giving up bank account information or the keys to Bitcoin wallets,
[0:02:33]things like that that you don't really want to have happened. So a bunch of people now are working
[0:02:37]on how do you bring an OpenClaw like experience to enterprises in a way that's safe for enterprises.
[0:02:43]This is still very much kind of a work in progress, but it is something to keep an eye on.
[0:02:47]In terms of what's actually happening in companies, there's a lot you may have heard, there was this
[0:02:51]MIT study that came out earlier this year saying that a lot of companies were failing to achieve
[0:02:58]return on investment from their use of AI, there was a statistic that 95% of AI pilots fail.
[0:03:06]I think we're kind of moving beyond that. I think that study is already fairly outdated.
[0:03:10]And you are seeing some really big enterprises seeing really significant productivity gains
[0:03:17]through the use of these agentic AI processes. This is a case study from Novo Nordus, the big pharmaceutical
[0:03:22]company that worked with Anthropic with their cloud model. They were able to take a process that
[0:03:28]used to take more than 10 weeks that had to do with compiling all of this data from clinical trials
[0:03:34]for these clinical trial reports that they then have to present actually to regulators.
[0:03:37]They took that down from 10 weeks to 10 minutes. So that's a huge savings that we're seeing.
[0:03:43]We're also starting to see the rise of these sort of solo entrepreneurs who are using AI to create
[0:03:50]huge companies with massive revenue with almost no employees. So this is a guy who runs a kind of
[0:03:57]medical supplement marketing company The New York Times wrote up. He's got $1.8 billion
[0:04:03]of users and annual revenue. He has two employees. So I mean these are the sorts of things that people
[0:04:08]are starting to see is possible with AI agents. And I think there's a big, a lot of big companies
[0:04:13]are saying how do we bring a little bit of that inside our own enterprise? What can we possibly do
[0:04:18]inside this much larger company that already may have billions of revenue to extend to new business
[0:04:24]opportunities? We're going to talk a little bit about this opportunity with my next set of
[0:04:28]panelists. Please welcome Andreas Mauer, the chief architect from Bosch of building solutions.
[0:04:35]Thomas Titsa, I hope you haven't mascured all your last names. The technical technology strategist
[0:04:40]at Microsoft and Nevin who is the VP of platform strategy and technology partnership at Autodesk.
[0:04:47]Please welcome. I'm not even going to try Nevin's last name. What is...
[0:04:51]Yuxsel Ikiji. Yuxsel Ikiji. So I should have gotten that's great. Thank you. Welcome. Please.
[0:04:58]Great. Andreas, I want to start with you because you have this fantastic vision for what you think
[0:05:07]is going to be possible at Bosch building solutions with the advent of AI and AI agents. Can you just
[0:05:14]sketch for us a little bit of what that vision looks like? Because I think it's tremendously exciting.
[0:05:18]Yeah, thanks. So at Bosch building technologies we creating commercial building solutions systems
[0:05:25]and services like you see here with light and HVAC in those commercial buildings and the fan
[0:05:31]fact is that these buildings are not really standardized in terms of how you operate them
[0:05:36]and even how do you plan dimension and design them so that you can bring them into those buildings
[0:05:42]and this is where we are now leveraging more and more the power of a generic system and trying
[0:05:46]to get that system design and planning and dimensioning so that we can create offers to customers
[0:05:53]using the power of a generic systems. That's one of the pieces. Then once we have that system
[0:06:00]designed and planned and offered to customers then we also bring them into the building. So we do
[0:06:05]provisioning installation configuring and this is also where we use the generic systems behind the
[0:06:11]scenes to get assisted configurations going, assisted installation for the engineers and the
[0:06:19]craftments that install all of this in the buildings. Then next we bring it into operation and then
[0:06:25]we operate, we maintain it and then sometimes Bosch building getting refurnished, restructured,
[0:06:32]sometimes even recycled and that is when the loop closes and starts from the beginning.
[0:06:38]All of those processes life cycle faces you need to have a high domain expertise and those people
[0:06:46]getting less and less. So we have a shortage of them and all of those affairs are not really
[0:06:53]highly standardized and this is also where our partnership with all the desk comes in and also
[0:06:58]the partnership with Microsoft. We're trying to civilize the wild west of commercial buildings as
[0:07:04]I always say and we have a big milestone today as well. We have a joint white paper composed,
[0:07:11]you find it in the flyer, we call this the open building service reference. This is trying to
[0:07:16]articulate, you can download the flyer as a QR code, this is really trying to articulate our dilemma
[0:07:23]that we have in those commercial buildings because some of you may know 30% of the CO2 footprint
[0:07:29]are caused by those buildings and the reason is they are not standardized, proactively operated,
[0:07:35]designed, constructed and this is where we use the whole range of everything we heard so far
[0:07:42]and there's even the new phrase called agent tile which is the combination of a tentic plus HIL.
[0:07:49]So because Ben was just deluding on what is the next kind of HIL look like, it's called agent tile
[0:07:55]and we are in the midst of this. Wow, that's a, no it's an amazing vision. One of the things even on
[0:08:00]that first piece on the system design you were talking about is that right now if you come to
[0:08:07]Bosch and you want a request for proposal for a fire alarm system for a building, it takes some
[0:08:14]time for Bosch to come up with that, it was a people intensive process and that limited the number
[0:08:20]of those that you could do in a given period of time and that itself limited the market opportunities
[0:08:25]that you could go after. You could only do a building of a certain size with a certain valuation.
[0:08:30]Now by automating that you can not only sort of take on more projects but also potentially serve
[0:08:35]market segments that you just couldn't serve before. Is that that's part of the vision? Maybe
[0:08:39]you can talk a little about what how that works. Yeah, absolutely. So we in this white paper you will
[0:08:44]also find something that we call the system plan our co-pilot. I give you some numbers. So in Germany
[0:08:49]alone we're getting per week 100 tender requests from customers for fire alarm systems and to
[0:08:56]plan such fire alarm systems depends of the complexity depends up on the side size. So it's
[0:09:02]a difference if you have a soccer stadium or an airport or a railway station or a building like
[0:09:07]this and so usually we cannot serve all of those 100 tender requests. So we lose business.
[0:09:13]And so by automating that now with a tentic system so we're getting all of those tender requests
[0:09:18]specification documents in including usually 2D plans. Never or seldomly do you get 3D
[0:09:25]build models. It's always 2D something. It's PDF, it's chain pack, it's maybe DVG, all of those
[0:09:32]funny formats. Sometimes it's even scanned in paper maps. And what we're now doing we really
[0:09:37]are transforming that in a whole pipeline of what people used to do in the process. So we take the
[0:09:42]information models we receive. We do the scoring. Is that business relevant for us? Yes, no. If
[0:09:47]that light goes green, we do the bill of material based up on product catalogs that comes from
[0:09:53]integrated enterprise master data management systems. Is a set of agents that are educated for
[0:10:00]that given task attend. And the next thing is what we're doing now stunningly is we're transforming
[0:10:05]to 2D maps we receive into 3D using the auto desk agents on the other side. So we're having an agent
[0:10:12]to agent integration. We create a 3D building model and then we design and project a fire alarm
[0:10:19]system in real time into that 3D model. And that is not just stunning and cool. It is also
[0:10:25]changing our industry. Because we're getting away from that wild west, what somebody thought
[0:10:30]should be planned and how it's being installed later is now in this fully life cycle chain,
[0:10:36]highly standardized from agents assisted and guaranteed to be safe. And that of course gives us
[0:10:42]higher business throughput. It gives us our standardizations and therefore we also make our
[0:10:47]business far more efficient. And we just start with a fire alarm domain. We have many other
[0:10:52]domains like HPEC, like interruption detection, like access control, video surveillance, energy
[0:10:58]management obviously which is a very important one. And so this is one of the examples,
[0:11:02]quite exciting times in industry. Yeah, never. I want to bring you in. When we were preparing for
[0:11:07]this panel you mentioned that a lot of companies when they first start thinking about using AI or
[0:11:12]using AI agents, they kind of look at their existing processes and they just try to make them a
[0:11:16]little bit faster, a little bit more productive. And you said that the danger is that you kind of end
[0:11:20]up with a lot of faster horses when you really want is a rocket ship. Can you talk a little bit about
[0:11:26]about that and how you avoid maybe that trap of falling into that sort of faster horses mentality?
[0:11:32]Yeah, and I think that's exactly right. I think when you start with, okay, these are the processes
[0:11:37]and the people and the outputs we have right now. And how do we apply AI to this? Like in our
[0:11:44]software development process, for example, can we use AI to automate documentation? We can,
[0:11:52]but that's getting you a faster horse. If you rethink the, okay, what outcome are we actually
[0:11:57]trying to get to? And how do we rethink that with these capabilities? Ben talked about that in
[0:12:01]the previous panel in terms of totally rethinking the process. Now what used to take the longest
[0:12:08]time was the coding, right, and software development. And that is, you know, you can now automate that
[0:12:14]with agents, but that changes the whole dynamic because now you could automate some
[0:12:20]pretty terrible mic outcomes, not really get to the outcome you want if you don't have the
[0:12:25]specifications very, very detailed. So now actually the bottleneck shifts, for example, to
[0:12:32]describing exactly the outcome you want. And then you can skip a few steps with agents. And then
[0:12:38]now testing also becomes very important, right? So the whole mindset needs to change from the top
[0:12:44]down. And it can't just be like a few, you know, excited engineers, like they on their own,
[0:12:49]they will not make such a big shift at the whole company. So rethinking it really requires a
[0:12:55]mindset shift at even leadership levels. Absolutely. Thomas, I want to bring you in here. What,
[0:13:00]you know, when you see customers maybe get stuck at that POC stage, you know, what is usually
[0:13:07]the issue there? Why do some of these projects fail to scale? And when you have seen ones that
[0:13:12]really have been successful and or really ambitious, like what Bosch is trying to do, what, what
[0:13:17]do they need to have done to sort of lay the foundation for that? And we see this very often,
[0:13:22]to be honest, that you start a lot of POCs, people build their own chat pilots and so on. And
[0:13:30]what then is missing really the transformation to the business? So what does it mean for my business
[0:13:34]process? I have to probably rethink my business process. What people will do in the future in this
[0:13:40]process. And then I have to think about, okay, what is the platform to host this? Do I have a
[0:13:45]scalable platform? If I use this, not in just one process, maybe in 100 process, if the
[0:13:50]chat pilot is not used by 10 people, by 10,000 people, or in the case of Bosch, 100,000 people maybe,
[0:13:57]can I scale? Do I have the security and so on? And we have to rethink also traditional IT
[0:14:02]processes. Can we adapt this processes for this new world? And can we bring this capabilities
[0:14:08]then also to all employees, not just the engineers who are used to use this? What happens if we bring
[0:14:14]to administrators to sales guys and so on? And how can we train them to use these tools? But we
[0:14:22]often see you provide tools for like a salesman and they are not used to use it. This training
[0:14:28]missing and so on. So you also have to invest in trainings and explain these things to your users.
[0:14:36]Andreas, how are you dealing with that at Bosch in terms of making sure people know how to use
[0:14:40]these tools effectively? Yeah, I mean, it always sounds like when I hear those C people talking
[0:14:46]about the 10X, 100X kind of savings, there is a real hard practice behind that you have to
[0:14:53]understand. To get as Thomas said in early POC, that is fairly fast done. But to bring this into
[0:15:01]a materialized highly automated engineering practice, you need to establish TeneiOps. TeneiOps
[0:15:07]is the same as this as DevOps does, but for the Tenei practice. And that ensures that you have
[0:15:13]end-to-end observability, which is very important. So you want to know who agent talks to other
[0:15:18]agents, what models are used, how many tokens did they consume, which prompts the day use,
[0:15:23]and what was the response in the collaborative interaction between agents. So you need to have
[0:15:29]this entire observability from the human operating with the digital agents all the way to the results
[0:15:36]and outcomes. And at Bosch, of course, we need to have this all and I guess for enterprise
[0:15:41]corporations as well, we need to have this compliant. So you need to adhere to ethical AI,
[0:15:47]responsible AI, transparent AI, ethical AI, right? Then you have funds like the FunFacts in Europe,
[0:15:54]like the EU AI Act that you have to adhere to. So to bring this all in practice, really, really,
[0:16:01]you have to take this seriously, an easy POC, that's, yeah, it's wipe, as they say. But to bring
[0:16:07]this materialized into an industry practice takes really effort and skills. So yes, you have now
[0:16:15]digital beings that do standardized and help you to get higher business process efficiency.
