In this episode of Software Spotlight, host Michael Bernzweig is joined by Maxime Vermeir, VP of AI Strategy at ABBYY. They discuss the transformative power of Intelligent Document Processing in the context of AI, exploring how ABBYY's journey from OCR technology to AI-driven enterprise solutions has evolved. Maxime shares insights from his own career, including lessons learned from his startup experience and how it informs his current role at ABBYY. Listeners will gain an understanding of why structured data is crucial for AI success and how businesses can effectively harness AI for their processes.
Maxime Vermeir is the Vice President of AI Strategy at ABBYY, where he focuses on integrating AI into enterprise solutions and guiding the company's technological innovations. With a background in networking infrastructure and experience founding a startup in low-code, no-code app development, Maxime brings a unique perspective to the challenges and opportunities in AI.
[MICHAEL]: You know, it's not every day you get to hear about a journey from networking infrastructure to leading AI strategy at ABBYY. Maxime Vermeir has seen it all, from founding a startup on the cutting edge of low-code, no-code app development to transforming the way organizations handle documents with AI. At our Software Oasis AI Summit, Maxime shared how ABBYY's evolution in AI, starting with OCR in the 70s, is now revolutionizing enterprise solutions. By the end of this episode, you'll understand exactly why Intelligent Document Processing is the key to making AI work for your business. I am your host Michael Bernzweig and this is Software Spotlight.
--- INTERVIEW ---
[MAXIME]: Max, I know a lot of organizations and individuals that are listening have arrived at where they are after many years.
[MICHAEL]: Some are working their way up through the ranks in an organization, others have founded their organizations, but can you share a little bit about your journey prior to joining Abbey and then bring us to the current day?
[MICHAEL]: Yeah, absolutely.
[MAXIME]: I'd be happy to, Michael.
[MAXIME]: It's been quite a journey, if I say so myself.
[MAXIME]: I started out studying specifically networking infrastructure, so if I would have gone on to the trajectory that my studies put me on, you wouldn't find me in a data center right now and that would be my background, not this office that I'm currently in.
[MAXIME]: From that, I actually went into consulting, specifically around enterprise content management.
[MAXIME]: Actually, that was a really interesting time.
[MAXIME]: I'll tell you why, because in that time, I learned so much in such a short amount of time because I was able to interact with a lot of different types of organizations from small-medium businesses to large pharmaceutical enterprises, these big third-party outsourcers, VPOs, all of them I came in touch with.
[MAXIME]: I definitely saw that they were all trying to solve the same type of problem just at a different scale, but it was that scaling component that made it so interesting to work with all of these people.
[MAXIME]: At the same time, even as I was going through that consulting phase where I actually learned a lot, I still had this entrepreneurial vibe in me that wanted to do something as well.
[MAXIME]: Simultaneously, I actually founded a startup that was a bit too early to market.
[MAXIME]: I'll admit that up front before I even tell the story.
[MAXIME]: The reason was too early.
[MAXIME]: We actually had created, as an organization, the very first no-code, low-code app development platform.
[MAXIME]: You could go in, you could drag and drop all these components.
[MAXIME]: This was shortly after the App Store and everything low-code, no-code had not happened yet, but we saw that that was where things were going.
[MAXIME]: We wanted to make sure that we got there and actually made it possible because back in the day, building an app was something that was a humongous undertaking.
[MAXIME]: It cost half a million dollars easily if you wanted to have any level of capability because it was so new and it wasn't commoditized yet.
[MAXIME]: We built that, but like I said, we were a bit too early to market.
[MAXIME]: The whole low-code, no-code craze started about five to eight years later.
[MAXIME]: But still, it was an extremely valuable learning experience for me to have done that, to have that entrepreneurial kind of experience as well in combination with consulting that still to this day lets me help all of my customers, everybody I talk to in a different fashion where I'm not just trying to figure out one particular problem with them.
[MAXIME]: I'm always seeing it at a bigger picture and how these decisions that they make could definitely impact at a larger scale the future that they're going to have because that's what I've learned to see in kind of all of those experiences over the years in doing both those things.
[MAXIME]: So after that, I ended up at Abbey where I've also had quite a long journey so far.
[MAXIME]: I've been with the company close to 10 years and it's been extremely exciting to be on this side of artificial intelligence for so long and seeing how it has evolved.
[MAXIME]: And again, also there, my first role was actually in pre-sales where I applied my knowledge from consulting and being able to work with customers directly to find the right Abbey solution for them.
[MAXIME]: And then I've evolved into working with our innovation team and actually starting an innovation team that works with the largest customers, figuring out where we could take them next and what our technology needed to do to be able to take them to that next place.
[MAXIME]: And then shortly after, I was sent in product management as well.
[MAXIME]: And then of course, now I run our AI strategy, which is both working with our customers and helping them on that path and at the same time looking inwards as well, where do our products need to go and what is that input that we need from our product teams in terms of what the next big thing is that we need to develop to stay in our leadership position.
[MAXIME]: And you know, it's interesting, we just had a guest recently on the podcast, Josh Haas over at Bubble and it sounds like during that same era where it sounds like the solution that you were building or were involved with was providing the building blocks for organizations to put together their own solutions.
[MAXIME]: You know, they had the same origin story, you know, really helping organizations that didn't want to get into fully coded solutions to come up with an end-to-end solution.
[MAXIME]: So let me ask you, obviously, ABI has been on the scene for decades and, you know, the origins in OCR and intelligent document processing is the core foundation of ABI, but obviously now, you know, AI and AI-driven enterprise solutions are, it sounds like the future.
[MICHAEL]: Can you kind of walk us through that journey from the origins to where ABI is headed?
[MICHAEL]: For sure.
[MAXIME]: And it's interesting that a lot of people view it that way, but, and this is also just a common misconception that I've recently dubbed it that the human context window to put it in today's technical terms is too small to remember everything.
[MICHAEL]: So, you know, when you think about artificial intelligence, the right now that gets labeled as everything that's known in the mainstream kind of, you know, thought process, right?
[MICHAEL]: So that's LMs, that's chat GPT, CLAWD, everything, vibe coding, but artificial intelligence started so much earlier.
[MAXIME]: Our very first OCR engine back in the seventies was already based on artificial intelligence.
[MICHAEL]: We use machine learning techniques such as convolutional neural networks, which also fall into that category, right?
[MICHAEL]: So it's been an interesting evolution for the organization to actually start out with dictionaries.
[MAXIME]: You know, if you want to know the true founder story of ABI, it's one of our founders was studying and he needed to learn a new language, but there were no dictionaries available online or actually I say online, but actually what I mean is digitally, they were not available.
[MICHAEL]: So what do you need to do to actually build that?
[MICHAEL]: Well, he needed to build an OCR engine because he wanted to build a digital dictionary.
[MICHAEL]: So the dictionary was the thing that he wanted to achieve.
[MICHAEL]: The OCR was actually a by-product, funny enough, but it actually turned into the mission of the company.
[MICHAEL]: Yeah.
[MAXIME]: And so first couple of years of ABI, you know, our OCR engine was always the one that was looked at for accuracy.
[MAXIME]: A lot of other companies have our technology embedded inside of their products, funny enough.
[MAXIME]: And shortly, you know, kind of as we evolved and there was more demand for this type of technology, especially, you know, late nineties, our organization decided that we can offer this to enterprise customers as well.
