Innovation Without Limits: The Blueprint for a Culture Where Anyone Can Build AI Agents

I'm the co-founder and CTO of Semble. I've spent a third of my life in Paris, a third of my life in New York, a third of my life here, and I still can't pronounce words that have R-L-D in a row. So what's Semble? Semble is a clinical system, and it's one of those systems most people have never heard of, but it powers a huge part of private healthcare in this country. If you've seen a private doctor at any point over the last few years in the UK, the odds are that we're the stewards of your personal data and of your health data.

And like most people in this room, for the last two years, someone has been coming into my office or into my desk, I don't really have an office, and has been saying... I don't know why the slides are moving without me touching anything, but okay. Someone's been coming into my office and has been saying, Michael, you're the CTO, you need to put AI everywhere. AI is the next thing. Our business is going to die if we don't stay at the forefront of AI. AI, AI, AI. And my response has been for quite a while that the tech for AI has been quite democratized by the big guys. Like, for most of us, some of us there's an exception, but for most of us, almost any use case that we need to deliver could be delivered through stuff that's been made available to our teams.

But finding the use case, that's really tough. And that's not a tech problem, that's an org problem. And I'm in a rare position to actually have two hats. I'm a co-founder, and I'm also a CTO. And so when there's an org problem, I just change hats and try to figure out what it is that I need to do in order to solve this problem. Now our mission is to enable health professionals to amplify their impact. And that's a bit of a weird thing when it comes to SaaS because virtually every SaaS platform you've ever met or spoken to or used, their goal is to maximize engagement.

They want you to use their tool as much as possible. The more you use it, the more money they make. Think Facebook, Google, all these things. They're trying to get you hooked. But for us, it's the opposite. I'm trying to minimize engagement because every minute a doctor, a customer of ours, doesn't spend on my system. They spend with their patient. They spend doing stuff that they actually went to school to do as opposed to play around with a system. And so if my goal is to minimize engagement, then AI is actually an extraordinary way to do that. Because effectively, the AI can do stuff on behalf of doctors when they understand what it is that they need to do. But I'm a technologist.

Most of my team are technologists. They don't understand the deep pains that our users go through every single day. And so I want to introduce you to Claire. Claire is on our customer success team. She spends most of her days with my slides are again moving on their own. She spends most of her days with our biggest customers. And she understands the problems of our customers in a way that I will never understand. She understands the tiny little issues with our software. She finds the smallest annoyances of our customers, and she knows what they are. And her whole life, just like all of us, she's been interacting with SaaS, whether she realizes it or not.

Effectively, whenever there is a problem that she needs to solve in traditional SaaS, all she needs to do is say, oh, I know another tool that does this thing differently. Or hey, when I use Gmail, there's a button there. Or when I use this other system, or when I use this other system when I was in school, there was another tool. And so her ability to ideate ideas is actually very high because she understands what it is that traditional SaaS does. When it comes to AI, she's not a technologist. She doesn't know how to write code. She's never interacted with it except for maybe chat GPT.

So this year, Claire actually built her very first AI agent. She built it on her own with no one's help. She just went and did it. What it does is it takes a load of communication from her customers that come from a whole bunch of different channels. It puts them into this big AI grinder, and it comes up with really strong insights that she can then go and discuss with her customer. Now, when she built it, the important part is that she actually did the building. It's not that she told someone in tech or she told someone in our various teams, hey, I really need an AI agent that does that. Please go and do it for me. She actually built it herself. She used her own two hands. And when you do that, you understand the strength of the tool you're using. You understand their limits. And you're able to actually really understand deeply what it is that they can do for you.

And so my goal and what I've always believed is that the best ideas come from the most unexpected places across your organization. But how can I expect the people that understand the problem deeply to come up with solutions if they don't understand the tools deeply? And that's actually a knowledge gap when it comes to AI, not to other stuff, but when it comes to AI that most of us in our orgs are probably feeling. You've got technologists and developers and people who understand the tools.

Now, I want to be really clear. I'm not talking about understanding what a transformer is or understand the latest machine learning model or whatnot. When we get into our car to get to the shop in the mornings, we don't understand how pistons get lubricated and how, for most of us, we don't understand how that stuff works. We just know when it is that we need to use a car and we go in and we use it. And that's the understanding that I want everyone at Semble to have and to understand really, really deeply.

And so the problem I'm solving is that I want the people closest to the problem to understand the technology, and I want the knowledge to live in a single person. And that's what happened at Semble, but it didn't happen by accident. It was a grueling piece of work. The way we made it possible is this four-pronged approach. The very first thing is we had to actually choose a platform that effectively allowed us to build the conditions for that transformation to happen. We bought an off-the-shelf system that's designed for non-technical users, not for a developer. We needed to run an intentional campaign to drive adoption with internal champions across the company.