[0:16:21]But in order to get there, you first have to invest. And there's also a fun fact behind that,
[0:16:26]without data strategies and data architectures, you can forget it. You better go and have a nice
[0:16:32]dinner with your wife instead. Because if you don't have data strategies and data architecture,
[0:16:37]then the AI is pretty much powerless, I would say. So you have to respect all of this and
[0:16:44]you have to invest for this and tell this to your C people because it takes, it costs money.
[0:16:49]Right. Yeah, and I want to pile on there with the building that trusts its infrastructure at the
[0:16:55]beginning. We've really invested a lot there to avoid kind of a wild west approach because every
[0:17:01]product team could have like gone and developed something. But it creates a lot of tech debt.
[0:17:07]And also how do you ensure the compliance, security, reliability that we commit to for our
[0:17:13]customers because we're operating industries where that matters. And so we don't want to skimp on
[0:17:19]that. So what we did was invest a lot in that trusted foundation first, really providing that
[0:17:25]visibility to our customers in terms of AI transparency cards for each feature,
[0:17:32]building a common shared infrastructures to make sure that each of our AI capabilities are built
[0:17:40]in a secure, governed, compliant way. And that actually really, that upfront investment
[0:17:45]really helped us speed up later because now each of the, like any, any team that wants to build an
[0:17:50]AI feature as long as they use that shared infrastructure, they can just go and they don't have to
[0:17:55]worry about all that stuff. It comes built in as long as you built it in that central place. And
[0:17:59]that really helps us speed up then. So it feels like you go slow, but you go slow to go fast.
[0:18:07]Go slow to go fast. That's interesting. Yeah, and there's one other hidden secret test
[0:18:11]ability because if you have non-ditaministic systems, there's no such thing like a unit test.
[0:18:17]So you need to establish domain specific evaluation pipelines. And these evaluation pipelines actually
[0:18:23]they need to reason against your domain expertise that you put in those digital kind of shapes.
[0:18:31]And so you better have also in your data strategy semantics included, meaning you have ontologies
[0:18:38]knowledge graphs in your universe because if the agenda systems don't have something to
[0:18:43]crown against and reasoning against domain specifically, then you also will not get to the best
[0:18:49]potential outcomes. Yeah, no, you really emphasize this point that that's essential kind of these
[0:18:53]knowledge graphs and having that domain specific knowledge embedded in a way that the system can
[0:18:58]access and reason against not and not just assuming that it sort of natively has this when it's
[0:19:04]pre-trained that it's going to know all these professional standards areas, you know, even in your
[0:19:10]own company, what's the meaning of a particular document or you know, what is the particular procedure
[0:19:14]that Bosch uses as opposed to a different company, right? And that's really a key. Absolutely,
[0:19:19]absolutely. And that is something you have to include in your data strategy. And thankfully,
[0:19:24]with our partner Microsoft, we use a lot of Microsoft Foundry goodness that gives you all of that
[0:19:30]capabilities as a platform, as a service because if you want to build this all by your own,
[0:19:34]how to use the evaluation pipelines to stock evaluation pipelines like you want to check if the
[0:19:40]outcome is any, any wild end or any, you know, bad things coming back. So you want to check this
[0:19:48]and you want to have this end-to-end traceability based on open telemetry standards. And then you
[0:19:53]want to have the models that you fingertips. You also want to have model routers so that you have a
[0:19:58]model that itself decides which model to go and get to question at hand, best potentially ask in
[0:20:03]terms of token consumption, throughput latency and all of that kind of thing. Yeah, well, that's a good
[0:20:07]point there about token consumption because someone actually at the break was saying, well, the big
[0:20:12]question I didn't ask in the last panel was about around cost. And, you know, how do you control
[0:20:17]the cost of these systems? And then also given that they do have a cost, how do you also convince the
[0:20:22]customer that the benefit they're getting from the AI system is worth potentially paying a premium
[0:20:28]for? Sometimes they said the problem is that the customer wants things done at AI speed, but does
[0:20:33]not want to pay for that at all. You know, it's, I think it's free. So how are you looking at that
[0:20:38]at Bosch? Both the cost containment on the production side, but also then what do you charge for these
[0:20:44]AI solutions to the customer? Oh, the business model story is also business architecture you have to
[0:20:49]have. And to be honest, we don't have all the answers yet because we're learning the entire
[0:20:55]industry is learning. I mean, if you ask 100, see people, how many agenda systems do you have
[0:20:59]lively deployed and make business with? Good luck. I mean, there is a lot of money burned right
[0:21:06]now still right? So we are on the exploration phase. We also trying to learn with Microsoft Thomas.
[0:21:11]Thomas, you want to get in here? That's actually one of the big challenges we have at the moment.
[0:21:15]So what we good now is I think we have this cover ability, how many agents are used, how often
[0:21:20]these agents are used, but the missing is still linked to the business KPIs. So what actually
[0:21:27]is the effect of this agent and is the business itself clear? What do we want to measure? So in
[0:21:33]this case, for instance, we want to provide more tenders per week. Is this measured? Where is this
[0:21:39]data measured? And how can we bring this data together? This is one of the challenges we see at
[0:21:42]the moment. There is a lot of good ideas, but does it really fit to business value? And this is
[0:21:48]where we need to get better, but this is not just a technology problem. It's also business problem
[0:21:54]that you think about what are the critical KPIs and how can I improve this KPIs with technology?
[0:22:00]Right, and for the managers then with Internet enterprises trying to figure out how do you manage
[0:22:05]human employees alongside AI agents to maximize these KPIs? I mean, the business KPIs
[0:22:12]presumably don't really change, but the question is how are you using a mix of people and agents to
[0:22:16]reach those KPIs? Digital employee thing is really growing now if you have agents which
[0:22:23]we see a lot of these chatbots. And I think we need to move away from these chatbots. They need to run
[0:22:28]by itself. So we have automatic agents. Then we have digital employees. So we have to control
[0:22:34]these agents, these employees. We have to monitor them. We have maybe performance discussions with them
[0:22:39]like normal employees. And it's important for instance, also for the HR department, how does this fit
[0:22:47]with my real employees and digital employees? Is the real employee becoming an agent boss?
[0:22:54]Manager of agents, what does this mean also for his performance itself? And so a lot of interesting
[0:23:01]questions coming up. Baby don't have an answer yet in all cases. The performance review is actually
[0:23:07]very interesting. We were literally just talking about it the other day with one of my colleagues
[0:23:12]who was using a lot of agents in his work. And he said I asked the agents to come up with a
[0:23:18]performance review for themselves. And then log that back into the master file so that they would
[0:23:23]then improve on those skills. So it's very, very relevant. Yeah, and we also talked about trust
[0:23:28]rights. How much do you trust such a genetic system about their outcome? So in terms of making
[0:23:35]a tender request to a commercial binding offer with the dimensioning and design and system planning
[0:23:40]included, which by the way, we manually when we do this, we have to do this as a pre-invest
[0:23:46]upfront. Even before we know that the customer is then ordering that system from us. So this is easy
[0:23:52]then, right? So there you can have an easy business model and you can see the value of what the
[0:23:57]token costs you and the entire gen I ops pipeline and the practice you need to establish versus what
[0:24:02]you get out of that. But there are other scenarios when we are on operations, for instance, in a
[0:24:06]building like that to do proactive heating ventilation and air conditioning. Usually these systems,
[0:24:12]they run in the basement, caved into cabinets and nobody cares about them until some system is
[0:24:19]no longer working. What we are now establishing is actually we have by the means of IoT goodness,
[0:24:24]you know, that was the hype cycle before the AI hype cycle. Anybody may be remembered that.
[0:24:29]IoT is still out there. So the connectivity and the crownedness and the semantic meaning of now,
[0:24:37]the systems continuously being observed by a tentic system in the background knowing exactly
[0:24:42]when energy consumption is not optimal any longer because there is an air handler unit filter that
[0:24:47]is sturdy that causes the fact that there is no air flow and air quality going to happen in that room
[0:24:53]causes additional energy consumption. You can now proactively in the background, order spare parts,
[0:25:00]planned service technician execution to happen in the building, notify the customer. This is no
[0:25:05]science fiction, no rocket science people. This is all the power of domain expertise put into standardized
[0:25:11]digital gold. And this is also where you can make a very easy reasoning about business model and
[0:25:17]innovation and impact that you have to the industry and society at all. Because remember,
[0:25:21]Borschtes invented for life. Yeah, absolutely. Again, vision is so compelling.
[0:25:27]Nevin, you had said that when we were preparing for this that really you need CCC sort of sponsorship
[0:25:34]and you kind of need AI native thinkers high up within companies to really drive some of the return
[0:25:39]on investment from this. Can you talk a little bit more about why you think that's so important?
[0:25:43]Yeah, I mean, at least understanding at the C level or at the decision making levels, I think
[0:25:48]is really critical. And if you know, unless they are listening to their AI native employees,
[0:25:56]they may not even realize what is possible, then they're going after the faster horse,
[0:26:01]but you're trying to get to the rocket ship. And so, you know, as long as there are influential
[0:26:06]people, you know, what happened at Autodesk, for example, we have a lot of very AI native VP level,
[0:26:14]you know, people. And so, they were all tinkering. As soon as, for example, the MCP protocol came out
[0:26:20]end of 2024, I believe. And then early 2025, we were already like showing examples of, hey, I
[0:26:27]built this infusion and make it made a little video about it like, hey, design a table for me.
[0:26:31]Well, make it cherry wood. Well, that like seems to be attached at the top. Let's put it to the
[0:26:36]bottom. You know, those types of conversations were, and then they were doing little videos and
[0:26:40]literally like, hey, my 10-year-old is now designing an infusion. 10-year-olds on video. It's like
[0:26:46]design a fidget spinner. Wow, that fidget make it blue. Like make those parts, and we're like
[0:26:52]watching the video, right? So, and then that we immediately took that to C stuff, obviously,
[0:26:56]like, hey, this is possible now. And that kicked off a whole flurry of activity. But, you know,
[0:27:02]unless C stuff can actually realize that, you know, you have maybe a little bit less direct impact
[0:27:09]potentially. Right. If again, if there isn't a mindset shift at the company. Right. Interesting.
[0:27:15]At Bosch, how did you decide where to start? I mean, you have this vision of, you know, so much
[0:27:20]can be done with AI agents. You can really offer a complete suite of automated building services.
[0:27:26]Where did you, but you're starting up with the fire system design? Why did you pick that? And how,
[0:27:31]what was that process like? Deciding, you know, because I think some people struggle with that,
[0:27:34]where do we start? What is, what's an imp, a valuable thing we can do, demonstrate impact,
[0:27:41]but that's achievable, because sometimes the vision is, it seems too big or you want to do something,
[0:27:46]it's very valuable, but it seems too hard to achieve. How did you pick the fire system designs?
[0:27:51]So, why we started with the fire alarm system domain was because we have the highest customer
[0:27:56]impact with that in Germany, obviously, because the fire alarm domain is highly regulated.
[0:28:00]So, and that gives us obviously the highest business throughput, followed by building automation
[0:28:06]and energy, because there's now regulations about energy monitoring and, you know, things like
[0:28:11]ESG reporting and stuff like that. So that's how, in the order we selected, that's the first domain,
[0:28:17]but for sure, we will all pick the other domains, because standardization gives you then also
[0:28:22]higher business efficiency visibility, which you usually don't have. So there's not that single
[0:28:27]pane of class that they see level guy can look at and see how good their bet is your business
[0:28:32]efficiency, and this is also something that we transport with that tool surface. Fantastic. Thomas,
[0:28:37]you have view across a lot of different standard Microsoft customers. Where are you seeing on this
[0:28:41]issue of sort of training, how you bring your people along, so they're using these tools effectively,
[0:28:47]and also how your, you know, how companies are thinking about continuing to train sort of the
[0:28:50]next generation of, of engineers and architects to use these tools and not maybe lose some of those,
[0:28:56]those key skills. What are you seeing across the industry? What do you see that people are
[0:29:03]use these tools at home, you have it on your phone, you use jetchypt everywhere or whatever tool,
[0:29:10]and what we need is to translate this in that daily work. How do you really use these tools
[0:29:17]effectively in your work? But we have very good successes within a day and a life of a salesman,
[0:29:24]an engineer, an administrator, whatever, and try to adapt the tools. How do you use AI for this
[0:29:32]real activities you have day by day? This gives them a better understanding what they can expect
[0:29:38]from the AI. There's also a lot of over-promising to be honest. People think I can do now everything
[0:29:44]with it. It's not yet the case, so we need to explain them also. Okay, how can you use AI,
[0:29:51]where you should be careful, how do you adapt it to your daily processes, and how you get better
[0:29:56]in your daily processes if you use these tools? This really helps to adapt it to specific roles.