[MAXIME]: And we wanted to really show the market that if you put your mind to it, you know, the way that we wanted to deliver it, we could actually do that in a huge amount of scale.
[MAXIME]: And there's so many solutions to this day that you've probably already touched when you interact with an organization that are being powered by ABI behind the scenes.
[MAXIME]: So fast forward to today, you know, not to make that, you know, those 35, 40 years make it sound too long.
[MAXIME]: This is definitely something that, you know, right now we're seeing that our technology is at the core of any kind of AI or agendic initiative that organizations want to go leverage.
[MAXIME]: You know, that structured data that is required to make sure that you don't get the right decisions happening and also the governance around it with our processing or process intelligence capabilities.
[MAXIME]: That is just something that's table six at this point in time.
[MAXIME]: So it's been really interesting to see over just the last four years, I would say, to go from being called a commodity to now being the most important part in any kind of AI pipeline that you can imagine, because everybody suddenly realized, again, that businesses run on documents.
[MAXIME]: And so that's where everybody has to start.
[MICHAEL]: Now, for individuals that are listening and saying to themselves, how does this all fit in with our organization?
[MICHAEL]: Can you maybe take a step back and share a little bit about your role within ABI and how that touches stakeholders internally and externally?
[MICHAEL]: Yeah, for sure.
[MAXIME]: So for me, one of the main components of my role is to make sure that we're able to communicate our value that we deliver to our customers extremely clearly.
[MAXIME]: And that is through having the ability to kind of showcase our technology in a way that it immediately makes sense.
[MAXIME]: Right.
[MAXIME]: That's also why I use invoices as an example in my presentation earlier, simply because it's something that we can all resonate with.
[MAXIME]: You know, I think the messiness of documents, the chaos of documents that we've all dealt with is something that everybody can relate to.
[MAXIME]: And the way that we then go about that is we turn that complexity into clarity that you can then use downstream.
[MAXIME]: And I think that's the simplest way I can put it.
[MAXIME]: You know, so you have something incredibly complex, but it's vital to your organization because that's where your knowledge is.
[MAXIME]: And you need somebody that actually turns that into something that's clear, understandable.
[MAXIME]: Well, it's that kind of clarity that ABI provides with its technology.
[MAXIME]: Now, I know you've as an organization recently just come out of your latest ABI Ascent, you know, the whole cycle and a lot of exciting new announcements and all of that.
[MAXIME]: For sure.
[MICHAEL]: For, you know, obviously, besides the clear black and white solutions and, you know, features that were announced, can you talk a little bit about the story behind some of those launches and things that customers may not see?
[MICHAEL]: So apparently on the front end as to why some of those those changes have happened.
[MICHAEL]: Yeah, for sure.
[MAXIME]: So I think for me, one of the most important things is that we continue to enable more and more use cases.
[MAXIME]: Right.
[MAXIME]: And sometimes you see this a lot where technology is just added for the sake of technology.
[MAXIME]: But we tried to not do that.
[MAXIME]: You know, every single choice that we make is quite deliberate in terms of what we add to our roadmap and why we add it.
[MAXIME]: And we definitely want to expand our ecosystem as well.
[MAXIME]: So putting all that together, some of the things that have happened behind the scenes is to make sure that we can accelerate faster and faster with our customers, but also with the market in terms of keeping up with the changes that are happening and continuing to be that kind of like that guiding, you know, I'm going to say the word guiding Sherpa as everybody's trying to climb the AI mountain.
[MICHAEL]: But then, you know, in doing it in such a way that it fits with our mantra, which is purpose-built AI, right?
[MAXIME]: You know, if I think back a couple of years and it's a great example of making a decision that actually goes against the typical stream where we decided that we're not going to add large language models immediately into our platforms because we wanted to figure out how we could make them work productively first and not have the side effects.
[MAXIME]: So that's something that we've announced at our Ascend.
[MAXIME]: There's a couple of other announcements that were extremely important, which is about defining the structure finally for what a document looks like in terms of AI, something that we're very happily working with together with the Linux Foundation.
[MAXIME]: So a lot of exciting things that are going to happen in the second half of the year based off of all of these starting points that we've now put into market.
[MAXIME]: And I'm extremely, extremely excited about what we're going to be able to do once the AI community starts to like bring everything more and more together.
[MAXIME]: Everything becomes more and more integrated and people start to realize what is possible.
[MAXIME]: Once you have those foundations, then things are going to get really cool and really interesting very fast.
[MAXIME]: Yeah, and it's interesting, you know, you look back at early machine model models, machine learning really is is the foundation of what we're calling AI today.
[MAXIME]: Obviously, it's expanded quite a bit to incorporate, you know, all kinds of additional data sets that, you know, may be external versus internal in the early days.
[MICHAEL]: But when you look at enterprise grade AI models and tech, you know, can you share a little bit about what that means to Abby?
[MICHAEL]: What are some of the components of, you know, putting out a solution that is truly enterprise grade AI?
[MAXIME]: Yeah, and so I think what we're seeing right now in the market is this rapid evolution of technology.
[MAXIME]: But what is often overlooked in all of that hype is actually that enterprises require so much more than just the technology itself.
[MAXIME]: I'm talking about governance, compliance, you know, the ability to provide auditability, the audit trails of how something was built, compliance with the AI acts both in the US and Europe and other regions.
[MICHAEL]: Data sovereignty is definitely becoming a thing right now where organizations are being asked both by other organizations, governments, their customers, OK, where does my data reside?
[MICHAEL]: But also, where did the data come from that is being used by the model or that was used to train the model that you're using to power some of these capabilities?
[MICHAEL]: Can you ensure to me that the decisions that you're making automatically are actually something that will not have a negative impact on my life?
[MICHAEL]: You know, to give an example, not too long ago, I think it was one of the first couple of years that Chappachee PT was available, a friend of mine actually went in and asked about himself like we all did back in the day, right?
[MAXIME]: We wanted to know what it knew about us.
[MAXIME]: Unfortunately, after a kind of a very accurate but somewhat long answer, it said, and you are deceased.
[MAXIME]: That was literally its answer, which was entirely incorrect.
[MAXIME]: And I see you smiling, Michael.
[MAXIME]: But if you think about it, if that same output would have been used to make a credit decision or a decision on whether or not he can actually be onboarded into a bank for a loan to buy a new house, to lease something, the answer would have been no.
[MICHAEL]: And try to prove into those organizations how that actually happened, right?
[MICHAEL]: So it's those kind of things that people at face value are like, well, you know, it's a mistake.
[MICHAEL]: It's quite funny how they didn't.
[MICHAEL]: It shouldn't have said that, but it can have very adverse effects.
[MICHAEL]: And those things still exist in today's models because the transparency, except for a very few, is not up to the level yet that they need to be so that it can actually be used in these types of high impact processes that organizations are looking to automate.
[MAXIME]: So I think that for me really embodies enterprise grade, having all of those boxes checked so that when our technology is used, our customers can use it with confidence and they don't have to worry about those kind of things.
[MICHAEL]: So when you're building around that, you know, as your design goals, I guess we could call it, are there some tradeoffs that need to be made in order to be able to achieve, you know, the predictability, the compliance, the governance, all of that?
[MICHAEL]: Not so much tradeoffs.
[MAXIME]: I think there's there's a lot more work that goes into it.
[MAXIME]: You know, compliance, governance is mostly about documentation, being able to prove how something was built, where the data came from.
[MAXIME]: We actually have the benefit as an organization that we have already existed for over 35 years.