And so the platform needed to be really simple to use. It needed to be able to be used by somebody who is not a technologist, for someone who doesn't need to understand what it means to orchestrate agents or how data actually gets into the agents in the first place. But effectively, it needed to be a tool that someone in their day-to-day could use. And so we bought an off-the-shelf system that's designed for non-technical users, not for a developer, and effectively anyone with a great idea and enough curiosity could actually use it to build a small safe agent that connects securely to our data.

The goal wasn't production stuff. The goal wasn't to get stuff into our application. The goal was that people should start learning by doing at Semble. And boy, did it work. We have something like over 200 agents internally at the moment. Over 103 of them are actually heavily used almost daily.

The second is we needed to run an intentional campaign to drive adoption. That stuff didn't just happen on its own. So what we had to do is we had internal champions. So we had a centralized group across the company who were internally AI champions. It started off at Slack with six people. And now I think there's like 80 people on that channel. All of them aren't champions, but they're asking questions from the champions. There's an AI champion in every single team. So someone who's curious, someone who's interested, someone, again, who doesn't need to understand algorithms, but someone who needs to understand how the tool can help their specific team.

So if it's in marketing, it's maybe to write some content. If it's in legal, it's maybe to summarize the news, the legal news or whatever people do in legal. And it needs to be... So we had champions. We then had easy to follow demos. So we have company-wide meetings every two weeks. Every two weeks, we showcase one agent, the people who built it. We celebrate their achievements. We run short workshops that show simple, real examples. So anyone can come in. It's a bit like office hours. You can come and you can see what other teams are building and how they're using it.

And we encourage people to contact the creators of the specific agents. And we can create really, really simple documentation. We want people's first interaction with an agent to be simple, fast, and rewarding. And the idea here is just like anything else. The faster you pick something up, the more you like it, the more likely you are to go and use it. And then, like anything you do in almost any company, is you need to invest in change management.

People are not going to experiment if they don't feel safe. So what we focused is on designing a space which allowed curiosity to thrive. And we made sure that people were deeply, deeply supported through the whole process. And so what we did is we gave really, really clear guidance. Like we handle medical data, so we can't mess around. But we gave really clear guidance to people, what's safe, what's not safe. The tool allowed us to give access to the right stuff to make sure that the agents that people were building couldn't access random data that we didn't want them to access.

We created sandboxes where people could make as many mistakes as they want. We made it really clear from leadership that it was okay to spend time learning how this stuff worked. And we really encouraged people to share early prototypes, not finished masterpieces. Just show us something as soon as it's ready. Just come and show it to everybody, and we'll celebrate that achievement. We'll celebrate the milestones.

And what we've done basically is we've normalized experimentation. It is normal to experiment at Semble with these technologies. And then the last thing is support has to come from every level of the organization. Now, whenever you go to CTO conferences, there's always a talk about how you manage the co-founders when you're a CTO. I'm lucky in that sense that I was able to convince the co-founder that it made sense to actually invest time in this.

But if it doesn't come from the top, if you're not able to convince whoever is the key person at your organization that this is worth spending time on, it will be very, very difficult. So effectively, when execs say that this matters, people pay attention. When managers say, make time to try this out, people actually tend to try this out. And effectively, when other people start showing you what they've built and how they've improved their specific problem, then people tend to go and try things out.

And so what happened as a result is that understanding the tools led to better product ideas. So when people were using the stuff every single day to solve their own problems, then they started building stuff that we hadn't even... Sorry, they started suggesting stuff to our product team that none of us thought of. So we thought of ways of summarizing letters or summarizing medical notes, the stuff that you read about in the newspaper that everybody thought of. But people like Claire started having really small ideas of how you can show really specific proactive insights to clinicians. So stuff that this particular clinician is going to care about, and you want to surface that at the top because you know that every day they go looking for it. We've automated identifications of sleeping workflows. So when doctors miss the step in the clinical pathway, we can identify that. Smarter triaging. These are all things that none of us in product or none of us in tech have thought of. And they all came from various places in the company that were not the people whose job it was officially to come up with ideas.

And the key here is not, I think probably all of us here have had someone come and say, can AI, you know, make my back hurt less? Because they have no idea what the tool does. But for us, it wasn't blue sky ideas. It was stuff that was grounded in real customer pain that people see every single day, paired with this understanding I've been talking about. And effectively, what ended up happening was that it wasn't just Claire that started producing agents, but started producing ideas. But it came from all over, from all over the company. And effectively, the moment this knowledge is shared across everything, then the company becomes truly intelligent, as opposed to just having a department that owns intelligence.

Now, there's always been a fear. Sorry, I keep on. So there's always been a fear that AI will replace SAS. I don't know how many times I've heard that over the last two years. It's like, oh, SAS is dead. And the truth is that some things are going to change, are going to change radically. If you think about it in a technologist way, what is SAS? It's basically a database with some business layer on the logic on top of it, and then some really thin AI layer for customers to interact with it. And some parts of this stack are probably going to change radically, most likely the UI UX side of it. Like, if you think that there won't be faster ways to interact with AI and filling out forms the way we've been doing for 20 years, I think you're really mistaken. But what we're seeing is that if you pair AI correctly with your existing piece of SAS, then it doesn't replace software, it just reveals new possibilities for it. It's a lot faster to interact with it. And so, how do you spot these abilities the fastest? You do it when the people who understand the customer the best are able to ideate ideas. And so, right now, internal experimentation has become our real R&D engine.