[0:30:04]Unfortunately, there's not one answer which works for everyone in the company. Your salesman needs
[0:30:09]a little bit different AI than your engineers. Andreas, I'll give you the last word very quickly,
[0:30:15]because we're almost out of time. If you had one sense for folks here, senior leaders,
[0:30:21]and where they should start with thinking about these agentic processes, what would it be?
[0:30:26]Yeah, I mean, identify your most critical business areas that you want to standardize and optimize,
[0:30:32]and also think about the level of autonomy you want to give those agentic systems. Because that is
[0:30:38]something for us, a very critical thing to decide how much autonomy do you let go, and how much do
[0:30:46]you still have the human in the loop? So, the scenario that I was describing with the HBAC system,
[0:30:51]do you still have a human in the loop to order a spare part? Probably not. Nothing can really happen
[0:30:56]other than maybe two parts ordered instead of one. But if you want to actually command and
[0:31:01]control the building automation system in order to make energy optimizations going on,
[0:31:06]maybe there's somebody that has the last word to do, do I want to shut down heating and
[0:31:12]cooling at the same time? You probably should. It's a bad idea in terms of energy consumption,
[0:31:17]but maybe you want to have something between because heating and cooling at the same time makes
[0:31:20]sense for whatever reason. So, that is one. And then the other thing is think about your software
[0:31:26]development practices if you have them, because what we observe now, there's a fun thing called
[0:31:31]so you can hire a super agent and you have to learn what spectrum development means because we
[0:31:39]no longer trying to have abstractions to talk to compute machines. We now express intent.
[0:31:47]And there are machines that will turn the intent around to talk to abstract the things like
[0:31:52]C++, GoRust, C-ShopJab, or you name it. And so spectrum development is key and you hire that
[0:31:59]super agent and then you can actually see the super agent hiring other agents for development,
[0:32:05]testing, creating IAC pipelines, doing deployment to the Microsoft Cloud, doing end-to-end user tests.
[0:32:12]And you can really in real time see how many agents are employed and when they go off the team
[0:32:18]again. So that is something explore, don't be shy and have the courage to fail because you will.
[0:32:25]And make sure you have sea level folks then understand AI is not just the bus word until you make it so.
[0:32:34]So if you really want to live it, you have to respect what it takes to make AI and industry
[0:32:41]reliable business outcome. Fantastic. We're out of time. Thank you so much for listening.
[0:32:46]Thank you to Nevin, Thomas, and Andreas. Thank you. We'll be back soon.
22 Apr 2026 · 61m 33s
Autodesk Devcon 2026 Keynote (Day 1): Powering The Next Era Of Innovation With Agentic Ai
The Autodesk DevCon 2026 Keynote showcased the transformative impact of Agentic AI on software development and the AEC industry, emphasizing AI-driven workflow automation and real-time interoperability with models. Demonstrations highlighted Autodesk Assistant's role in enhancing efficiency in tools like Revit and Fusion, with innovative use cases from partners like Andrids and Deutsche Bahn. The session underscored the future of design and engineering powered by AI, data connectivity, and dynamic orchestration of connected systems to boost productivity and innovation.
Agentic AI in software developmentAutomation and AI-driven workflowsData-driven engineering transformationInteroperability and AI integration in AEC workflowsIntegration of AI in Autodesk productsConnected systems and workflowsOrchestration through Autodesk Assistant
AndridsDeutsche BahnDuriverMirBuilding and Construction Authority (BCA) SingaporeGeneral MotorsNour ConsultVictory AIArquedes
APS (Autodesk Platform Services)ACC (Autodesk Construction Cloud)RevitFusionAutoCADCivil 3DNavisworksGenerative DesignPlatform ServicesAutodesk Platform Services
Agentic AIGenerative DesignAssistantMicrosoft Co-Pilot Studio
Demo Segments
- 0:00:00 Demonstration of Andrids using APS for workflow automation, reducing manual workload significantly.
- 0:00:00 Introduction of Autodesk Assistant and AI integration in design workflows.
- 0:28:06 DX Live streaming interoperability demonstration with model updates in Revit and Rhino.
- 0:47:23 Conceptual demonstration of Autodesk Assistant creating a door schedule in Revit
- 0:48:51 Autodesk Assistant generating visual review-ready images in Fusion
- 0:52:12 Autodesk Assistant in collaboration with Victory AI optimizing HVAC designs
- 0:55:35 Arquedes using Microsoft Co-Pilot Studio agent with Autodesk Assistant for project evaluation
Key Frames (59)
The slide contains a message emphasizing scalable design with the phrase 'Design what scales.'
A person is presenting at Autodesk DevCon with a slide showing a celestial diagram or map.
AUTODESK
A presenter is onstage with a slide showing a solar system animation, possibly demonstrating a feature at Autodesk DevCon.
AUTODESK DevCon
The frame shows a presentation slide featuring images of construction professionals using digital tools, with the Autodesk DevCon logo.
Autodesk DevCon
A slide from an Autodesk DevCon presentation showing themes: Build, Orchestrate, and Scale.
Autodesk
A presentation slide featuring the name and title of the EVP, Chief Technology Officer at Autodesk.
Autodesk
📹 0:06:00
code
The frame shows programming code along with function descriptions related to input characters and hex digit handling.
AUTODESK DevCon
The frame shows a speaker presenting at Autodesk DevCon alongside a slide showing a silhouette of a person in front of computer screens.
Autodesk
The frame shows a presentation slide with the text 'More meaningful work' and images of a speaker and a person working at a computer.
AUTODESK
📹 0:09:00
demo
An aerial view of a large industrial complex with the text about custom solutions and Autodesk branding visible.
Autodesk
📹 0:10:30
diagram
The frame shows a person presenting a technical diagram on a screen to a small group, wearing construction safety attire.
Autodesk
A presentation slide at Autodesk DevCon showing a graph with customer value versus time/maturity including task automation, workflow automation, and systems automation.
AUTODESK
A presentation slide showing a chart about automation with a speaker at Autodesk DevCon.
AUTODESK
A presentation slide at Autodesk DevCon showing different levels of automation from 'Autodesk AI Foundations' to 'Systems Automation.'
Autodesk
A presentation slide showing a person speaking and an image of a robotic hand fist bumping a human hand.
AUTODESK DevCon
A presentation slide showing a close-up image of a hand holding a gold nugget, with a speaker on the left side.
AUTODESK DevCon
A presentation slide shows 'AUTODESK ASSISTANT' with a speaker on the left side.
Autodesk Assistant
A presentation slide with the text 'What will you build?' alongside a speaker on stage.
Autodesk
The slide highlights a message about developers moving up the stack, with a presenter on stage.
AUTODESK
📹 0:20:30
demo
A speaker is presenting at Autodesk DevCon with an image of an old fort on the screen.
Autodesk
A speaker is presenting about 'Structured Granular Data' with keywords 'Grounding', 'Context', and 'Focus' visible on the slide.
AUTODESK
A slide from Autodesk DevCon showcasing the AEC Data Model with upcoming features like Plant 3D, Import & Export IFC, and Civil 3D.
AEC Data Model, Plant 3D, Import & Export IFC, Civil 3D
A presentation slide at Autodesk DevCon discussing interoperability as an industry problem.
Autodesk
A slide from Autodesk DevCon showing a DB train with text 'Engineering and Consulting'.
Autodesk
A slide from Autodesk DevCon featuring Dura Vermeer with an image of a bridge and infrastructure work.
Autodesk
📹 0:28:30
demo
A presentation slide displays a comparison between Revit and Rhino 3D usage in architecture.
Autodesk Revit, Rhino 3D
A presenter stands next to a slide featuring a cityscape and the Autodesk DevCon logo.
AUTODESK
This frame shows a presentation slide with the logo of the Building and Construction Authority at an Autodesk DevCon event.
Autodesk
The frame shows a speaker next to an image of a historical fort or castle by the sea, under the Autodesk DevCon banner.
AUTODESK
A presentation slide from Autodesk DevCon is shown alongside a speaker on stage.
AUTODESK DevCon
The frame shows a presenter speaking at Autodesk DevCon with a slide featuring a person using VR equipment.
AUTODESK DevCon
📹 0:37:00
ui
The frame shows a tablet screen displaying an interface with a prompt 'What's on the agenda today?' and various menu options like 'New Chat' and 'Settings' under the label 'ChatGPT'.
ChatGPT
📹 0:38:30
demo
The frame shows a presentation with a talking head and two women engaging with virtual structures on a tablet, likely discussing software or technology applications.
Autodesk DevCon
A presentation slide at Autodesk DevCon showing topics 'Discovery' and 'Orchestration' with illustrative images.
Autodesk DevCon
A slide from Autodesk DevCon showing a 'Discovery' theme with a colorful abstract design emanating from a laptop.
AUTODESK DevCon
A presentation slide showing various team collaboration and design scenes with the Autodesk DevCon logo.
AUTODESK DevCon
The frame shows a presentation slide with the title 'Orchestration' and a speaker alongside.
AUTODESK DevCon
A presenter is shown next to a slide displaying a complex metallic structure at Autodesk DevCon.
Autodesk
A slide from Autodesk DevCon showing a comparison of '40% Lighter' and '20% Stronger'.
Autodesk
A presentation slide showing images related to Autodesk DevCon, including people working with technology and a speaker on stage.
AUTODESK DevCon
A video frame showing a presentation slide related to a construction project by Norconsult Digital at Autodesk DevCon.
Autodesk DevCon, Norconsult Digital
The slide presents information about over 60 million data points coordinated in real-time, featuring designs and diagrams of infrastructure projects.
Autodesk
The slide showcases a variety of professional scenarios, including construction, office work, and teamwork, under the Autodesk DevCon theme.
Autodesk
The frame shows a slide discussing Discovery, Orchestration, and Distribution with visuals and text.
AUTODESK
A presentation slide from Autodesk DevCon featuring a speaker and an illustration of teamwork with digital effects.
AUTODESK
📹 0:48:00
demo
A speaker is presenting at Autodesk DevCon with a demo of the Autodesk Assistant interface shown on screen.
Autodesk Assistant
📹 0:49:00
demo
The frame shows a speaker presenting at Autodesk DevCon with a building floor plan and a text analysis with door specifications on the screen.
AUTODESK
📹 0:49:30
demo
A person is presenting at Autodesk DevCon, displaying a digital keyboard rendering on the screen along with a chat interface.
AUTODESK
A presentation slide showing '1000s of data points' with a network-like background. The frame is part of an Autodesk DevCon event.
Autodesk
📹 0:54:00
demo
A presentation slide at Autodesk DevCon showing a 3D building model with duct optimization analysis and interface featuring Autodesk Assistant.
Autodesk Assistant, Revit
A slide from Autodesk DevCon showing that AI generates insights and suggests further actions, with improvement opportunities identified.
AUTODESK, Arcadis
A slide from an Autodesk DevCon presentation showing a construction site with the title 'Hours'.
Autodesk
The slide highlights AI integrations from the Autodesk community, showcasing different tools and plugins available.
AUTOM8LABS MCP Connector, NonicaTab PRO. Professional member, Sentinel QC with MCP AI Assistant for Revit
The frame shows a presenter on the left and a collage of images on the right related to architecture, manufacturing, and technology. The event is Autodesk DevCon.
AUTODESK
The frame shows a slide from an Autodesk DevCon presentation featuring images related to construction, architecture, and technology.
AUTODESK
The frame shows a slide from Autodesk DevCon featuring three executives and their roles.
Autodesk
►
Full Transcript (786 segments)
[0:00:00]Yeah.
[0:00:04]Don't leave!
[0:00:44]And now, please welcome to the keynote stage, Autodesk's VP of Developer Enablement, Ben
[0:00:54]Cochran.
[0:00:55]Yes.
[0:00:56]Good morning, DevCon.
[0:01:05]I have to say, this is my favorite event of the year, and it is great to see so many of
[0:01:10]you again.
[0:01:12]Whether this is your first DevCon or you've been hitting double digits like me, welcome
[0:01:17]to everyone.
[0:01:19]DevCon has always been a place where we take on the hardest problems.
[0:01:23]We look to figure out what's possible.
[0:01:27]And in this time of great change, I can't help but look back at times of other significant
[0:01:32]change.
[0:01:34]So if you will, imagine this.
[0:01:37]You're an astronomer in the 1600s.
[0:01:39]You lived your entire life knowing that the Earth is at the center of our solar system.
[0:01:46]You worked with tools and formulas that confirmed this model.
[0:01:49]You're so confident in this thinking.
[0:01:51]Then there's this Galileo guy.
[0:01:54]He comes along with a few glass lenses, points him at the night sky.
[0:02:00]Now with a telescope, something is wrong.
[0:02:03]He discovers it's the sun, not the Earth.