[MAXIME]: So our data set of documents that we've seen that has the messiest kind of things that you could imagine in terms of documents makes us very well specialized.
[MAXIME]: And we've been training on that data set that has been freely shared, that we have been able to accumulate with permission of our customers and everybody that's worked with us over the years.
[MAXIME]: And we've carefully curated that data set into something that we now can confidently say, this is exactly where it came from, this is lineage and this is how it's being used to train.
[MAXIME]: In terms of tradeoffs, I think the only one is, is that it takes a little bit longer to bring something to market.
[MICHAEL]: But if you think about the typical cycle that, you know, definitely large enterprises go through, they don't change their technology overnight, right?
[MICHAEL]: This is something that happens in the term time span of several months at minimum, if not years in some cases where before true change is enacted, POCs, MVPs, all those things that happens fast.
[MICHAEL]: But oftentimes they highlight it, especially now that this is something that customers are looking for, enterprises are looking for.
[MICHAEL]: And so then they're happy to kind of wait for us to deliver on that because they know that it's going to tick all the boxes, not just from the technology side, but also from that enterprise weight level that they really, really require.
[MICHAEL]: Otherwise, they simply cannot go into production.
[MICHAEL]: Now, having worked with so many organizations in the space over the years and especially in this latest transition, are there some statistics or numbers or data that you think most enterprises would be surprised to understand related to this area of AI that you think are eye-opening?
[MAXIME]: Yeah, I think there's two that immediately come to mind, one that is always an eye opener.
[MAXIME]: You know, I recently presented in Dublin to an audience that was very unfamiliar with what we do as an organization.
[MAXIME]: And I showed them the 90 percent statistic, 90 percent of your business processes run on documents.
[MICHAEL]: And I literally asked the audience, you know, is there anybody that can name a single one?
[MICHAEL]: You know, you don't have to give me 10 out of 100, just give me one that actually does not include a document.
[MICHAEL]: And nobody raised their hand, either they were scared or they just didn't have an answer.
[MICHAEL]: So, but that statistic, you know, it's not, once you hear it, you're like, yeah, that actually makes sense.
[MICHAEL]: But it's something that is not typically top of mind for everybody, because for us, we are so used to working with these documents as a medium, right?
[MAXIME]: It's what we do every single day.
[MAXIME]: So you don't think about it that way.
[MAXIME]: So that's definitely one.
[MAXIME]: And I think that the other one is, it's not just one statistic, but it's something that I think is definitely worth for a lot of organizations to go look at, is the transparency index of the standard university when it comes to large language models.
[MAXIME]: What I think that very easily captures is the different level of transparency across multiple facets that is available and documented for the most commonly used ones.
[MAXIME]: Right.
[MAXIME]: I'm not talking about an obscure open source model somewhere or, you know, you name it.
[MAXIME]: But no, the really big ones from coming from the big tech giants, that's where you see that the transparency level is not exactly where you would have thought it would be at at this point in time, especially because there's already so much emphasis on these types of things.
[MAXIME]: So those would be my two recommendations.
[MAXIME]: And one to keep in mind, there's always documents involved in your process and keep that one and kind of feel your back of your mind that when you're thinking about using a model and you want to be able to deploy it to production, that's something that you do want to check, that it ticks all the boxes that you're going to need down the line, not from a technology perspective, but from a compliance and governance perspective.
[MICHAEL]: And if you were to break down, you know, some of the major goals in terms of getting towards where the marketplace is heading as an organization, what are some of the major major opportunities in the years to come for AVI?
[MAXIME]: So I think for us in the years to come, because we have this incredible DNA of both understanding documents and processes, so intelligent document processing, which I talked about, and process intelligence combined, I see us being able to really provide almost like the the full blueprint of an organization's process.
[MICHAEL]: And in turn, once you know the blueprint, you can turn that easily into something that is then executed, right?
[MICHAEL]: Whether that involves multiple systems or something else, because this is the thing that a lot of systems struggle with and why everybody's so excited about agentic AI is because it allows us to go discover how to work the process.
[MAXIME]: But if we can already create a blueprint of what the process is and how it should be optimized, both from document data, system data and combine all of that, that makes for a very interesting shortcut to get to that path of success where a process can simply become automated.
[MAXIME]: And the thing is, and this is why I'm going to refer to the 90% statistic again, business happens inside of these documents.
[MAXIME]: So you can have the most wonderful RPA, iPass, agentic AI or agent running as long as they can't see what actually is being put into your PDFs, your TIFF files, your Excel sheets, your Word files.
[MAXIME]: They can't make all the decisions.
[MAXIME]: There's always going to have to be a person.
[MAXIME]: And lo and behold, if you actually went and went into the whole six-segment thing, right, where you were working with the process, asking people questions about how the process works in the organization.
[MAXIME]: Even if you do that, we've seen this in numbers as well.
[MAXIME]: And then you start looking at the gaps in the data.
[MAXIME]: You've quickly uncovered that there's one person in a corner somewhere that knows everything about the process that everybody else does not know.
[MAXIME]: And as soon as you take that person out, everything stops working.
[MAXIME]: There is that in every single organization.
[MAXIME]: And so I think because we have this visibility level to also kind of see that because we can see what happens in the documents, we can translate that into event data.
[MICHAEL]: What happened?
[MICHAEL]: What was the data before?
[MICHAEL]: What was the data after?
[MAXIME]: That gives us a very unique position of where we could take ourselves next.
[MAXIME]: So I don't want to give away too much because I'm not going to divulge our full 10-year roadmap to you, Michael.
[MAXIME]: But I hope that kind of gives you a little bit of a tip of the iceberg there.
[MICHAEL]: And as we're wrapping up a couple of final thoughts, is there, and I don't know if you want to mention a specific organization or more of just a use case, but a specific project or engagement that the team has worked on that has been really aided by AI and just amazing in terms of the final results achieved?
[MICHAEL]: Oh, there's a lot.
[MICHAEL]: So now you're asking me to pick favorites, which is always the hard thing to do.
[MICHAEL]: So one, because it's close to home and it kind of really shows that technology can help an organization persevere through regulatory changes.
[MICHAEL]: In this particular instance, that change was Brexit.
[MAXIME]: And it was one of our partner's customers that had to quickly transport fast-moving consumer goods.
[MAXIME]: So really, really delicious pastries that they needed to move from Ireland over into Europe and the UK as well.
[MAXIME]: And so, but they were getting stuck basically due to customs.
[MAXIME]: So they were sitting there for days.
[MAXIME]: Of course, you know, that's not something that you want to happen to food, obviously.
[MAXIME]: And we were able to leverage our technology to actually bring that down to five minutes.
[MAXIME]: You know, the whole process was transformed completely.
[MAXIME]: And now everything in terms of custom clearance happens in five single minutes where it used to be days before, especially post-Brexit.
[MAXIME]: So that is how, you know, like I said, technology can help you persevere as the world keeps changing, as the market keep changing and as you can keep that competitive advantage.
[MAXIME]: Yeah, that's amazing.
[MICHAEL]: And final thoughts as we're wrapping up here, is there from where you sit and what you say, is there something that's really exciting to you in the space that you're on the edge of your seat watching to see what happens?
[MICHAEL]: I'm watching to see what happens when more worlds start to collide, right?
[MAXIME]: So we have everything in terms of agents happening right now.
[MAXIME]: We also see a lot of push in terms of robotics, you know, what's possible now.
