Why AI Agents Fail Without Document Governance | Stéphan Donzé, CEO of AODocs

Hey everyone, welcome back here to Techstrong TV. Our next guest is Stephane Danze. He's the CEO of a company called AODocs.

Don't worry if you never heard of them, we're going to tell you about them. But first, let's hear a little bit about Stephane. Stephane, welcome to Techstrong TV.

How are you? Thank you for having us. I'm great, thanks.

So Stephane, as I mentioned, you're the CEO of AODocs, and we're going to talk about AODocs in just a moment. But give people a flavor for how you got here, what your journey's been like. Sure.

So I'm an engineer initially, and my first professional experience was with a company named Exalead. It was a French search engine, both on the web and on enterprise documents. And so I built the core of the search engine for a few years, from 2000 to 2006, '7, and then I moved to the US to help commercialize it here.

So that's how I got... You can hear from my accent, I'm French originally, but I live in the US since more than 16 years. And then the company was acquired by Dassault Systems, a big PLM company.

Sure. So that's taught me for a few years how to sell business-critical software in America. And it was an interesting experience because I realized that helping companies manage their business-critical information had a lot of value, be it 3D design documents or others.

So I was there, and I wanted to do my own company, but I didn't know exactly what to build. I was very concerned of the risk of having a genius idea that nobody would want to build. So I was kind of running in circle, and I ran into my previous investor, who was still in France and who had started something else in the cloud services company.

And I realized that the cloud was the architecture of the future, it was something interesting to build, and that he with this services company could have access to customers who would have problems, who would generate an idea. So by grounding my nascent business in ideas coming from companies, I was like, okay, there is a chance to build a product market fit since the idea comes not from me, but from customers. And very quickly, those customers of this service company were saying that they wanted to put not only their email and calendar and so on in the cloud, but also their business documents.

And there was no document management system in the cloud. All of the document management markets till today is old on-prem technology that was not designed for the cloud. So there was all the ingredients, a business problem, something business critical, document management, new technology in the cloud.

There was all ingredients to build something new, and there was a population of early adopters, the companies who had already chosen to put their email and collaboration files in the cloud. So we took all of this, and I started AODocs associated with this service company, so that we could bootstrap the whole. And to this day, AODocs has never raised money.

We have built a company entirely on bootstrap revenue until AODocs was profitable in 2022. So it's a very uncommon story of a startup. It's the old-fashioned way, right?

You earned it. Yeah, exactly. Yeah.

But it gave us some, let's say, a down-to-earth approach that I think our customers appreciate. They appreciate that we don't BS them, so to speak. We always down-to-earth, including on AI, what works, what doesn't work.

We're very reliable and very trustable people. And for the business we do, taking care of the business critical documents of our customers, it's very important to build trust. We work with companies like Airbus, like Google for their data center construction plans, with Veolia for building water treatment plants.

So the documents we manage for our customers are the most important documents of the company. We help them ensure that they use the right version of the right document. We help them put these documents in process, put traceability everywhere, and of course, add AI on top of it in a reliable manner, which I think is in line with your next questions.

Excellent. I love it. Hey, just quickly, the website?

com, A-O-D-O-C-S. And if you wonder what it means, Ao is the Polynesian god of the daylight, and by extension, it's the god of the cloud. So docs like documents, Ao like the god of the clouds, documents in the clouds.

Well, I meant. Love it. Excellent.

That's one of the nice things about interviewing founders and stuff like this, and CEOs, because you get these little backstories, where the name come from and so forth. All right. When you started the company, you didn't think you'd be out here talking about AI and its effect, though, right?

That's right, yeah. Not this quickly, anyway. But it's changing everything.

I just got back. The whole plane ride home, I'm playing in my own agentic AI, playing, working in my own agentic AI- Yep ... platform, doing things.

What has this meant for AODocs? Yeah. I think there's been two big moments, right?

And people don't realize that-- Everybody knows about the ChatGPT moment. Okay, 2022, all of a sudden, we can chat with something that looks intelligent, and all of a sudden, there's a machine that understands text. So for us, it was a first revolution, and frankly, that's something I've been waiting for all my career, because even back in my search days in Exalead, we were trying to get some semantic from the text, right?