[0:02:07]That's at the center of our solar system.
[0:02:10]This changes everything.
[0:02:12]Everything needs to be recalibrated.
[0:02:14]Every assumption, every model.
[0:02:17]This idea is so disruptive, they throw them in jail for it.
[0:02:22]But, tries they did.
[0:02:25]There was no going back to that old Earth-centric model.
[0:02:28]Astronomers from that time forward couldn't go back.
[0:02:33]You see, when tools change, the boundaries of what's possible change within them.
[0:02:38]Agenetic AI is to the developer community what the telescope was to 17th century astronomers.
[0:02:45]And I think many of us are already feeling that.
[0:02:48]Can we bring up the lights in the room?
[0:02:52]Raise your hand if you believe software development is changing.
[0:02:58]Even in the overflow room.
[0:02:59]Go ahead.
[0:03:00]Take a chance.
[0:03:01]Go around.
[0:03:04]It's almost everyone.
[0:03:07]And if you're still skeptical, that's okay.
[0:03:10]At least you won't be thrown in jail like Galileo.
[0:03:15]So, all right.
[0:03:22]For decades now, we have moved data by hand.
[0:03:26]We built static algorithms and we wired workflows ourselves.
[0:03:31]Now we have a new opportunity.
[0:03:33]We get to express intent.
[0:03:35]And the system it partners with us to achieve these outcomes.
[0:03:39]This means rethinking how products are built and workflows are designed and how value
[0:03:44]is created across the design and make ecosystem.
[0:03:48]This is a new model where success will not be defined by a collection of algorithms and
[0:03:53]features.
[0:03:55]It's going to be defined by the complex outcomes that we deliver that were once out of reach.
[0:04:01]Now this requires us to do three things.
[0:04:05]Build, orchestrate, and scale.
[0:04:09]And this community of developers has been innovating on the platform for over a decade,
[0:04:15]taking APIs and building real world solutions that customers depend on each day.
[0:04:21]And now, Agente.ai expands what we can build.
[0:04:26]The tools you build won't just be standalone.
[0:04:29]They're going to be orchestrated inside workflows, called by agents to achieve outcomes.
[0:04:34]And build an orchestrated?
[0:04:36]That's the foundation of what you're going to hear more about today.
[0:04:39]And then tomorrow, we're going to talk about scale.
[0:04:43]How agente workflows move beyond one team or project to become new ways that organizations
[0:04:48]work.
[0:04:49]Today, you're going to hear from Shelley, who's going to focus on data and Vikram on how
[0:04:55]discovery works in the Agente era.
[0:04:58]We're attached to showing us the show us orchestration and action.
[0:05:02]But first, let's look at how this shift changes the ecosystem opportunity for us.
[0:05:09]Please join me in welcoming to the stage, Rajee Arasu, our CTO and someone who has never
[0:05:15]stopped being a builder.
[0:05:19]Thank you, Ben.
[0:05:29]And good morning, everyone.
[0:05:32]Good morning.
[0:05:35]Every time I stand in front of this community, I get goosebumps.
[0:05:40]Because this room is filled with builders and visionaries, connectors, the people who
[0:05:45]turn ideas into reality.
[0:05:50]About three decades ago, when I started my journey in computer engineering, development
[0:05:56]looked very different from today.
[0:05:59]I saw during circuit boards and writing assembly code.
[0:06:05]And honestly, thank goodness I don't have to do that anymore.
[0:06:09]My programming was basically the developer's painful version of doing a taxes.
[0:06:15]A simple typo, and it doesn't just trigger an error, it silently collapse the entire
[0:06:21]system.
[0:06:23]Nothing worked.
[0:06:24]And nobody knew why.
[0:06:27]But here's something.
[0:06:28]I have observed over the few decades.
[0:06:32]Developers always move up the stack.
[0:06:35]We automate what is repetitive.
[0:06:38]We abstract what is complex.
[0:06:40]And we spend our time solving higher order problems.
[0:06:44]First we automated the machine coding.
[0:06:47]Then we automated circuit board assembly.
[0:06:49]Thank goodness.
[0:06:50]And then infrastructure moved to the cloud.
[0:06:53]Each step didn't make developers less important.
[0:06:57]It made us move on to things that were more impactful and more valuable.
[0:07:03]And now we are in this AI era.
[0:07:08]Things like this have always felt overwhelming, even scary, especially this one.
[0:07:15]Because they challenge the stability structures we have built over the decades, the jobs we
[0:07:20]have created, the digital architectures we have designed, the companies we operate, and
[0:07:24]the systems we trust.
[0:07:27]But history shows something important.
[0:07:31]Every time the technology automates work, humans move up the ladder to more meaningful
[0:07:36]work.
[0:07:37]AI will definitely fill many gaps where humans struggle.
[0:07:42]Its pattern discovery, repetitive reasoning, automation at scale, you name it.
[0:07:48]But there are still things that AI cannot figure out without our help.
[0:07:54]Context, judgment, responsibility.
[0:07:58]So the real question is not, will AI replace developers?
[0:08:02]The real question is, what higher order work will we as developers do next?
[0:08:09]And that is something this community can and will define.
[0:08:14]And when I look at this community, I already see the early pieces of future that have already
[0:08:19]put in place.
[0:08:21]For years, you have powered automation.
[0:08:24]You have connected workflows, reducing manual work that was delivered through our APIs,
[0:08:31]Autodesk platform services, and what is referred to as APS.
[0:08:36]So now let's watch how Andrids is using APS to automate at scale.
[0:08:48]With Autodesk construction cloud, we want to bridge the gap between design, manufacturing,
[0:08:53]construction, and service operation.
[0:08:55]The ultimate goal is that these different worlds are really interconnected with each other.
[0:09:05]Andrids is an engineering company doing large scale state of the art engineering projects.
[0:09:10]We have around 30,000 employees in over 80 countries around the globe.
[0:09:15]We built custom solutions based on APS.
[0:09:17]The biggest project for us was integrating ACC into our system landscape.
[0:09:22]Before starting with ACC, it was really hard for our users to get all the related documents
[0:09:28]they need, for example, on site.
[0:09:30]When we started with this integration, we wanted to do it differently than before.
[0:09:34]So building a centralized integration pipeline with a centralized business logic where the
[0:09:39]sewer systems does not need to know anything about ACC.
[0:09:44]They just send the data in their format with their metadata they have, and we have a centralized
[0:09:50]business logic based on that information map it to the destination system in ACC.
[0:09:56]So we used all of the APS components to interact with ACC.
[0:10:01]Is it ACC API to interact with projects, admin workflows, and so on, but also data management
[0:10:07]to upload data, then design automation, because we also want to process data on their journey
[0:10:13]from the sewer systems to ACC.
[0:10:15]So our most important topic here was to build an integration pipeline to bring all these
[0:10:21]documents from these systems automatically to ACC.
[0:10:24]In our first projects, we had already more than 30,000 documents, and nowadays in our
[0:10:30]biggest project, we have close to 500,000 documents in.
[0:10:34]Doing that manually is two time consuming and error-browing, because if some person forgets
[0:10:40]to upload the latest version, people on site assemble it based on an outdated version.
[0:10:44]We save more than 85% of manual work load in automatizing this.
[0:10:51]Amy will remind me that it's former, not ACC.
[0:10:58]That's putting APS in action.
[0:11:00]With 85% manual work, Andreds now has fully automated workflows at scale.
[0:11:06]Now imagine what happens when AI builds on top of that.
[0:11:11]Two APIs and AI this community will drive that future of automation.
[0:11:17]I'm certain about that.
[0:11:19]But that future doesn't arrive all at once.
[0:11:22]It happens in waves.
[0:11:24]And every time a new wave arrives, developers move up the stack.
[0:11:29]The first wave is task automation.
[0:11:32]AI helping with individual tasks, like generating designs, summarizing information, and accelerating
[0:11:39]and repetitive work.
[0:11:42]This is incredibly valuable today, but over time, it becomes table stakes.
[0:11:49]The second wave is workflow automation, which connects those tasks into end-to-end processes
[0:11:55]across design, engineering, and construction.
[0:11:59]That is where real differentiation starts to happen, when these workflows are solving
[0:12:05]with context, business rules, and domain expertise.
[0:12:09]And then the third wave is systems automation.
[0:12:13]Entire systems operating with coordinated intelligence projects, factories, production
[0:12:19]pipelines.
[0:12:20]They continuously are adapting and optimizing over time.
[0:12:26]To summarize, task automation is a starting line.
[0:12:30]Workflow and system automation is what will help you deliver value through this higher
[0:12:36]order work in this AI era.
[0:12:39]Today across our industries, there are still countless broken workflows.
[0:12:44]And solving them is bigger than any one company can do it.
[0:12:49]It's the opportunity for this community.
[0:12:54]Last year, at DevCon, when we talked about the shift to AI-driven automation, two questions
[0:13:00]came up.
[0:13:01]First, will AI work inside workflows?
[0:13:06]Second was how will AI solutions be discovered and monetized?
[0:13:10]These questions are even more acute now.
[0:13:15]So let me start with the first question.
[0:13:17]Will AI work inside real workflows?
[0:13:20]Yes.
[0:13:21]AI has moved from experimentation to taking action where work happens.
[0:13:28]So let's look at an example of how this has transformed the world of coding, which
[0:13:32]I know you will be able to associate to that.
[0:13:36]Think about where we started with tools like AWS, Ciro and CloudCode.
[0:13:42]When they first appeared, they could generate a few lines of code.
[0:13:46]Impressive, but still just a tool.
[0:13:50]Then something changed.
[0:13:51]They started to understand the repository, the project structure, the dependencies, the
[0:13:56]test.
[0:13:57]They could propose changes.
[0:13:59]They could run risks and even open pull requests.
[0:14:04]And that's when something clicked for developers.
[0:14:08]That wasn't just about AI generating code anymore.
[0:14:11]It was an agent that was participating in the workflow.
[0:14:15]Suddenly, it felt less like a tool and more like a coding partner.
[0:14:22]What made that possible wasn't just a smarter model.
[0:14:26]It was context.
[0:14:28]Context of the code base, the workflow, the system it operates in, and this applies everywhere.
[0:14:35]In design and make workflows, agents need MCPs and context to access and understand the
[0:14:42]3D design model, the project state, the engineering intent, and that's exactly what we are enabling.
[0:14:51]We are starting with Autotest Public MCPs.
[0:14:55]And as you build agents, so you can actually do that and operate inside a design and make
[0:15:00]workflow.
[0:15:01]We're introducing MCPs for Fusion and Revit with many more on the way this year.
[0:15:08]And with the right MCPs, an associated context, AI moves from demos to real automation.
[0:15:17]Now let's look at the second question.
[0:15:20]How will AI solutions be discovered and monetized?
[0:15:24]I know this is on your mind.
[0:15:26]AI is shifting discovery from traditional search pages into the workflow.
[0:15:33]It's surfacing solutions at the moment where users express their intent.
[0:15:39]This changes how we build.
[0:15:41]It's not just about being found, but it is about being understood, trusted, and chosen
[0:15:47]by the system.
[0:15:49]Later, Vikram is going to be up here and he's going to dive deeper into this shift and
[0:15:53]how you can start building for it.
[0:15:57]I want to focus on the two big steps we're taking to help you show up in this new model.
[0:16:04]First, we are opening Autotest Assistant to third party MCPs.
[0:16:11]Assistant is the connective AI layer across Fusion, Revit, Autocad, Civil 3D.
[0:16:17]It is embedded directly in the flow of work.
[0:16:20]And when your MCPs integrate here, it becomes part of how customers designed, automate,
[0:16:27]and execute their work.
[0:16:30]Sit tight.
[0:16:31]You'll get to actually see this in action later in the keynote where Ritesh will be on
[0:16:36]the stage showing this off.
[0:16:38]Now, in addition to plugging your MCPs into Autotest Assistant, we are doing one more thing.
[0:16:45]We are launching a design and make marketplace.
[0:16:49]An AI first destination for our customers, where your solutions show up alongside ours.
[0:16:57]It elevates the app marketplace that you know and you trust today, with a new destination
[0:17:02]for agents and MCPs bringing everything together in one experience.
[0:17:09]This marketplace will expand to include monetization capabilities, enabling your
[0:17:15]agent workflows to scale, to reach more customers and generate revenue for you.
[0:17:21]As we design these systems, your distribution and discovery are always front and center
[0:17:27]for us.
[0:17:28]When I think about this moment in history, I think back to those early days of soldering,
[0:17:36]circuit boards, and writing assembly code.
[0:17:39]That felt like the hardest engineering work imaginable.
[0:17:44]It did really like that.
[0:17:46]But over time, we automated it.