[MAXIME]: And then, you know, something that is bubbling up again, even though it's also been already in research for many, many years, but there have been lots of interesting advances and, of course, quantum computing.
[MAXIME]: So I see it as if we can make robots that are as nimble as we are, we give them the tools to actually take action and make decisions based on this kind of agentic framework.
[MAXIME]: And then we have the computing power, which starts to resemble a Star Trek computer inside of, you know, the Enterprise Class-D, then, you know, I like to imagine that we're going to live in a sci-fi world pretty soon.
[MAXIME]: And I'm actually quite sure that's not going to be a bad thing.
[MAXIME]: I love it.
[MAXIME]: I love it.
[MAXIME]: Well, on that note, I really appreciate your joining us on the podcast this week and very much looking forward to your presentation at the live event.
[MAXIME]: Awesome.
[MICHAEL]: Maxime just gave us a fascinating look into his journey from studying networking to leading AI strategy at ABBYY. Now, he's about to walk us through how ABBYY's Intelligent Document Processing connects to AI, and why this evolution matters for businesses today. This ties back to our core insight about transforming unstructured data into actionable insights.
--- PRESENTATION ---
[MAXIME]: It's a pleasure to be here.
[MAXIME]: I was actually talking to a large European organization that is in trade and logistics not too long ago.
[MAXIME]: They were telling me about what seemed to them an extremely novel idea, which was to modernize the way that they work with documents within their organization specifically for the purpose of actually enabling artificial intelligence.
[MAXIME]: What was their idea actually that they wanted to do?
[MAXIME]: Well, they wanted to create an LLM pipeline that took their invoices and I'm using invoice as an example, but it's something that everybody can relate to.
[MAXIME]: But this, in this particular instance, was actually more of a complex document.
[MAXIME]: But anyway, we'll stick with invoices.
[MAXIME]: During the pilot, they pushed everything through.
[MAXIME]: They were getting all the data out of their documents that they were looking for.
[MAXIME]: The total amounts which vendor had actually sent this invoice.
[MAXIME]: They were seeing really great results.
[MAXIME]: They were actually seeing about 95 percent accuracy, and accuracy in this particular instance means that 95 percent of these documents actually went through from input all the way to the ERP system without anybody having to look at them, and when they were using a large language model to do it.
[MAXIME]: Now, after about two weeks, reality started to kick in.
[MAXIME]: Production meant that they suddenly had to deal with a lot of more complexity than they actually saw during a pilot.
[MAXIME]: Because I'm pretty sure that we can all agree, documents are incredibly chaotic.
[MAXIME]: Payment terms, data points, actual tables that span across multiple pages, overlapping columns, actually items of data that you want to get from the descriptions inside of a column, inside of a particular row.
[MAXIME]: You can't imagine how complex it can actually get.
[MAXIME]: I haven't even mentioned barcodes and everything else that comes with it.
[MAXIME]: So what actually was happening was as soon as production hit, and the chaos started to increase, they were seeing a lot less results than they were actually hoping for.
[MAXIME]: About two weeks later, only 12 percent of invoices were being posted incorrectly, actually posted incorrectly, and it was hallucinating payment terms.
[MAXIME]: We had inconsistent table extraction, vending mismatches, and the most important thing of all, as they were going through that process, is that their auditing team started to ask the question, can you tell me how we got to this result?
[MAXIME]: Because this is something that we have to fully understand and document for all of the regulations and governance that are coming into place.
[MAXIME]: Is this the thing that we're going to be looking for?
[MAXIME]: Now, if you think about that in the context of what else is happening in the world, where else the market is heading, and what we are experimenting with, and what I'm referring to is actually Agenta KR.
[MAXIME]: If you listen to Gartmell, for example, they say that about 42 percent of organizations are going to be deploying Agenta KR in the next 12 months.
[MAXIME]: Agenta KR makes decisions based off the data that you feed it.
[MAXIME]: It interprets it, and actually then makes a decision of what the best next step is.
[MAXIME]: Problem is, like we've seen with that example that I just talked through, 75 percent of all of our organizational data is completely unstructured.
[MAXIME]: Not just unstructured, because that might sound like something that an agent, a large language model could actually work with quite well, but it's also not structured in a way that they can actually confidently work with it.
[MAXIME]: From our own research, and this was validated as well, 90 percent of business processes run on documents.
[MAXIME]: I would love for you to name me a business process where there is not a single document involved.
[MAXIME]: That's simply because we as humans are actually running the business.
[MAXIME]: This is how we communicate with each other between different organizations and even between different departments.
[MAXIME]: I can't help but conclude as I was thinking about seeing this example, how the pilot went great, but then when it actually got to the point where production needed to happen, all of that sound, that ground-truth data was not available, and it turned into something that was less than what they had desired.
[MAXIME]: Fast-forward that into exactly what a lot of businesses are now looking for with Agenta KI to take action on that ground-truth data, but they don't have it.
[MAXIME]: What does that actually equate to?
[MAXIME]: Well, like this.
[MAXIME]: Agenta KI on bad document data is like putting FlexTape on a broken pipe.
[MAXIME]: It's a bit of a fix, but it's not going to hold for very long.
[MAXIME]: The thing that everybody's struggling with is, how do LLMs actually work inside of this environment?
[MAXIME]: Now, unstructured inputs means unpredictable outputs.
[MAXIME]: The probabilistic nature of this technology really means that you have to provide that level of structure if you want to get the desirable outcomes that you're looking for.
[MAXIME]: The way to do that is actually to leverage Intelligent Document Processing.
[MAXIME]: So Intelligent Document Processing is much more than just the ability to extract data from a document.
[MAXIME]: The way that it works is through having the ability to transform any kind of input file, doing image enhancement, because those TIF files that you still have from back in the 1990s might not actually be really good for data extraction.
[MAXIME]: So that's what image enhancement does.
[MAXIME]: It brings it back into today's world and actually it heightens the ability to read the image.
[MAXIME]: Deterministic OCR, ICR.
[MAXIME]: So both hand-printed text and printed text are being extracted using deterministic technology.
[MAXIME]: This is based off of machine learning models as well.
[MAXIME]: But then of course, specifically fine-tuned for exactly what we're trying to do here and not a general purpose model.
[MAXIME]: Document classification, assembly, data extraction, validation, all of these capabilities come with a variety of modern technologies that you can leverage to achieve the results.
[MAXIME]: The important bit is here, is that you use the one that's actually needed specifically for what you're trying to do.
[MAXIME]: Human in the loop is one of the most important things that I talked about with a lot of people today.
[MAXIME]: Human in the loop means that you actually have the ability to have that explainability in your process as you're extracting valuable information from your documents, and at the same time, we can deliver input as to how it should have done better.
[MAXIME]: Further upscaling and training the model as it actually is continuously improving.
[MAXIME]: On the analytics, now what you show, okay, this is what we're getting.
[MAXIME]: This is the consistency level, this is our accuracy level, and this is where we're having fallout.
[MAXIME]: The important bit here is, because we're using both the best of breed technologies in terms of deterministic and probabilistic, so purpose-built AI together with GenAI and large language models, we can actually show where things are going to fail.
[MAXIME]: That might sound a little bit weird, knowing that something's going to fail and highlighting that as an actual USP.
[MAXIME]: Well, actually it's not.
[MAXIME]: If you are an organization, if you know that this is the fallout that you're going to have, the exceptions that you're going to have, that makes a huge difference, because catching that exception can be extremely more valuable than actually having to figure out what the exception was later down the line.