Now, if you want real innovation, you don't start with the engineers. You start with people who own the knowledge, you give them the tools that they need, you support them, you, I now decided not to move forward, to guide them, and you champion them. Because when someone who knows the customer the best and understands what AI can do, they can put two and two together in a much, much better way than you can. So ideas don't come from hierarchy, they don't come from complex roadmaps, they don't come from knowledge, effectively, they come from knowledge meeting capabilities, usually when it's all in the same person. And for us, that's when innovations without limits began at Semble. The ideas that people are generating now are vastly of higher quality than the stuff that they've been producing, they've been suggesting in the past. And so, for us, it's been a huge success.

Because you were working in a data-sensitive environment, right? So, how do you regulate and safeguard your highly regulated data? So at the moment, our highly regulated data cannot be accessed by any AI or any model, right at the moment. And what we're building is, we've actually spent a lot of time, actually, partnering with AWS, they've been invaluable on that front, helping us build a clinical safety engine that's going to sit, or that is now going to start to sit between, effectively, our data stores and our AI agents. And so it's the stuff you'd expect. It's the classic eval platform, but it's also a custom clinical safety module, which actually tests for checks for every single change that we do to any AI. How does that impact the clinical safety of the product? And we have our clinical safety officer that actually writes in plain English what tests he wants the framework to run and to make sure that it's clinically safe. So that's the next frontier for us.

Wow, that's impressive. And just to clarify, for anyone that may have missed it, what platform did you use to do this? It's a platform called Dust, but there's like 50 on the market. But Dust has served us really well. And one of the things I really like is that it uses your credentials to whatever systems that you have within your company. So it has all the security built in natively. So you can connect it to HubSpot, for example. And if you ask information about a customer, it will only give you information about a customer if your user who's logged into Dust also has access to that customer in HubSpot. So it has security quite deeply ingrained from day one. And is that why you chose to go with that particular platform? How did you come up with that decision? Yeah. So for me, the number one criteria was how easy it was to use for non-developers. That was the primary concern. And it had to do that without sacrificing security and safety. So it's actually really, really intuitive to use. And the security is really top-notch. So the way you can segregate data, the way you can give it access to data to some people but not other people, they have a concept of workspace, which really allows fine-tuned segregation.

And the security is really top-notch. So the way you can segregate data, the way you can give it access to data to some people but not other people, they have a concept of workspace, which really allows fine-tuned segregation. So yeah, that was one of the main criteria.

So I'm guessing you had a trial period with non-technical people and see how they got on the platform. Yes. OK, lovely. They were part of the decision-buying process. OK, yeah. That sounds great.

So how did you manage the transition from internal prototype to something that scales to production, that scales to production scale and security? So we didn't reinvent the wheel on that front. So we're an AWS shop. It sounds like I'm advertising for them. But we follow their best practices. They actually lend us engineers to actually make sure it was production ready. We ran a workshop with them, a bootcamp with them over the summer where a bunch of our engineers went actually to their offices for, I think it was a three-week program. And they really helped us a lot to productionalize stuff on the... We use Bedrock. So on AWS Bedrock. So yeah, that's how we did it. Wow, impressive.

OK. So this is just for clarification as well. So someone said that you mentioned your colleague building their own AI engine. Did your colleague do this off their own back? Or how did you get the rest of the people in your company to embrace it and not just by using chatJPT? Yeah. So I feel that embracing stuff is kind of linked to trusting stuff. And it's very much linked to how any startup starts, where you'll have something crazy hard to use that's not connected to anything. And you have a few crazy early adopters that are somehow passionate about it. And they're like, oh my God, I always wished I could understand what this thing is. And effectively, they're your trailblazers. And they improve their lives by a certain percentage, and they won't shut up about it. And then the next person is like, oh, I'm skeptical about this thing, but I work with her every day, and I know her really well. So I'm actually gonna... I want to know what this is about. And skepticism kind of slowly melts one step at a time. If you come in, and maybe we can think of what happened for us on the development side, we actually got it wrong. We came in and we said, cursor is gonna change everything. Here's cursor, we're paying a hundred grand for it. You gotta use it. And we're gonna measure how well you use it. And we're gonna give you training. And of course, that didn't work great. But if we then took this new approach, which we took with the rest, and we started showcasing, okay, this is what I've done with cursor, and this has changed my mind. And if you want to continue using your old way, you can. But this is a measurable improvement it's had on my development process. And that's how we melted skepticism there as well. So it's really skepticism melting by micro joyous moments that you experience with the technology. Micro joyous moments. I really like that as a concept. Michael, we're out of time, but thank you so much for doing this. Thank you. Wonderful segment.