[0:17:48]We abstracted it and moved up the stack.
[0:17:52]And sure, still even today, there are some brilliant people who work in chip design and
[0:17:57]Nvidia and AMD shops.
[0:17:59]And the rest of us have moved on.
[0:18:01]We buy GPUs from them and are doing higher auto work.
[0:18:04]Thank goodness.
[0:18:06]So I believe you with a simple question, as you climb up to the next level of the stack,
[0:18:13]what will you build?
[0:18:17]Now, whatever amazing things you will build, you always need good data.
[0:18:25]And that brings me to the next speaker.
[0:18:28]He's not just passionate about data.
[0:18:30]He's obsessed about it.
[0:18:33]He spends his day thinking about how to make it more granular, more connected, or how
[0:18:38]to drive intelligence from it.
[0:18:40]Which is why we internally call him our data king.
[0:18:44]So please welcome Shelley Mustafa.
[0:18:47]Yes.
[0:18:48]Yes.
[0:18:49]Yes.
[0:18:50]Yes.
[0:18:51]Yes.
[0:18:52]Yes.
[0:18:53]Yes.
[0:18:54]Yes.
[0:18:55]Thank you, Rajee.
[0:18:58]Look.
[0:19:00]Look, I'm no king, but I do love data.
[0:19:06]And if I might add, I actually look pretty good in red.
[0:19:11]But seriously, I do love Rajee's framing.
[0:19:15]Moving up the stack.
[0:19:18]Because every time we move up the stack, we take on more complex higher auto problems.
[0:19:24]And those problems demand data that is connected, reliable, ready to be used.
[0:19:31]No matter how powerful your AI becomes, it only works as well as the data foundation
[0:19:37]beneath it.
[0:19:38]And when you think about building on that foundation, it's worth asking what actually
[0:19:45]makes systems last?
[0:19:49]I had this moment of clarity recently on a family trip to Puerto Rico.
[0:19:53]The highlight of our trip was visiting Castillo, San Felipe del Moro in Olsen Juan.
[0:19:59]It is considered a masterpiece of Spanish engineering.
[0:20:03]But it began as a single cannon on a cliff.
[0:20:08]It evolved over 250 years into a six level structure spread across 75 acres with walls
[0:20:16]that are 40 feet thick.
[0:20:21]Looking along those walls, I thought about the engineers who passed their knowledge from
[0:20:25]one generation to the next to adapt to a changing world.
[0:20:30]Fixed components were replaced with dynamic mechanisms.
[0:20:34]Circular tracks enable structures to rotate with flexibility.
[0:20:39]Sequential layers were designed to ensure access control in stages.
[0:20:44]These engineers built for the future.
[0:20:48]Interability to adapt is what has kept El Moro relevant for centuries.
[0:20:54]While tools and techniques might change, good engineering will adapt to endure.
[0:21:02]These last few years, we have been witnessing another turning point for engineering as a
[0:21:06]discipline.
[0:21:08]As engineers, we are trained to drive determinism, precision, and repeatability.
[0:21:16]Now, suddenly, we are working with these probabilistic, large language models that can give
[0:21:20]you three different responses to the same prompt depending on the day.
[0:21:25]We have all seen it.
[0:21:27]When it is wrong, it is confidently wrong.
[0:21:30]Look, we love the potential.
[0:21:35]We just want to be able to trust the outcomes.
[0:21:38]It's almost as if we have been handed a magic wand with no user manual.
[0:21:43]So we have had to adapt to come up with ways to harness this power while containing its
[0:21:50]weaknesses, to guide these systems with context, focus, and grounding.
[0:21:59]We are in an era where AI promises incredible possibilities, but realizing those possibilities
[0:22:05]depends on two things.
[0:22:08]Your engineering expertise and the data you build upon.
[0:22:13]It's the same principle we saw at El Morro.
[0:22:16]You need a strong foundation that can evolve over time.
[0:22:20]And today, that foundation is your data.
[0:22:24]We are building capabilities so your data can move across product and company boundaries
[0:22:29]without friction.
[0:22:32]Because when that data becomes accessible, your expertise becomes the driver and AI becomes
[0:22:37]the accelerator.
[0:22:40]So let's start with how we are evolving to make data granular and available through
[0:22:44]the AAC data model.
[0:22:47]First, we gave you revert properties in the data model.
[0:22:51]But then we heard, okay, but what am I supposed to do with our geometry?
[0:22:57]So we built the next layer.
[0:22:59]With granular revert geometry coming to the AAC data model, we are transforming what used
[0:23:04]to be locked inside files into live cloud-based elements.
[0:23:10]But then some of you said, it's great to have geometry.
[0:23:15]But I would like to write my enterprise data back into the model.
[0:23:20]Fair point.
[0:23:22]With the general availability of the AAC data model extensibility, we are turning the data
[0:23:27]foundation into something you can shape.
[0:23:30]Data that fits your workflows instead of the other way around.
[0:23:35]Teams and partners can build apps, automations, and custom workflows by extending the data model.
[0:23:42]So what do you need next?
[0:23:45]More data types, better sync capabilities.
[0:23:50]I was hoping you would say that.
[0:23:53]Today we are announcing a plan 3D public beta for AACDM.
[0:23:58]I have seen port is coming soon, and later this year we plan to have civil 3D data available
[0:24:03]too.
[0:24:05]And with revert sync capabilities, your foundation becomes more responsive.
[0:24:11]You no longer need to wait for published cycles to access your data.
[0:24:16]So to recap, granular geometry, extensibility, faster sync, more data types, these are not
[0:24:25]separate features.
[0:24:27]There are the building blocks to ground your AI, the data foundation for your agent workflows.
[0:24:37]Once teams have structured data, something powerful happens.
[0:24:43]It changes what becomes possible.
[0:24:48]It's the same parent of evolution we saw at El Moro.
[0:24:52]Once engineers mastered precision, they didn't stop there.
[0:24:56]They engineered ways for components to move and respond together.
[0:25:01]They turned static strength into dynamic capability.
[0:25:06]Granularity gives us that strength.
[0:25:09]Interoperability gives it motion.
[0:25:13]If you were here last year, you might remember that I shared the stage with Paul Heddedell.
[0:25:19]Paul's director of technology partnerships and digital transformation.
[0:25:24]During our conversation, he mentioned something that stuck with me.
[0:25:28]Interoperability isn't just a customer problem.
[0:25:33]It is an industry problem.
[0:25:37]And that is why we are building solutions not just for our products, we are building solutions
[0:25:43]for the entire industry.
[0:25:47]Last November, we announced the general availability of the data exchange cloud connector for
[0:25:51]IFC.
[0:25:52]Now, come on.
[0:25:54]I know how big of a deal IFC is for this audience, so I was expecting some applause.
[0:26:01]Thank you.
[0:26:04]And today, we are extending this approach to one of the most critical tools in construction
[0:26:09]coordination.
[0:26:11]Navisworks.
[0:26:13]The Navisworks connector is now generally available, enabling model data to sink across
[0:26:17]Revit, Tecla, Rhino, and Mentor and Power BI.
[0:26:22]No more manual exports, no more version control issues.
[0:26:26]And there's more.
[0:26:28]With the civil 3D connector coming soon, we are bringing infrastructure into the same connected
[0:26:33]systems.
[0:26:34]So, large complex projects can finally scale with automation.
[0:26:40]And we are already seeing customers put these capabilities to work.
[0:26:44]Take Deutsche Bahn, one of Europe's largest mobility and logistics companies.
[0:26:51]They receive a huge volume of IFC files from third party applications and managing those
[0:26:56]files inside Revit was creating time-consuming bottlenecks.
[0:27:01]Now when the files arrive, their teams use the IFC connector to turn them into clean data
[0:27:06]which can be loaded into Revit for documentation.
[0:27:11]Or look at DuriverMir, one of the most innovative construction companies located right here
[0:27:17]in the Netherlands.
[0:27:20]They are using the AC data model and data exchange ecosystem to build structured, reusable design
[0:27:25]data that can inform model checks and dashboards.
[0:27:29]That means earlier, smarter decisions across every project.
[0:27:36]And as powerful as these connectors are, they are only the beginning of what true interoperability
[0:27:43]can unlock.
[0:27:45]Quick check.
[0:27:47]How many of you feel like you spend too much time in meetings?
[0:27:51]Okay, anyone who doesn't have the hand up, please come talk to me after the conference.
[0:27:55]I need to learn what you are doing.
[0:27:58]But if you're like me, the meetings that are most satisfying are the ones where we are
[0:28:03]actually getting something done.
[0:28:06]Not assigning tasks but completing them.
[0:28:09]And that is why I am so excited to share with you all an early look at data exchange life.
[0:28:18]This is a chance to see evolution in real time.
[0:28:23]DX Live brings streaming interoperability directly into your workflows.
[0:28:28]As you can see, when someone makes the changes in Rino, the user in Revit can see those updates
[0:28:34]seamlessly.
[0:28:36]This means teams can resolve clashes together, see updates in seconds, and close out entire
[0:28:42]issue sets in a single working session.
[0:28:45]All in their tool of choice.
[0:28:49]Downstream reporting evolves into a live shared experience where your data, models, and decisions
[0:28:55]can all move at the same speed.
[0:28:59]This is what fewer meetings on your calendar looks like.
[0:29:05]And that sets the stage for the next evolution.
[0:29:09]Data that is not just connected but fully accessible to AI that can act on it.
[0:29:16]Accessible data shifts the center of gravity in your workflow from managing information
[0:29:21]to making your data work for you.
[0:29:25]Nowhere is that more important in places where public utility and safety depends on getting
[0:29:31]decisions right.
[0:29:34]That is exactly what one customer Singapore is working on today.
[0:29:40]For building construction authority.
[0:29:43]Reliable data isn't just helpful.
[0:29:45]It is essential.
[0:29:48]To tell us more about how they are turning their data into solutions, I would like to invite
[0:29:52]BC as deputy CEO, TechTie, hang to the stage.
[0:30:04]After months of seeing you on a Zoom screen, it is great to be here in person with you.
[0:30:08]Absolutely.
[0:30:09]Likewise.
[0:30:12]So, Tech, welcome.
[0:30:15]Can you tell the audience a little bit about BC as mission and what kind of projects you work
[0:30:20]on?
[0:30:21]Right.
[0:30:22]So BCA is a government agency in Singapore.
[0:30:25]And at BCA, our mission is twofold.
[0:30:29]One is to make sure that Singapore's built environment is safe and code compliant and
[0:30:34]to make sure that it is so future-ready.
[0:30:37]Let me speed you through some history.
[0:30:39]In 2013, BCA has been championing industry-wide adoption of BIM.
[0:30:45]Three weeks ago in Singapore, Autodesk former received its BSI kind-mark certification that
[0:30:51]recognises its support for ISO 19650 standards, workflows and functionality.
[0:30:56]Nice.
[0:30:57]BSI has also developed an ISO 19650 Singapore National NX to support its use case in Singapore.
[0:31:05]So Singapore is now moving into mandated 3D model-based regulatory approvals called
[0:31:10]CONNETX.
[0:31:13]CONNETX sits at the intersection of policy, technology and industry transformation, bringing
[0:31:19]together multiple public agencies and the industry into a single trusted digital process for
[0:31:25]building approvals.
[0:31:26]I see.
[0:31:27]So, when you're dealing with this level of complexity, what are the challenges you face
[0:31:32]when it comes to coordination and compliance?
[0:31:35]Well, let me give everybody an idea.
[0:31:38]So for CONNETX, we work with six other government agencies and each of us are responsible for
[0:31:45]an aspect of the build environment.
[0:31:47]When we commence the CONNETX project, we realise that all together among seven agencies,
[0:31:52]we have 6,000 rules of which 4,000 are deterministic.
[0:31:57]On the other hand, the 3D models need to be coordinated and federated between the
[0:32:02]architects, civil and structural engineers and the mechanical and electrical engineers
[0:32:07]approved by the developers and ideally blessed by the builders before submitting the CONNETX
[0:32:12]for approval by the authorities.
[0:32:15]Imagine the cognitive load on the designers and the offices.
[0:32:20]To assist our developers and processing offices, we need a model checker.
[0:32:26]This is the holy grill of our BIM modelling world.
[0:32:29]Those models that are not compliant cannot be built.
[0:32:33]Yeah.
[0:32:34]So, and I know you're working with Autodesk to solving this problem.
[0:32:37]Can you tell us more about that?
[0:32:39]So it turns out that the CONNETX mandate to force everybody to use 3D BIM models for the
[0:32:44]process.
[0:32:45]It actually lays the conditions of outcome-based BIM.
[0:32:49]Now, the REVIT 3D models are very rich in data.