[MAXIME]: Then of course, the data output.
[MAXIME]: You need to be able to interconnect all of this data in a certain format that is needed for the next system, that's going to use all of that data.
[MAXIME]: That we see is something that's more and more is going to be a Gen2K AI.
[MAXIME]: So I talked a little bit about the hybrid approach to document processing.
[MAXIME]: The typical challenges that you see when you're working with large language models in terms of documents, is accuracy, compliance, the explainability, prompter-generated version drifting, everything is constantly updating in that world, which means that you constantly have to update your prompts as well because they might not be getting the same results.
[MAXIME]: Infrastructure, a lot of maintenance overhead, cost and scalability is still a question.
[MAXIME]: Tokens are getting cheaper and cheaper, but actually we see that it's not down to the level yet, where you could actually deal with the cost fluctuations that you might not expect.
[MAXIME]: We've seen so many examples where there actually was that approach and a huge amount of tokens were burned through for actually lackluster results.
[MAXIME]: So this is something that we continuously push for, and that I've talked to a lot of people about as well, to combine intelligent document processing with LLMs together.
[MAXIME]: So you can provide that very valid, very ground-truth data type of structure, and then let a language model do what it do is best, and actually work with the more complex and dynamic pieces of information that you know are correct and let it make its interpretations.
[MAXIME]: I'll talk through the use case in just a second.
[MAXIME]: But overall, this is what we call our Phoenix model, where you start deterministic to have that reliability.
[MAXIME]: This is where you want to get going first with your documents.
[MAXIME]: You want to use machine learning when there's structure that actually matters, and you want to be able to have that available to extract the data.
[MAXIME]: But then once you have all of that, then you can actually use Gen AI, large language models or agents to make that meaning and provide that understanding.
[MAXIME]: So when meeting matters, when there's nuances in the data that you've extracted, this is something that you want your Gen AI to actually go to town with.
[MAXIME]: So how do we put all of this to practice?
[MAXIME]: What's a great example of where it went down this path, where this particular customer went from a story where they were working down the path of just using a large language model for data extraction from their lease agreements.
[MAXIME]: I'm going to take you through how they got to the point where they now have a fully working solution.
[MAXIME]: So the challenge that they had, this is a large retailer of food, both in the US and in Europe, all across the world really, you can see them everywhere.
[MAXIME]: But specifically in the US, they have about so many retail locations that they need to service, and actually they're adding more and more every single year.
[MAXIME]: So they also have a lot of lease agreements that they have to process, about 30,000.
[MAXIME]: They have a team of about 25 FTEs, full-time employees that were working on getting all the data points that they needed to have to ensure compliance with the new financial accounting standards, boards, regulations.
[MAXIME]: Now, this regulation actually implied that they have to get 350 different data points out of each and every lease agreement.
[MAXIME]: So they started down this path.
[MAXIME]: They went for what is often called the silver bullet approach of using a large language model to actually get to that point where they were able to extract that.
[MAXIME]: Similar to the story that I said at the front, they saw some great results in the pilot, but as soon as they got to a point that we're actually processing things and started to validate what the outcomes were, because they were so plausible, they just kind of took it at face value, they immediately saw that it was not actually correct.
[MAXIME]: And accuracy was only getting to a mere 60%.
[MAXIME]: And this was about the same as they had with 25 FTEs doing it fully manually.
[MAXIME]: So as I worked through it with them, they actually went for my guidance with the hybrid approach, where we used our deterministic technology, purpose-built AI to actually extract almost about 340 data points automatically using our capabilities.
[MAXIME]: And then where there was interpretation that was needed, for example, there was actually this particular paragraph in each of these lease agreements that came from an adjudicator.
[MAXIME]: They gave an opinion about the specific lease.
[MAXIME]: This lease opinion had to be then translated into a certain type of category of information.
[MAXIME]: And so this is something where LLMs are done perfect for.
[MAXIME]: So combining everything together, including our manual review, so that we continuously update the accuracy level.
[MAXIME]: At the finish line of the project, and this was a couple of months ago already, we have 82% accuracy, but only five FTEs were required for their other 18% of data that had to be validated.
[MAXIME]: And all 350 data points were being provided out of the box.
[MAXIME]: So now they're definitely already in the high 90s when I last checked in with them.
[MAXIME]: And they're extremely happy about how quickly they're able to process all of this data, completely touchless with this particular hybrid approach.
[MAXIME]: So that's what I wanted to share with you guys today, a quick story about how you need to look at everything in terms of technology that's available to you and apply it in such a way that you actually have it available in your arsenal of components that you can put together in the right way and use the technology that's going to be best fit for purpose for what you're trying to solve as a business case.
[MAXIME]: Happy to take some questions later.
[MAXIME]: Max, with that fantastic presentation and coming up in a few minutes at the top of the hour, I know you'll be on the Q&A panel.
[MAXIME]: So for everyone that is listening live, just type your questions into the Q&A box that you see in front of you.
[MAXIME]: Max, from my end, one question that came to mind, I know a lot of organizations are making this transition and have, shall we say, tried to do some of these things internally and have quite honestly failed.
[MAXIME]: So what are some of the foundational elements that organizations should be looking at to decide on that buy versus build versus outsource type of decision when it comes to AI?
[MAXIME]: That's a great question, Michael.
[MAXIME]: And it's definitely something that a lot of organizations today are thinking about because, I mean, just speaking from personal experience, it's gotten so easy to build, basically to build.
[MAXIME]: I think, you know, full stop, it's gotten so easy to build.
[MAXIME]: But there's a huge difference between building a POC, an MVP, and actually a fully deployed end-to-end enterprise-grade application that allows you to fully automate an end-to-end business process and then expand upon that onto the next one.
[MAXIME]: So I think that's something that you have to take into account.
[MAXIME]: My recommendation would always be, if it's not core to your business, go out onto the market, find a vendor that actually allows you to not just tackle one particular use case, but then also lets you tackle the next one so they show you that breadth of capability and also the proof in terms of outcomes that they've had with their customers.
[MAXIME]: And if you truly can't find what you're looking for, or there's too much customization that you would need to do, then you can consider going down that path of building yourself.
[MAXIME]: But, you know, with the caveat, keep in mind that the longevity of this program that you're embarking on might be something that you come to regret down the line.
[MAXIME]: We've seen this in the past as well, where enterprises have paid consultancy companies and a lot of third-party resources so much incredible amounts of dollars simply to get them off their self-built applications onto something standard from the market so they can make it a third-party vendor, their problem of how to maintain and uptake all of those applications.
[MAXIME]: So keep that very well in mind that you're not just walking down this path blindly because history is prone to repeat itself.
[MAXIME]: Makes a lot of sense.
[MAXIME]: Well, as I mentioned earlier, type your questions into the Q&A box.
[MAXIME]: We'll get to as many as we can during a live Q&A coming right up.
[MICHAEL]: Maxime's presentation really highlighted the importance of Intelligent Document Processing in today's AI landscape. This was part of the live panel at our Software Oasis AI Summit, where real audience questions took center stage. These questions will help us further explore how businesses can unlock the potential of AI with structured data.
[MICHAEL — CHECKPOINT]: So, what Maxime made very clear was that without structured data, AI can't deliver the consistent results businesses need. Now, you might be wondering: how do companies transition from pilot success to full-scale deployment? And how can they ensure their data is ready for AI? And that's exactly what this audience brought to the table in our Q&A panel.