It was understanding the difference between, orange the color and Orange the company and all of that kind of stuff. So all of a sudden, AI was able to understand a text, summarize a text, and it changed-- It was a first revolution for us as a document management, because instead of asking humans to tag documents manually, to put it in the right folder and in a huge amount of manual work, which was forever of the biggest hurdle of adoption of our kind of product. All of a sudden, we can tell people, "Hey, just put the document there.

" So first revolution, the ChatGPT moment, AI computers are able to understand text and do something with it instead of only work on structured data. But then there's a second revolution, and I don't think people realize that it's as big as the first one. 6, and Claude.

It's OpenAI Codex. It's the ability of agents to really start doing things for a long period of time. Before the end of last year, an AI agent was able to do a single task.

Do this. Ten seconds later, you have to give another instruction. Ten seconds later, another instruction.

And now they've reached a point where you can give an agent a mission. " And it works alone for 10 minutes. " And it works for two hours.

And you come back and it's done. It's not perfect. It's still junior-level work.

But you can give missions to an agent to work for a long period of time. So it changes really everything, because now you can treat agents like end users. And that's what's provoked the SaaSpocalypse issue with the valuation of SaaS software, right?

The actions that are usually done by humans can be delegated to agents. So maybe now I need 10 users instead of 100 users to do a certain operation in a certain software. For us, I think it's beneficial in the sense that those agents, to work correctly, they need a solid foundation.

They need to work on the right documents. They need to be able to put their work with traceability somewhere. So we provide, let's say, the concrete slab on top of which you build your things.

If you write code, a business application, and you let the agent decide itself what is the storage, how the documents are managed and so on, you're in for a lot of trouble because you're asking a junior developer to make all of the hard choices regarding how you manipulate business-critical documents. But on the other hand, if you ask an AI agent to build an application, let's say, to manage your maintenance work request in your hospital, right? Something that you shouldn't mess up with.

But you ask the agent to put everything in a specific document management workflow in this AODocs system that is reliable and so on, you know the documents will be taken care of, then your agent can focus on the user interface, on the mobile app, on all of the things that matter to your business users while you protect the company by putting the data somewhere solid. So we enable the use of vibe coding business applications. We enable the autonomous agents that will run in your company do stuff because you know that they will not mess up with the data as long as they are putting it in a safe, centralized system.

And you can have 20 different agents, 200 different agents, and vibe-coded applications all relying on the same centralized system. You don't have this, you fragment everything into 200 different databases with lots of poetry and improvisation on how the data is handled. But if you put those 200 applications on a single solid foundation, now you have something that can scale in number of applications and remain compliant and traceable and all the good stuff that you need for your business documents.

So we position ourselves in this era of vibe coding and autonomous agents as the foundation on which we can build things that manipulate business-critical documents. I got it. But, I can't help, I'm listening to you, Stephan.

So, they talk about replacing all the junior people. Yep. Right?

And I understand why one would say that. But until our AIs canHave a little more common sense. You know what I mean?

Yes. Can distinguish between certain things, because even a junior person can do that. And so this goes to something...

And as a CEO myself, much like you, I'm learning, we're all learning as we're going here, right? We're making it up. 100%.

We're seeing what works. Day by day, week by week, things change. But I still believe in my heart, whether you're a junior or a senior worker, if you embrace this technology, but don't think it's replacing you or just let it run amok or run wild, but if you channel it, use it correctly, you 10X yourself.

You can 10X yourself and- Yeah ... make you, I don't care whether you're a junior, senior, or post senior, you make yourself much more valuable in the market to your employer, to your next employer, and everything else, right? Yeah, absolutely.

I think the multiplier depends on your seniority. Juniors will 2X themselves. Yes.

Senior can 10X themselves, but it doesn't matter. It's a new skill to learn, like I said. To me, the first step, and I had that conversation with some junior staff in our company before asking the AI to produce stuff, use it to review your stuff.

You build something, you ask AI to review, give you ideas of improvements and so on, and then it speeds up, it increases the quality of your work. And then once you start to understand how the AI, where the AI is good, where it fails, then you can upgrade little by little and make it do stuff. But it's a learning curve.

And I think there is a shock right now because everybody is discovering the technology and adjusting their hiring plans based on that. But once we pass the shock, we get into a new normal where, okay, onboarding a junior is no longer let them do the boring work to learn the trade. Is that you take them directly to more advanced work and learn to use the AI as a sidekick.