[0:32:53]An extraction of that data was really given a legum with the launch of the AAC data model.
[0:32:59]The data will fill the Autodesk assistant.
[0:33:02]Now our big hairy audacious goal is to really let the system do what humans cannot and automated
[0:33:08]model checker to consistently check thousands of deterministic rules across every submitted
[0:33:15]model.
[0:33:16]So this allows our processing officers to focus on judgment and accountability and delivering
[0:33:22]a consolidated response within 20 working days while accommodating, let's face it, our
[0:33:28]architects out there who still love to ship spaces a little bit too creatively.
[0:33:31]So, if you can achieve that, it is conceivable that an automated model checker that is embedded
[0:33:39]into a generative design engine can one-day generate models which are truly born compliant.
[0:33:46]For now, we will just settle for faster approvals, faster reviews, higher confidence in our
[0:33:51]outcomes, and no compromise on quality and regulatory rigor.
[0:33:55]Wow.
[0:33:56]So, that's the only way I can sign up for.
[0:34:00]So last but not the least, for folks in the audience who are facing similar challenges,
[0:34:06]where should this start?
[0:34:07]Well, a few independent estimates placed the market size of this opportunity at between
[0:34:13]$100 million to a few hundred million dollars in Singapore alone.
[0:34:17]With the Singapore agencies organized now to help you be a partner to tackle this complex
[0:34:23]problem and the strategic partnership between Autodesk and BCA, we really hope to present
[0:34:28]the flywheel to all of you in the audience for you to plug into, to develop, test, and
[0:34:33]launch your tools, perhaps on the design and make marketplace.
[0:34:38]So well, AI tech is evolving rapidly and what once felt like a moonshot is really now within
[0:34:43]beach.
[0:34:44]Indeed.
[0:34:45]Indeed.
[0:34:46]Thank you, Tech, for your time here today and for those of you wanting to learn more,
[0:34:49]check out BCS class during DevCon.
[0:34:51]Yeah.
[0:34:52]Looking forward to seeing you guys there.
[0:34:53]Thank you.
[0:34:54]Thank you.
[0:34:55]Thank you.
[0:34:56]For centuries, great engineering has thrived on one principle.
[0:35:08]Design to adapt.
[0:35:11]The success of intelligent workflows won't be defined by flash AI features, but by the
[0:35:17]foundation of structured data that you will build.
[0:35:22]This is your opportunity to build that foundation, grow the ecosystem, and scale your impact.
[0:35:29]Next year, you could be the one on stage telling us all what is possible.
[0:35:36]Now that we have talked about building the foundation, the next step is understanding
[0:35:40]how your work gets discovered and used.
[0:35:44]And there's no one better to talk about it than Vikram.
[0:35:48]Vikram began as an engineer and now he brings the same lens to analyze how AI is reshaping
[0:35:54]consumer behavior and marketing.
[0:35:56]Vikram, take it away.
[0:36:01]Yes.
[0:36:04]Yes.
[0:36:07]Yes.
[0:36:08]Yes.
[0:36:09]Thanks, Shelley.
[0:36:10]You do look good in red.
[0:36:13]He's right.
[0:36:14]I started my career as a software engineer building a virtual reality toolkit that worked
[0:36:20]across multiple operating systems and headsets.
[0:36:25]Now this was the 90s.
[0:36:27]640 by 480 was best in class resolution.
[0:36:31]And Vertigo, that was essentially a core feature.
[0:36:36]So let's just say that experience accelerated my move into marketing.
[0:36:42]And it also taught me something.
[0:36:45]Engineering builds the tools how people discover and use them determines whether they succeed.
[0:36:52]And now these two are directly connected.
[0:36:56]So let me start with a quick question.
[0:36:58]Quick show of hands, how many of you are mostly using chat GPT or the AI answer on Google
[0:37:02]instead of clicking on all those links in Google search?
[0:37:06]Let's see.
[0:37:08]Let's see.
[0:37:09]A lot of hands.
[0:37:11]This is what we are seeing everywhere.
[0:37:15]And the data backs it up too.
[0:37:16]I want to share a sneak peek from a 2026 spotlight report on AI.
[0:37:23]So out of the 2500 industry leaders we benchmarked, 98% are already using AI tools.
[0:37:30]98% but here's the catch.
[0:37:35]Only 19% have actually integrated AI into their core workflows.
[0:37:41]Half-CAP tells us something important.
[0:37:44]Consumer behavior has already changed but how we built hasn't caught up yet.
[0:37:51]For a long time the question, how do people discover and use what we build was mostly a marketing
[0:37:57]question.
[0:37:58]Now it's a developer question too.
[0:38:02]Because the way people find and use things has fundamentally changed.
[0:38:08]From searching to simply asking, from defining steps to defining outcomes and then getting
[0:38:17]them done.
[0:38:18]Look, this change isn't coming folks.
[0:38:21]It's already here.
[0:38:23]Organic traffic from Google has dropped across B2B.
[0:38:26]In some sectors it's reported to be down by almost 70 to 80%.
[0:38:32]In and company reports that click through rates for B2B software categories are down as
[0:38:37]much as 30%.
[0:38:40]And look, our address is in the moon.
[0:38:42]We are seeing it in our numbers too.
[0:38:45]We are having to rethink how we show up, how we get discovered, how we get used.
[0:38:52]So I would challenge all of you.
[0:38:54]Are you building for how people used to work or how they work today?
[0:39:01]So I'm here to break down what's changing and how to build for it.
[0:39:05]So we're going to walk through three key shifts.
[0:39:09]First, how discovery is evolving, how orchestration is reshaping, how work gets done, and then
[0:39:17]how distribution determines where you show up.
[0:39:21]Understand these changes and you can build for how work happens today.
[0:39:27]The teams that move early won't just adapt.
[0:39:31]They will define what comes next.
[0:39:34]All right, so are we ready?
[0:39:36]Let's talk about discovery.
[0:39:38]For a long time discovery works through search.
[0:39:41]You would get a list of links, click through a few, and then pick one.
[0:39:47]That's not what people expect anymore.
[0:39:49]Now people ask a question and they expect a recommendation right away in context inside
[0:39:57]the tool they're already using.
[0:40:00]And that answer is increasingly generated by AI.
[0:40:05]Which changes the game, not just today, but for what comes next.
[0:40:09]You are no longer optimizing for ranking.
[0:40:13]You are optimizing to be the answer.
[0:40:17]So let me make that real.
[0:40:19]Imagine a developer asking an AI tool.
[0:40:22]What tool can I use to design a component, maybe automated workflow?
[0:40:27]Or analyze a model?
[0:40:30]In the old world, whether you showed up dependent on SEO and ranking.
[0:40:36]If you did that well, you had a shot at getting clicked.
[0:40:41]But in this new world, AI tools, chat, GPT, Claude, Gemini, audit assistant, return answers.
[0:40:49]No second chances.
[0:40:52]Which means the AI tool has to understand your capability and trust it enough to include
[0:40:57]it.
[0:40:59]That's what generative engine optimization or GEO is all about.
[0:41:05]So as Shelley said earlier, this only works when grounded in expertise and reliable data.
[0:41:13]Look in these workflows, something has to be chosen.
[0:41:16]And the answer that gets chosen gets used.
[0:41:21]You can be that answer.
[0:41:24]Look AI tools learn the same way developers do.
[0:41:27]From articles, documentation, code, and real examples of how things are used.
[0:41:33]Which is good news.
[0:41:35]The system rewards what actually works.
[0:41:39]Not just what ranks.
[0:41:41]To make sure you show up as the answer, be clear about what your capability does.
[0:41:47]There are real examples that actually work.
[0:41:51]Be present where developers and customers solve problems, including community forums and
[0:41:57]discussions.
[0:41:59]Now discovery is only the first step.
[0:42:03]The next step is orchestration.
[0:42:06]How your capability gets chosen and used as part of a workflow.
[0:42:11]Now in this world, orchestration starts with intent.
[0:42:15]Instead of looking for tools or designing solutions step by step, users are simply expressing
[0:42:22]what they want done and the outcome they're looking for.
[0:42:25]They might say, optimize this design.
[0:42:29]Analyze this model.
[0:42:31]Have a stroke waffle delivered to my hotel.
[0:42:34]I did that.
[0:42:35]It was delicious.
[0:42:37]And the AI tool decides what to do and how to do it.
[0:42:42]So let me give you a real example.
[0:42:44]The group of engineers at General Motors moved away from designing step by step and started
[0:42:49]by defining the outcome.
[0:42:52]So in designing a seatbelt attachment, they set constraints like weight, materials,
[0:42:57]and manufacturability.
[0:42:59]And they used Autodesk Generative Design.
[0:43:02]In one case, age separate parts became one.
[0:43:07]40% lighter.
[0:43:08]20% stronger.
[0:43:10]That's what intent looks like.
[0:43:13]And it changes the role of the engineer and what we build as developers.
[0:43:18]So as Ben said earlier, when the tools change, everything has to be recalibrated.
[0:43:24]The job is no longer to design each component of the tool.
[0:43:28]It's to define the outcome and guide the system to get there.
[0:43:33]That's higher order work.
[0:43:36]And for companies, that's where real value gets created, not in executing individual
[0:43:41]steps, but in achieving better outcomes across the entire process.
[0:43:48]Now, in the real world, that intent has to be carried across connected systems, orchestration
[0:43:55]connects capabilities, data, and workflows to get that work done.
[0:44:01]And that's exactly what Autodesk Platform Service is enables.
[0:44:07]Now let me give you an example here, too.
[0:44:09]From the Soetra Bridge project, Nour Consult was dealing with a massive coordination challenge.
[0:44:15]Global teams, complex infrastructure, huge, huge volumes of data.
[0:44:21]So instead of managing everything step by step, they connected the systems using Autodesk
[0:44:26]Platform Services.
[0:44:28]So design models, project data, visualization tools were all working from the same connected
[0:44:33]environment.
[0:44:35]You know at one point, over 60 million data points were being coordinated in real time,
[0:44:41]not step by step, but across the entire workflow.
[0:44:47]That's orchestration and action.
[0:44:50]And what it shows is this, when your systems are connected like this and intent is expressed,
[0:44:56]the work can get done faster, better, and at scale.
[0:45:02]So what does this mean for all of you?
[0:45:04]It means you need to build for this world.
[0:45:09]Design reliable capabilities AI can understand.
[0:45:13]Make sure they can plug into unified systems or platforms to be executed as part of a workflow.
[0:45:22]And for those of you leading businesses, this means prioritizing connected systems and
[0:45:27]structured data that allow that work to happen.
[0:45:33]Okay, now let's talk about the final change, distribution.
[0:45:39]Even if your capabilities really, it still has to show up where decisions happen.
[0:45:45]To support you with this, as Raji mentioned, we are introducing the design and make marketplace.
[0:45:51]It's a curated destination where customers can discover AI-powered agents, tools, and workflows.
[0:45:59]In other words, a place where your capabilities can be used and discovered and used at scale.
[0:46:06]We are also opening up Autodesk Assistant to third party MCPs.
[0:46:11]Autodesk Assistant is how customers get done, get work done with AI inside Autodesk products
[0:46:17]like Revit, AutoCAD Fusion, and many, many others.
[0:46:22]This is where your capabilities show up at that exact moment of need inside real workflows.
[0:46:32]This is how you win in the Agentec era.
[0:46:36]Be the answer.
[0:46:39]Turn intent into action and be chosen by the system.
[0:46:46]Now, to show you what intent to action looks like in Autodesk Assistant today,
[0:46:51]please welcome to the stage, Ritesh Pansar.
[0:46:55]Yes.
[0:46:59]Yes.
[0:47:03]Thanks, Rikam.
[0:47:05]As Shalee said earlier, we have been handed this magic wand without a user manual.
[0:47:11]Creating that user manual is where I spend most of my time these days,
[0:47:16]figuring out how to make it all work in real workflows.
[0:47:20]And what you just heard about turning intent into action,
[0:47:25]that's exactly what Autodesk Assistant is built to do.
[0:47:28]How many of you in your day-to-day work are still jumping between tools and chasing information?
[0:47:36]A lot of you, right?
[0:47:38]And all of that work is necessary.
[0:47:42]But it's not where your time should go.
[0:47:46]You should be building designing solving real problems.
[0:47:51]Assistant lets you move up the stack.
[0:47:55]Instead of switching between tools and chasing down information,
[0:47:59]you can just ask and get the work done.
[0:48:04]Behind the scenes, Assistant acts as the orchestration layer, bringing together the right systems,
[0:48:10]data and agents to understand your intent and turn it into action.
[0:48:16]It's context-aware, secure by design and built to work across products so everything connects
[0:48:24]without the friction. And because it's built on your existing data and workflows,
[0:48:31]it fits how you already work.