--- Q&A PANEL ---
[MAXIME]: I hope everyone is enjoying the event so far.
[MAXIME]: For everyone that's been with us since first thing this morning at 8 Eastern time, we've come to our first Q&A panel, and we'll have the opportunity to ask a lot of the questions that have been submitted to each of the panelists.
[MAXIME]: After this panel, we'll have a presentation from Lucas over at LL Informatics.
[MICHAEL]: So why don't we get things started?
[MAXIME]: I'll introduce everyone on the panel.
[MAXIME]: Joining us from NTT Data, we have Niraj Singhal.
[MAXIME]: He's the group senior VP.
[MAXIME]: Welcome, Niraj.
[MAXIME]: Next we have joining us from Superbow AI, we have Dimitri Papazisis.
[MAXIME]: He's the co-founder and CEO.
[MAXIME]: And finally, joining us from Abbey, we have Maxime Vermeer, who is the VP of AI Strategy.
[MAXIME]: Welcome, Max.
[MAXIME]: Hey, everyone.
[MAXIME]: So let's get started.
[MAXIME]: I want to try to get to as many of these questions as we can.
[MAXIME]: We had just a ton of questions come in from all around the world.
[MAXIME]: So the first question that came in was for Niraj, and, Niraj, this question comes to us from right in this area where we are, just outside of Boston.
[MAXIME]: This is from Ethan, who's, as I said, just outside of Boston in Massachusetts.
[MICHAEL]: And Ethan is asking, when you advise executives to move from discrete AI pilots to business model reinvention, what's the very first metric you ask them to change so everyone feels a shift towards outcomes and not experiments?
[MICHAEL]: Yeah, that's a very, very interesting question, Ethan, and thank you for that one.
[MAXIME]: I think a lot of people confuse AI as any other technology.
[MAXIME]: My feeling is AI is more a mindset than technology.
[MAXIME]: So I do not believe that you can get the full power of AI unless you change perceptions, you change mindsets, you train people up, and you make them more outcome-oriented.
[MAXIME]: So I think the starting point is really sensitization of the teams that are involved, the extended leadership within the client organization in terms of what needs to change from behaviors, attitudes, perceptions, before they really have the full impact of AI felt within the organization.
[MAXIME]: So you start with people, and then ultimately you get into process and technology as you scale up and move towards business process reinvention.
[MAXIME]: That's really the way we look at it.
[MAXIME]: Makes sense.
[MAXIME]: The next question came in for you from Harper, who is in, looks like just outside of Denver, Colorado.
[MICHAEL]: And Harper is asking, when a team shows you a successful LLM invoice pilot at 95% accuracy, what specific questions do you ask to stress test whether it will survive real production document chaos?
[MICHAEL]: Well, the very first question is, how many times did you run through the same invoice and did you get the same result every single time?
[MAXIME]: I think the probabilistic nature of this technology really shines in the fact that it can create incredibly plausible results.
[MAXIME]: And it's not that you should not use LLMs in any type of production pipeline when it comes to documents.
[MAXIME]: It's actually the inverse, where you need to combine it with other types of technologies that truly ground it inside of the data that is truly inside of documents.
[MAXIME]: You can have all of those references and you can trust the output that you're getting.
[MAXIME]: Where we've seen much more success in leveraging that kind of capability is when you have that ground through data and you are then able to leverage a large language model, for example, or vision language model for a specific part of the task that you're trying to do in your full document pipeline.
[MAXIME]: Fantastic.
[MAXIME]: And Dimitri, this next question came in for you.
[MAXIME]: This is still here in the U.S.
[MAXIME]: We have from Avery, who is in, looks like just outside of Miami, Florida.
[MAXIME]: So all the way in the southernmost point in the U.S.
[MAXIME]: Not quite the most southernmost.
[MICHAEL]: Avery is asking, when a CIO tells you their agents are working fine behind an API, what red flags do you look for to judge whether they actually have the sovereign governed infrastructure you described?
[MICHAEL]: All right.
[MICHAEL]: So I think everything starts from a different standpoint, right?
[MAXIME]: So first of all, the way we believe and approach is, first of all, we need to make sure that the enterprise has the proper data structure.
[MAXIME]: Most enterprises, they are, you know, magic results and making sure that their data is very well structured.
[MICHAEL]: Most of the time, data is chaotic, scatters here and there, back and see your PRMs or different media, you name it, right?
[MICHAEL]: So for us, before you reach the AI agents, you need to make sure that your vendor has isolated systems and data.
[MAXIME]: He has an abstraction layer to make sure that we vacuum the unstructured data and we structure the data in a way that at least our system will understand and make sure we deploy a system of agents, not one agent, a system of agents, agentic AI.
[MAXIME]: Then we need a knowledge foundation layer, which is basically all the knowledge, sovereign data control and the governed orchestration layer.
[MAXIME]: If you don't have those in place, in our opinion, your AI agents are not really working.
[MAXIME]: And also we need to make sure that your AI agents co-work together.
[MAXIME]: So it's different when a task-oriented standalone tool and still a different thing if you have an agentic AI, which is a system of agents, they communicate, they work together, they reason together, and they're goal-oriented.
[MAXIME]: So these are very important.
[MAXIME]: Fantastic.
[MAXIME]: The next question came in from Neeraj, but before that, I will stand corrected.
[MAXIME]: I see a lot of comments came in.
[MAXIME]: Yes, Key West is the southernmost point in the US.
[MAXIME]: So for a lot of our fans in that region.
[MAXIME]: True, very true.
[MAXIME]: Correct.
[MAXIME]: Neeraj, the next question comes in to us from Priya, who's in London, over in the UK.
[MICHAEL]: And Priya is asking, how do you keep business stakeholders focused on reimagining customer outcomes when the loudest internal conversations are still about model choice and vendor selection?
[MICHAEL]: So that's, again, a very, very interesting question.
[MAXIME]: I think a lot of times you've got to take a step back and really reflect on the reason why you're there.
[MAXIME]: The reason why every business is there is to make it easier for our customers to get the outcomes that they desire.
[MAXIME]: That's the sole purpose of any business.
[MAXIME]: So it's very important to start reimagining your business with the customer in the center.
[MAXIME]: That's the starting point.
[MAXIME]: I firmly believe that customer dissonance and customer fiction become the starting point of reinvention.
[MAXIME]: So what AI and agentics and everything new that's coming in from a technology perspective really helps you do is reimagine the business with the customer at the center, which essentially means the way you were dealing with the customer and the way you had outlined your processes may not be relevant today, and there might be an easier way of making the customer do business with you.
[MAXIME]: So I think it's true self-reflection of the reason why you're there, keeping the customer at the center, and then thinking about the business and reimagining the art of the possible is the approach that one needs to take.
[MAXIME]: The models will flow.
[MAXIME]: The models will be commodity.
[MAXIME]: I have this very interesting thing in my mind.
[MAXIME]: If you look at Java as a technology that came in about 25, 20, 30 years back, everyone was talking about Java, but today Java is pretty much involved in everything, but nobody's talking about Java.
[MAXIME]: It's about the outcomes that you can achieve with Java as the underlying basis for that.
[MAXIME]: So I think AI in the next five years is going to become commodity.
[MAXIME]: What's going to be really differentiating is the outcome.
[MAXIME]: AI is going to be native in everything that we do for our clients, everything that we do in our businesses.