But if you stop hiring juniors, who are going to be your seniors in five years, right? So I think there's a temporary turbulence. There are position that disappears.

It's a fact. You don't need people to translate text anymore, so that kind of specific tasks will disappear. But the new equilibrium between humans and AI will settle, and then it will be the new normal.

But right now, it's very turbulent because everything is changing so fast. Like I said, in the last six months, it's a complete revolution. And I'm very surprised talking to people.

I was at a conference last week in Chicago. Most of the people in the room, meaning CIOs and IT leaders, were not aware of the fact that everything changed before between, let's say, November and February this year. They still- Absolutely ...

have integrated. I agree with you. So as I said, I was at RSA C, Security Conference.

I actually gave a talk, Mitchell Ashley and I, about developing and security, and it's to this exact point. The whole AppSec, application security market, was based on finding bugs. We scanned, we tested, we fuzzed, and we found bugs, and we gave you bugs and say, "Fix the code.

" So the emphasis was on finding those bugs. Now with Claude Opus, Claude Security, and not just Claude, they all. With AI, we can find so many bugs.

Not all of them are critical, but we find so many bugs that it overwhelms you. Yep. So the cheese has moved.

The emphasis is no longer on finding bugs. It's on governance of code. The bottleneck used to be an average developer, what did he make?

A couple of dozen, maybe a couple of hundred lines of code a day at most. Oh, sure. A really good.

Yeah. Now you got these machines that are turning out thousands of lines of code a day. What does that mean for the world of security?

We couldn't secure the amount of code coming out before. How are we possibly going to secure the amount of code that we're turning out now? Not without AI, not without autonomy and autonomous scalability.

Yep. And I think that's exactly what you're talking about as well. Governance is the new- Yeah ...

bottleneck. I think there's an aspect, and going back to what you were saying about common sense, an aspect we didn't cover is using AI to find information, right? " Right?

No, because AI assumes that everything you give it is- Is equal ... reliable. Is equal, exactly.

Yeah. And AI is very good at semantic matching. I have a question, I will find all the documents that relate to that question, contain a potential answer to that question.

The N minus five version of the maintenance manual of your industrial robot, and everything has changed because now it's not this button first and this button second, it has reversed. And if you don't follow the right procedure, boom, or you lose your arm or whatever. So AI is enabled to distinguish between something obsolete, something not validated, something bluntly incorrect.

It just assumes that everything you give it is correct. And everything I talk about this, I ask the room, "Okay. " Nobody raised a hand.

" Two, three hands in the room. It's 90+ percent temporary files or whatever. So if you ask critical question, if you want reliable answer, your banker answering about an interest rate, your maintenance technician asking about what should I do with this or that error code, right?

If you are asking questions where you cannot afford to make a mistake, you need to make sure that your AI is working only with the golden copy, the correct version of the documents, and for this, you need governance. AI is not smart enough to find the right document. Why?

Because the information is not there. xls. Which one is the correct one?

If you don't have the tag, if you don't have the metadata that tells you this is the price applicable today because the end of your discount is over and we're in such and such region and so on and so forth, all of this contextual information is not in the file somewhere. You need to give it to the chatbot otherwise it will pick one randomly because all five spreadsheets have the same semantic score in terms of relevance to your question. Right.

I noticed this and I'd say, pick the best. It has very hard... Like one of the...

You know how it always suggests. This is what's the best. It doesn't- How do I pick the nice-- Right.

Does it have a criteria? It tells you it picked-- Well, that's a whole nother discussion we could have. Exactly.

That's where the common sense comes into question. Yeah. As a human person, if you see five files that have exactly the same name, you're like, "Whoa, danger.

" It's like, "Cool. " Who wins? Not the AI.

So I wrote an article. When I was a kid, there used to be a Starkist tuna commercial, right? Charlie the Tuna was always trying to get picked by Starkist by showing them he had good taste.

Yeah. " And AI, that's the problem we have here. High taste, yeah.

We want AI that tastes good, not good taste, but we need it to learn good taste. Stephane, we're over time. I got to end this one up.

But great conversation. Appreciate it. Good luck with AODocs.

Come back, keep us posted on what's going on. Yeah, of course. Thank you for having me.

Bye. All right, thank you. It's our pleasure.

Stephane Danto, CEO, AODocs, here on Techstrong TV. We're going to take a break. We'll be back.