[0:48:35]But enough explaining it.
[0:48:37]Let's see it in action in a couple of Autodesk products.
[0:48:40]That's where the fun is.
[0:48:43]Now let's imagine I am a project engineer working in Revit on a large campus project.
[0:48:50]I need to create a complex door schedule.
[0:48:54]Traditionally, that takes time.
[0:48:58]Manual setup, multiple steps.
[0:49:01]Now watch this.
[0:49:03]This is not just a list.
[0:49:05]I need to set up multiple parameters, align families and make sure everything groups correctly.
[0:49:13]Now I just ask.
[0:49:16]And in seconds, Assistant builds the full schedule for me.
[0:49:21]Correct fields grouped properly.
[0:49:24]No setup, no guesswork.
[0:49:27]That's documentation done in seconds.
[0:49:31]Now, let's take a look at Fusion.
[0:49:35]For this, let's pretend I am a product designer preparing for a review.
[0:49:40]The model is ready, but creating a visual usually means setup, exports and waiting.
[0:49:49]Now I just ask.
[0:49:52]And in seconds, I have review-ready images from my model.
[0:49:58]No switching tools, no delay.
[0:50:02]Across both workflows, this is the shift from time-consuming work to simply asking.
[0:50:13]From thousands of data points to one informed decision in seconds.
[0:50:20]That's Assistant across design and make data today.
[0:50:25]And we are continuing to expand where Assistant shows up.
[0:50:29]I am excited to announce that Assistant will be coming to Autodesk platform services.
[0:50:35]It will help you take action with APS in new ways.
[0:50:40]But Assistant doesn't stop with Autodesk.
[0:50:44]There's deep domain expertise available outside the Autodesk ecosystem.
[0:50:50]And that's where all of you come in.
[0:50:53]Autodesk Assistant is agent-taking nature and able to call into third-party MCPs agents
[0:51:02]built by this community. Because real work happens across systems, not in silos.
[0:51:10]The design and make marketplace will connect your tools, services, MCPs, agents and expertise
[0:51:19]directly into Assistant. So it just doesn't assist. It acts.
[0:51:26]The marketplace will certify your solution so that they can be trusted, discovered and
[0:51:33]invoked inside real workflows by Assistant. And we are investing in helping you build these,
[0:51:41]starting with MCPs, with MCPs over. Resources, templates and training.
[0:51:48]So you can go from idea to working capability quickly.
[0:51:54]What you are about to see is where we are headed. This isn't shipping tomorrow.
[0:51:59]It's our true north. And we are building it together with all of you.
[0:52:06]Now let's look at an example we are developing one of our partners where Assistant reaches
[0:52:11]beyond Autodesk. Victory AI builds tools that automate complex engineering workflows like HVAC
[0:52:21]design. They recently developed an MCP to integrate with Autodesk Assistant.
[0:52:29]Alejandro is an engineer at a construction company. He's leading an HVAC air supply project
[0:52:37]and needs to validate and update his model against company efficiency, quality standards.
[0:52:44]So the design is buildable, compliant and ready for production.
[0:52:49]Today that means extracting data cross systems, running a lot of external checks,
[0:52:56]in external tools and fixing issues, element by element. It's slow and it doesn't scale.
[0:53:06]With Autodesk Assistant and the Victor MCP, that changes completely.
[0:53:12]Autodesk provides Assistant, the orchestrator and the MCP framework which allows external tools
[0:53:19]and agents to securely connect and work together. Victor brings their domain expertise through their
[0:53:26]MCP. Once Victor lists their MCP in the marketplace, Alejandro can easily discover and subscribe to it,
[0:53:35]bringing Victor's expertise directly into Assistant's ecosystem.
[0:53:40]Where it all comes together in one place for him to use in his project. Let's take a look.
[0:53:46]Alejandro opens Autodesk Assistant and selects the Victor agent. He asks Assistant to identify and
[0:53:56]optimize all the ducts in the air supply system. Assistant connects to the relevant MCP server,
[0:54:04]analyzes the model and highlights all the ducts. It generates a recommendation based on Victor's
[0:54:11]standards. Alejandro reviews the insights and takes action. With one click, Assistant
[0:54:19]optimizes the ducts, adds metadata and links to the Victor app for a summary of the HVAC metadata
[0:54:27]and cost analysis. No manual steps, no friction from hours of work to minutes.
[0:54:35]This is the power of Assistant, not just surfacing insight but executing workflows end to end.
[0:54:46]I just showed you how a certified MCP will work from the marketplace.
[0:54:53]Now let's talk about agents because I know a lot of you have your own data, your own standards,
[0:54:59]and you're making big investments in platforms like Microsoft.
[0:55:05]We are putting this to work. Let's see a real customer example.
[0:55:11]Arquedes is a global engineering firm delivering complex infrastructure projects.
[0:55:17]Yosha, a product owner is leading one, managing multiple teams, systems, and staying on top of
[0:55:24]compliance. Today, getting a clear view means chasing data across systems. Arquedes built a
[0:55:34]maturity model powered by the Microsoft Co-Pilot Studio agent to evaluate project performance across
[0:55:41]data quality, compliance, and security. Here's how it all comes together. Autodesk provides
[0:55:50]Assistant the orchestrator and the agent framework that allows external systems to
[0:55:55]securely connect to platforms like Autodesk Forma. Arquedes brings the domain expertise,
[0:56:02]their company policies, project standards, and project data.
[0:56:07]Co-Pilot Studio through its work IQ intelligence layer powers the agent Arquedes has built.
[0:56:15]All of this connects through Assistant in one place. Let's see how this works.
[0:56:23]Yosha opens Assistant, selects his agent, and runs a maturity check.
[0:56:30]Assistant connects to the Co-Pilot Studio environment and invokes the right agent.
[0:56:36]It evaluates the project across data quality, compliance, and security.
[0:56:41]In seconds, the full evaluation comes back. Yosha sees exactly what needs attention and takes action.
[0:56:52]He asks Assistant to create these issues and assignments.
[0:56:58]And even schedule time with Arquedes experts by checking calendars and looking
[0:57:03]at the meetings. The one's Shelley hates. From insights to actions.
[0:57:11]Do you see why this is so cool? What used to take Yosha
[0:57:16]are opening Autodesk Forma, working step by step, manually scheduling follow-ups one by one.
[0:57:25]Now happens in seconds. By connecting Autodesk tools, Arquedes expertise and Microsoft AI,
[0:57:34]we turn project data into real immediate action. No friction, no boundaries, you just ask.
[0:57:44]Assistant's goal is simple. Solve the customer's problem. To do that, it calls the best capability available
[0:57:55]whether built by Autodesk or one of you in the community. That's why the design and make marketplace
[0:58:03]is so critical. When you have, you list your certified MCP, you're not just adding an integration,
[0:58:13]you're making your capability discoverable and callable by Assistant. Together, we are building
[0:58:21]a trusted orchestration layer. Assistant understands intent, manages context, enforces permission,
[0:58:30]and connects to your expertise. This is the shift. From fixed integrations to
[0:58:38]and predefined workflows to dynamic orchestration across MCPs, agents, and models. Unlocking entirely
[0:58:49]new possibilities. But here's the key. Assistant doesn't have your domain expertise. Your local building
[0:58:59]codes, your supply chain constraints, your industry nuances. You do. Autodesk provides the
[0:59:08]platform, the orchestration, the security, and the customer reach. You provide the specialized
[0:59:15]context, tools, and agents. And together, we are redefining how work gets done. This is your
[0:59:24]moment to move up the stack. From building tools to delivering real outcomes. And this is your
[0:59:32]opportunity to define what comes next. Thank you. Ben, back to you.
[0:59:49]Thank you, Ritesh. That's the Assistant turning data into intelligence powered by you all.
[0:59:57]Reducing complexity for the team. Getting them to better outcomes faster. That's the future.
[1:00:05]And our job, it's so great. We get to build it. We get to write the manual for
[1:00:11]King Data King, Shelley's magic wand. We design the workflows. We decide how agents work across
[1:00:19]disciplines and industries. And this change, the experience for the architect, the product designer,
[1:00:26]and the BIM lead. We can solve the complex problems. Moving up the stack. And everything you saw
[1:00:34]today, build and orchestrate. This is the foundation for what comes next. Then, tomorrow, we're going
[1:00:43]to look at how we scale this automation. And these workflows across the entire system. You're going
[1:00:49]to hear from our industry leaders, Amy, who's synonymous with innovation and AEC. And Shrinath,
[1:00:55]who's one of the most passionate voices for design and manufacturing. Together, they're going
[1:01:00]to show us how these industries are evolving with a genetic AI and where our customers need you
[1:01:06]to lead with innovation. Then, Daniella, known to many of you here in Europe, will show us how these
[1:01:12]industries are converging and how we can accelerate that. It's amazing. I love having all of you here.
[1:01:20]I hope you enjoy all the sessions and I'm going to see you at the reception tonight. Thank you.
16 Apr 2026 · 23m 22s
Revit Api | Mcp: How To Build A Deterministic Ai Agent For Bim Qa
The video explores the development of a deterministic AI agent for BIM quality assurance using the Revit API and Model Context Protocol (MCP). It demonstrates leveraging large language models like Claude and GPT to automate tasks such as extracting information from Revit models, creating architectural elements, accessing Autodesk Form data, and generating reports of issues within BIM projects independently of specific platforms.
Deterministic AI Agents for BIM QALLM-driven model information extractionDeterministic AI in BIM Quality Assurance
RevitAutoCADCivil 3DForge
ClaudeGPTLLM
Demo Segments
- 0:07:21 LLM receives a PDF plan and creates grids, levels, and columns for a model automatically.
- 0:15:42 Using GPT to extract and verify information from Revit models.
- 0:18:05 Interacting with cloud to access Autodesk Form information and produce reports.
- 0:20:55 Listing and reporting issues on a BIM project using an AI agent
- 0:22:25 Accessing and generating a report of issues without specific platform reliance
Key Frames (7)
A slide with an agenda displaying topics like context, architecture, and a live demo.
AUTODESK
A presentation slide with the title 'Context and Problem' and a speaker on stage, with AI in the background.
AUTODESK
The frame shows a slide discussing the context and problem related to Revit models and AI assistants.
AUTODESK
A slide titled 'General Architecture' is displayed alongside an image of a person using a tablet in a workshop setting.
AUTODESK
A presentation slide showing a general architecture diagram with elements like user input, LLM, and tools.
Autodesk
📹 0:07:00
demo
The frame shows a demo of integrating Claude MCP with Autodesk Revit for architectural planning.
Autodesk, Revit
📹 0:12:30
demo
A demonstration showing the integration of GPT agents with Autodesk Revit, including a Revit model view.
AUTODESK, Revit
►
Full Transcript (104 segments)
[0:00:00]Thank you for joining us today. My name is Enrique Mineses and I'm going to present from problems to more of the changes I read the AS system with MCP.
[0:00:11]The agenda for today is the context. Well, we are going to see a little bit about what was the problem and well the context.
[0:00:22]How I developed this application, then we're going to see what is an MCP, an MCP, one part of the A8 called modern context protocol.
[0:00:35]And then after that, we're going to see the general architecture of the MCP and after that, we're going to see a little bit of some examples with videos that I prepared for today. So it's pretty simple.
[0:00:48]Well, this is one of the phrases that inspired me. It says,
[0:00:53]Sinari Garantisa el Maniana, el Oyse webing means, so if no ones of warranties today, sorry, tomorrow today it's in meant so it is like something that inspired me every time to check if I'm doing the things that I need to do every day to continue.
[0:01:14]Okay, let's see a little bit about the context and the problem. The context, well, we have a lot of different kind of models, a lot of different information on that.
[0:01:28]Every single model has a purpose that sometimes you need someone managing that information, also producing that information, that it's too difficult to access to that information when the people are doing the same.
[0:01:43]And then people, you know, just working on it. And also you need to check everything in an easy way to, you know, share the information with the external members.
[0:01:57]Sometimes the people is already involving what they are doing and they don't want to share or they don't know exactly how to do the quality check or they don't know exactly what the other members needs.
[0:02:11]So I want to democratize this process and to leave everyone at the studio information. I guess on AI integration, yet good option.
[0:02:22]Even if the user is not an expert user, if you have an LLM connected to the rabbit and you have some tools posted, you can extract information from rabbit to be used.
[0:02:35]Even if you are not a specialist or not, I really good user, you can just interact with the LLM and then extract information that you want.
[0:02:46]This is one of the biggest options that you want to find MCP. If you are not a big, well, a good user of an specific tool, but you have an LLM that use some tools to extract information or to interact with that software, then you got it.