[MAXIME]: And I would agree with you that the models, even at this point, are getting so good that it's in many cases hard to differentiate.
[MAXIME]: So this next question comes to us from Max, who's in...
[MAXIME]: Actually from Jackson, for Max, from Raleigh, North Carolina, and shout out to any NC State listeners in the Raleigh, North Carolina area.
[MICHAEL]: The question is, how do you decide which fields in a complex document should be handled by deterministic IDP versus left-to-gen AI so you get both reliability and flexibility?
[MICHAEL]: No, that's a great question, and that truly speaks to thinking about what the right approach is for the task at hand, right?
[MAXIME]: So in essence, my recommendation is always that you go for the majority of all the data points that you want to extract from a document, or basically any piece of content through deterministic capabilities, deterministic AI, machine learning models.
[MAXIME]: And it's only when you actually can basically limit the amount of context that needs to be actually analyzed or processed by a large language model that you can actually then also minimize the amount of risk that you might have for hallucinations based on the original document that you're processing.
[MAXIME]: In a great example, we've seen this a couple of times where one of our customers had to process lease agreements, massive documents, thousands of pages sometimes, and a specific use case where they did use a large language model was after we already had extracted the specific segment that contained the answer, but then it required interpretation.
[MAXIME]: It required the deeper level of transformation that you can get with deterministic AI.
[MAXIME]: And that's, for example, a perfect use case of where you want to use the tandem of the two and all other, actually this was 350 data points, were all extracted using deterministic capabilities because it gave them that assurance that if there was something wrong, we could actually be able to tell them, whereas the probabilistic will always give you an answer.
[MAXIME]: It's just much harder to detect whether it's the right one.
[MAXIME]: And so in this particular instance, that was also linked to compliance and governance that they had to provide to the regulators.
[MAXIME]: So in this particular instance, the probability of a wrong answer had a much higher impact in terms of down the line for the business than having to go through and do something manually.
[MAXIME]: So that was actually their preference.
[MAXIME]: So as always, there's more than just the technical aspect that makes the decision of what you're going to use and when.
[MAXIME]: Dimitri, this next question comes to us from Oliver, who is in Paris, France.
[MICHAEL]: And Oliver is asking, when agents need to work across PDFs, spreadsheets, ticketing tools and eargap systems, how do you design the orchestration so data never silently leaks outside sovereign boundaries?
[MAXIME]: All right, so we have our own AI stack entirely within the sovereign boundary.
[MAXIME]: We begin with a full governance chain, which is for every action, which is traceable to source.
[MAXIME]: These are systems and data.
[MAXIME]: Then we go to policy enforcement, governance rule, enforced within the boundary and not the...
[MAXIME]: Then we have the governed model integration.
[MAXIME]: Actually, it's model outputs governed.
[MAXIME]: And we have, for any data, it's governed data control.
[MAXIME]: So we don't have any sensitive data at all on our language model.
[MAXIME]: So we make sure that we are not in mind.
[MAXIME]: And then when we are talking about deployment, we're talking about air gap deployment, which actually operates without external network dependencies.
[MAXIME]: And these were our AI systems, which is just AI agents really work.
[MAXIME]: So the way we approach it, it's not a fast way to go to market, but it's the only way to go to market.
[MAXIME]: And what I'm trying to say, it's not a secure way to go to market, and that is why we decided, this is where, you know, we've been around since 2010, so before AI revolution, we began from doing NLP.
[MAXIME]: So, you know, we always took under very serious consideration to have everything...
[MAXIME]: ...include...
[MAXIME]: So...
[MAXIME]: ...is the player.
[MAXIME]: What I...
[MAXIME]: ...were actually layers...
[MAXIME]: ...police enforcement, the governed model...
[MAXIME]: ...deployment...
[MAXIME]: ...happened within the layer...
[MAXIME]: ...round up...
[MAXIME]: ...isolated...
[MAXIME]: ...layer to the resource layer, to knowledge foundation layer, to governed orchestration layer, and eventually to AI agents, which are going to be air-gapped, it's going to be the air-gapped deployment.
[MAXIME]: Makes a lot of sense.
[MAXIME]: So, Neeraj, this next question comes to us from Grace, who's in Austin, Texas.
[MAXIME]: And Grace is wondering, and I gather that she's in the banking industry, for a bank that has already automated pieces of the mortgage workflow, how do you decide which end-to-end journeys are mature enough for agentic AI...
[MICHAEL]: ...versus those still too broken in their underlying processes?
[MICHAEL]: Yeah, excellent question again.
[MAXIME]: So, you know, this is the part that I think we should be moving away from.
[MAXIME]: So, if you really look at the initial adoption of AI and predominantly agentic AI, most of the adoption was for cost and efficiency.
[MAXIME]: So, if you look at an end-to-end process and you say, I'm going to streamline these processes, I'm going to deploy discrete agents across different aspects of this value chain.
[MAXIME]: And I'm going to achieve a compression of about 30 to 40 percent, I'm going to reduce cost and I'm going to improve efficiency.
[MICHAEL]: But I think the fundamental question that one needs to ask is, what's the outcome that the customer is looking for?
[MICHAEL]: Well, the customer is trying to buy the best home for himself or herself.
[MICHAEL]: So the outcome that the customer is looking for is, can you help me buy the best home possible?
[MICHAEL]: And if financial institutions start reimagining their businesses to make customers get the best possible home rather than provide the customer the best possible mortgage, you start looking at the art of the possible.
[MAXIME]: So, I would say, look at what the customer is trying to do and reimagine your business end-to-end in terms of how you can cater to that and how you can help the customer rather than cherry picking pieces of the value chain where you can drive discrete efficiency.
[MAXIME]: So, it's a completely different paradigm.
[MAXIME]: I don't think many financial institutions are there yet, but that's really the way to get the full power of agentic AI.
[MAXIME]: It's not to look at discrete compression, but to look at the end-to-end lifecycle with the customer at the center and becoming more outcome oriented in terms of what the customer is trying to do.
[MICHAEL]: The next question comes to us from Max, and it's from Chloe, who is just outside of Phoenix in Arizona, and Chloe's question is, for leaders under pressure to do agents now, how do you persuade them to invest first in document quality and structure instead of skipping straight to agent orchestration?
[MAXIME]: Because I think that the reason, or at least the persuasion has already been kind of called out by the other participants, Neeraj and Dmitry as well, right, you need to have that level of data and perception available for your agents to actually be successful.
[MAXIME]: You know, I often compare it to when they first started working on the self-driving car.
[MICHAEL]: Did you actually figure out first how to make it turn left or right?
[MICHAEL]: No, you actually needed to give it that perception layer to understand where the hell it's going on the road, what lanes are, if there's pedestrians.
[MICHAEL]: So if it doesn't understand the business world that it's operating in, how on earth can you actually have it do anything that's going to actually further your business, right?
[MAXIME]: So I think that's the important thing to remember.
[MAXIME]: Makes a lot of sense.
[MAXIME]: The next question came in for Dmitry, and this is from Jaina, who is, looks like in Warsaw, Poland.
[MICHAEL]: And Jaina's asking, what are the benefits of having a self-driving car?
[MICHAEL]: Makes a lot of sense.
[MICHAEL]: The next question came in for Dmitry, and this is from Jaina, who is, looks like in Warsaw, Poland.
[MICHAEL]: And Jaina's asking, as model providers ship new capabilities every few weeks, how do you maintain model agnostic governance so that policy and accountability don't have to rebuild, be rebuilt with every upgrade?
[MICHAEL]: Well, first of all, yep, that's true.
[MICHAEL]: Models are getting smarter, and this is exactly what makes them a commodity, right?
[MAXIME]: So and this is, that's why we just use them.
[MAXIME]: Today, we, as you know, in Superbo, we are LLM agnostic, so we use more than one LLM, depending on what we are trying to succeed.
[MAXIME]: For example, a different LLM for reasoning or a different LLM for analytics depends.
[MAXIME]: Each and every model has its own unique superpower, right, in terms of strengths and weaknesses.
[MAXIME]: And now the security does not, you know, and the policies don't happen inside the model.
[MAXIME]: They happen outside the model in your infrastructure.
[MAXIME]: That's why you need to have your own AI stack.
[MAXIME]: Otherwise, this is, you know, otherwise you're just an LLM robber and it's going to go south.
[MAXIME]: So you need to make sure that you have your own AI stack, which is going to be a full layer above the model.
[MAXIME]: What we do in Superbo, besides being LLM agnostic, we're also infrastructure agnostic.
[MAXIME]: It means that we can deploy in any kind of cloud, private cloud or on premise without a cloud.
[MAXIME]: So we don't depend on one.
[MAXIME]: It's very important, we cannot go out there and claim sovereignty if we depend on one model or one cloud infrastructure.
[MAXIME]: Not only that, but I'm going to take it a step further.
[MAXIME]: Even the framework that we build our agents is something that we've built from scratch.
[MAXIME]: So we don't use hyperscalers to build our own AI agents.
[MICHAEL]: What I'm trying to say is that, for example, it would be very easy and very fast, actually, to build AI agents on Google Vertex, right?
[MICHAEL]: Go there and it would be very fast, go to market.
[MAXIME]: The reason that we chose the hard way was we know very well the hyperscalers are built horizontally.
[MAXIME]: It means that we're going to hit a wall at some point when it comes to bespoke tailor making and customization.
[MAXIME]: That's why we build our own agendic framework so we can build our own AI agents in there.
[MAXIME]: We had to make the whole framework to be infrastructure agnostic and LLM agnostic.
[MAXIME]: So anything around security, the governance, the orchestration, everything happens after the model.
[MAXIME]: So it has nothing to do with the model.
[MAXIME]: And this is what air gap deployment more or less means.
[MAXIME]: Neeraj, the next question came in for you.
[MAXIME]: And this question came in from a few different people in slightly different ways.
[MAXIME]: So I'm going to pick one.
[MAXIME]: I'm going to pick one from Daniel who is just outside of Chicago, Illinois.
[MICHAEL]: And Daniel's asking, when tech debt is high and budgets are tight, how do you prioritize which parts of the legacy stack must be modernized now to unlock AI value versus what can safely wait?
[MICHAEL]: You know, this actually reminds me of a quote from Edward Deming.
[MICHAEL]: And Deming, who was known for probably the guru in quality used to say, you don't have to do this.
[MICHAEL]: Survival is not mandatory.
[MICHAEL]: You don't have to do this.
[MICHAEL]: Survival is not mandatory.
[MICHAEL]: If you're living with tech debt, with all the influx of technology and the changes that are happening, it's going to be really, really, really difficult to streamline your business.
[MAXIME]: So I think the starting point obviously is tech debt.
[MAXIME]: And I would say even from a tech debt perspective, look at client facing, customer facing systems first.
[MAXIME]: Those are the ones that are going to give you the maximum return.
[MAXIME]: Those are the ones that are going to have the maximum impact and then work your way to the back end.
[MAXIME]: So I would say address tech debt holistically if possible.
[MAXIME]: If you can't and you've got only 10 percent budget available, which usually is the case, look at making the maximum impact for the customer.
[MAXIME]: And that's the way to prioritize processes.
[MAXIME]: Great answer there.
[MAXIME]: And, Max, I'm going to send this last question to you as we're coming to the end of our time.
[MICHAEL]: Sarah, who's in Stockholm, Sweden, asks, when you embed Human-in-the-Loop review into document workflows, how do you keep it from becoming a permanent bottle deck instead of shrinking safety net over time?
[MICHAEL]: Awesome.
[MAXIME]: Well, I love having questions from fellow Europeans.
[MAXIME]: And it's actually a really good one.
[MAXIME]: So Human-in-the-Loop is one of the key components that you want to have inside of your document workflow, even inside of your Gentic workflow, any workflow whatsoever.
[MAXIME]: Because I think that human component, depending on what type of organization you are, what type of industry that you're working in, the regulations, compliance, you're going to need it.
[MAXIME]: Now, whether it becomes a bottleneck or not, that's something else that you can help determine.
[MAXIME]: But actually that can also be determined by the regulations and compliance that you have to deal with.
[MICHAEL]: Now, specifically to the more question of, OK, how do you avoid that it just becomes more and more and creates a choking point for your actual workflow?
[MAXIME]: I think that's not particularly a problem that we see often happening inside of our customers when they actually deploy our technology for two reasons.
[MAXIME]: One, the capability to actually take that human input and further train the system on how it should have dealt better with that information that was provided is the thing that truly sets it apart in terms of it being an educator for agents, for the models that are being used.
[MAXIME]: And the other component is absolutely that the fact that you are able to leverage it now in combination with, for example, like Dimitri was talking about, the Gentic capabilities, allow you to actually also accelerate that.
[MAXIME]: So we're no longer just having human in this room, but there actually is assistance from agents happening too.
[MAXIME]: So they're making the humans actually be able to work through it all faster in terms of, you know, not creating that choke point, but actually assisting them, making them work more faster.
[MAXIME]: And that's what we see all the time, that there's a high request.
[MAXIME]: So to my earlier point, it does really depend on regulations, industry, the specific use case.
[MAXIME]: But the main thing is, is that you're always going to want to have it.
[MAXIME]: And I've never seen it become a true choke point because the systems that you leverage it in should be able to learn from the additional input.
[MAXIME]: And now there's even more and more techniques to actually further assist the human in the loop to move faster and faster.
[MAXIME]: I think a lot of great insights there.
[MAXIME]: I'd like to thank Neeraj, Dimitri, and Max for their fantastic presentations earlier in the event and for joining us for these audience questions on this panel.
[MAXIME]: Coming up in just a moment, I want to keep everybody on target for the next session.
[MAXIME]: But if anyone in the audience has additional questions or we didn't get to your question, at the top of the site you'll see that big orange button.
[MAXIME]: You can submit your, you know, your question there and someone on the team or one of the panelists will respond to you after the live event.
[MAXIME]: So with that, I will send everybody off to the next session.
[MAXIME]: Thank you.
[MAXIME]: Thank you, Michael.
[MICHAEL]: We've heard some compelling insights today. First, the journey from OCR to AI-driven solutions at ABBYY shows the power of evolution in technology. Second, the importance of transforming unstructured data into actionable insights is key to leveraging AI effectively. And third, the role of Intelligent Document Processing can't be overstated. These takeaways really underscore why it's so crucial for businesses to focus on data readiness for AI. Maxime, thank you for sharing your expertise with us. If you'd like to connect with Maxime, you can find him on LinkedIn. And if you haven't already, come join us over at softwareoasis.com — that's where all of this really comes together. If you enjoyed this episode, consider subscribing and leaving a review. Until next time, I'm Michael Bernzweig, and take care.