[0:03:05]Very simple. Well, that was the problem that it would be up to the context, to be honest, is very simple.
[0:03:16]I'm trying to, you know, improve the process, delivering some tools using artificial intelligence to struct information from the models and check that information out of the chem process without passing to the external members.
[0:03:35]I'm just going to talk about the specialist, just to interact with that information in an easy way.
[0:03:41]Now, let's see what is an MCP. Well, the model context protocol, it's one of the options that you have to interact with the LLM.
[0:03:50]The large language model, it's one of the tools that we have to interact with these agents, with these big agents that are companies like GPT or like chat GPT or like a cloud or like pilot are producing.
[0:04:09]The model context protocol is one tool that brings us that opportunity to interact with them in a special way.
[0:04:16]I always pull this technology like let's imagine that the LLM is a chance.
[0:04:25]Well, that chef cannot produce the dish. It doesn't know exactly the recipe that he needs to apply to produce that dish.
[0:04:35]It's not in the correct environment, even if in the recipe he will be able to produce them.
[0:04:47]Well, here the LLM is a chef and the tool that we're going to post to the LLM is the recipe.
[0:04:55]So now that we know how to act with the information that we deliver and also well in the correct environment that's imagine, rabbi, tautaca, dotus, cout, that the environment to be the correct space to produce or to, you know, put it on to run that specific tool.
[0:05:20]So in this way we are combining LLM's, the tools, the environment and well, the outcome meets whatever we want.
[0:05:29]The tool is not, you know, to use trade or if you can do whatever you want, the process can be adapted by you and well, by your team.
[0:05:43]So it's pretty simple.
[0:05:45]Okay, let's see a little bit about the general chapter of the solution. We have the M, sorry, we have the LLM, the LLM, it's the brain, it's the model that is going to do all the processing of the information.
[0:06:04]Then we will have the tools, the tools that we're going to present to the LLM to be executed, the tools can be attached to one specific software or to one specific environment and going to say with the file firm, it's out there, social plot and then we will have the output.
[0:06:26]So it's pretty simple. The process to balance is pretty simple and well, you have the option to add more and more, sorry, I pulled this up.
[0:06:37]For more information to, so now let's go to the example videos.
[0:06:46]Okay, hmm sure.
[0:06:48]Okay, well, here to be showing us one case of study here, I'm interactive with cloud, cloud is going to receive the videos.
[0:07:01]Sorry, the sheets that I want to produce are exposed to some tools, not a lot of create reads, create levels, create walls, create a super frame, means create calls.
[0:07:14]So here the LLM is receiving the plant that I want to model.
[0:07:20]The model, it's going to be produced without any kind of human interaction.
[0:07:28]Well, what is happening here is that them is taking the tool of create reads, it's taking the information that is in the plants and it's producing all the all the great.
[0:07:42]As I said, I have the specific plants to be the model, but it's not just for that kind of PDF plants, you can, you know, just split out the information in an image or in a sketch or something like that.
[0:08:02]And then cloud is going to produce the elements, it has the correct tools. So here cloud is producing them, the levels also, I have the levels in another PDF and produce the levels.
[0:08:21]So now using that creates a levels, what are these going to produce all the columns.
[0:08:29]After that, we're going to produce framing.
[0:08:33]The short important part is that after I shared the information of PDF and the base is already in the model, like creates and levels.
[0:08:45]If you know what you want to model, you don't need to pass it again at the end for something that is just prompting it, you're just using the current, you can interact with cloud produced that modeling.
[0:09:00]In this specific example, I'm trying to interact with cloud and say, hey, I can you model the columns on the specific rate.
[0:09:11]And across each with the other, the verticals, the grid. So it's finishing all the columns with that specific information.
[0:09:22]After it produces it knows exactly where they were produced.
[0:09:29]It has a memory in the background a little bit, the knowledge of what's what he produced in the model.
[0:09:37]So you didn't reach every time the context for sure if you have an image, you have something to add to the LLM, well sorry to the conversation with the LLM.
[0:09:49]It can be told to the conversation, but per C del LLM is going to have the information.
[0:09:57]After you produce the columns, we are going to use the structure of framing tools.
[0:10:02]And here comes the tricky part, even if you're from this good and even if the tool exists, sometimes the LLM is not having the correct or the specific location, because well maybe the location is in the point to us here on the LLM is taking the latest column and it's going to produce the true curve for instance, another part or maybe this scale is different.
[0:10:32]You need to be sure and take care of it. That's why the human needs to still interact with the model and say hey you're producing it where we're going to do girl framing, they start pointing this one on the end point is this one.
[0:10:47]So it is pretty simple, you can adapt and check the prediction, the model prediction to be honest in real time, but you need to take care and have an eye on what it's being produced.
[0:11:05]I think it's a good option to interact with the LLM's, but just to if you want on a specific sample, I produce just a module in an hour maybe for a little for an especially for a rabbit on all the litter for many well trained person.
[0:11:27]Just know that big field you know well the person kept producing just called it maybe in a half an hour or something like that and you're going to spend time prompting it well maybe you're going to spend too much time prompting it and you're going to live time just you know trying to explain to the LLM what it needs to do.
[0:11:48]So you need to be really smart you know in the use of this LLM even if you have the tool sometimes it's better you know just check and review if we need to use it or not.
[0:12:03]After that said well as you can see we'll talk about the model and as done sometimes the LLM is failing but at the end of the day well it can be adjusted.
[0:12:18]This is just one example of what you can do using it.
[0:12:24]Let's pass to the next video that's another example about how we can use the other context protocol that now we GPT.
[0:12:36]So well here is another example using GPT okay well here I am using chat GPT to you know instruct information from the model.
[0:12:51]So what is going to happen here is this is another LLM and this is another large language model and what is going to happen is and need this LLM to extract information from rabbit.
[0:13:07]This is another option because I guess in this specific scenario the user can interact with the model not modeling proceed just struct information and well as you can see the LLM is going to check if the rabbit is open and which kind of elements we have.
[0:13:27]I'm going to start you know just trying to get what just getting information from the floors and the walls and the kind of things so I ask for a communication of floors and then well that LLM is going to give me that specific information.
[0:13:46]Also if I want to start the other information like floors or like other kind of items for sure it can be done.
[0:13:55]But here this special part is what I try to put in the table to have a specific information about this example.
[0:14:07]The other thing that I want to talk about is LLM is not just for cloud LLM is also for GPT or for Microsoft compiled but the use of it it's a little bit different also as you can see the time that chat GPT is taking to think about the question and apply the specific word flow to execute the tool.
[0:14:33]It's bigger than in cloud that well this is something curious to be honest I prefer to get GPT because in cloud we have our limit of use and sometimes could be not a good option if you have a limit but well on agents in GPT you don't have that restriction.
[0:14:56]So it's a good option also to use GPT to struct information also if you want to add information to some specific elements for sure you can select elements and well you can you can check if everything is going correctly or not.
[0:15:13]You can struct information about the levels about the I don't know floors or walls or any kind of any kind of elements.
[0:15:24]Also you can add information to them on if you select floors and you want to add information to want to specific field well for sure you can do it.
[0:15:35]You can also override the visibility graphics in for a bit just to check if everything is going to do it.
[0:15:45]In this example what I did is okay give me all the levels give me all the floors and then please let's check if everything is correct.
[0:15:57]Here the human interaction is a little bit different you need to take care of what you are requesting and what you want to present to your well what you want to strike into
[0:16:13]to your team because even if the process is very simple here I ask for for a dollar's communication you need to check if the information that the LLM is collecting from rabbit it's correct.
[0:16:29]So the user is not going to be replaced every time that you use an LLM use a new tool for sure the user needs to check if the information that is being pulled out from the model is correct.
[0:16:47]Well after that if you evaluate that information you are pretty sure that this information is correct.
[0:16:59]Well that's something that you want to share with your team in the dashboard or any other kind of presentation.
[0:17:09]So this is a good option also if you want to start information from rabbit if you are not on specialist or not a really good user or if you just want to interact with rabbit to add more information.
[0:17:25]I taught on this example for people like project managers or cast to controlling managers or people that is just using rabbit to distract information.
[0:17:43]So well that's another option let's pass to the next video.
[0:17:49]Here's the last video. Here well as you can see the interaction is going to be with cloud and well the interaction is going to be different.
[0:18:05]Here what I am doing is first of all I will do the login on my account using my user and password I'm going to be authenticated.
[0:18:17]So with that specific notification I can check if I don't know if I have the information in a specific project or a specific model.
[0:18:29]So here comes a good part of it even if you're not passing in auto format your information is already there let's let me explain this you are not using auto format per see you are not in the platform but cloud exploring out that information because that information is living in auto format.
[0:18:57]I'm trying to know asking for okay give me all the projects that I have on my account and then give me all the models in one specific project.
[0:19:07]So what is happening is okay let me check you know those former I have access to your projects I'm going to list them and after I list them I'm going to check all the models in one specific project.
[0:19:23]So again the question is going to be addresses to out this close to cloud and well the information is going to come from the LLM and that's it.
[0:19:33]As you can see I already request the qualification of walls and the qualification of some other elements and they can produce reports using that specific information.
[0:19:45]I just pulled out the information of this specific tool because I want to show you okay this is what you can do using this specific workflow.
[0:19:57]So let's pass to the last part of the presentation this is the live demo.
[0:20:03]Here I have a cloud and I already have to give myself and I'm going to interact a little bit with this specific tool.
[0:20:19]So I'm going to request can you tell me the projects that I had on my account.
[0:20:31]The LLM is going to do the request to the out to special class and well he's going to pull out all the elements as you can see his you know checking the tool that we have and after that he's going to take all the information of that tool.
[0:20:53]So here is the list of projects but that's many that I want use this product and tell me the issues we find out.
[0:21:09]So again the information is going to be pulled on the LLM.
[0:21:15]It's going to check if it has the tool to extract that information about the issues and then well the information is going to be pulled in this conversation.
[0:21:27]Sometimes if you have a lot of information well the interaction is going to be difficult because the context window it's really small.
[0:21:38]And well this is going to pass also in rabbit or another so far so you have just.
[0:21:44]And small context window to interact with the with this specific LLM and also sometimes is if you have a lot of a lot of information that time that is going to be taking to produce that report or produce that answer can be more than the time that you maybe can spend.
[0:22:07]If you access to the to the platform so here well it's checking all the all the issues we have to 35 issues on this project and it's doing the list of issues I can produce ultra report of that issues what I'm trying to explain is.
[0:22:25]You are not passing proceed to the specific you know platform you are not on that specific one firm that the request of information is going to work because at the end of the day you are just using the tool and well producing the report.
[0:22:45]So here's a report of the issues and well here's information about them so as you can see we can access to that information without the use of out this culture club and this is not limited to models this is for all the tools that we have trained out this.
[0:23:05]So that was live demo now I guess question answers and then we can finish thank you very much for your time and that's it.
16 Apr 2026 · 21m 12s
Automating Sustainability Metrics: Aps, Ms Fabric, And Ai Co-Pilot
The video discusses the automation of sustainability metrics in construction projects using Autodesk's APS, Microsoft Fabric, and Co-Pilot technologies. It showcases a data pipeline that integrates these systems to provide real-time environmental cost analyses of unique assemblies. The demonstration highlights how these tools can enhance design decision-making early in infrastructure projects.
Automating sustainability metrics in construction using APS and Microsoft technologies
Heimungst
APSACC
Co-Pilot
Demo Segments
- 0:00:00 Introduction to data pipeline from APS to Microsoft Fabric and Co-Pilot
- 0:08:29 Example of coordination model in Autodesk format shown in 3D
- 0:11:49 Executing data pipeline in Microsoft Fabric with authentication and medallion structure
- 0:14:35 Allocating product codes with a custom dashboard for durability analysis
- 0:16:10 Running calculations for environmental cost with SQL database
- 0:17:25 Displaying results in a dashboard showing environmental impact of design decisions
- 0:19:00 Using MCP server in Microsoft Fabric for querying data with Co-Pilot
Key Frames (13)
A slide from an Autodesk DevCon presentation introducing a BIM Data Engineer from Heijmans.
Autodesk
A slide titled 'What we do' featuring images of construction workers and machinery.
Autodesk
A presentation slide with the word 'Why' and an aerial view of a landscape, part of an Autodesk DevCon presentation.
Autodesk
A slide from Autodesk DevCon presentation highlighting that design data is rich, but sustainability insights are often manual and delayed.
AUTODESK
The frame shows a presentation slide about APIs, specifically the Data Management API and an example of a GET request.
AUTODESK
The frame shows a slide from an Autodesk DevCon presentation discussing various APIs including Data Management API and Autodesk Construction Cloud APIs.
Autodesk