🎉 Introducing Appmixer 6.0: New features bringing the future of automation closer to today
Learn more
In this episode, Maxime Vermeir, Senior Director of AI Strategy at ABBYY, takes us through his journey with AI, from his early experiences in machine learning to the transformative power of generative AI today. Maxime explores the growing role of AI in product development and how it’s become an essential part of everyday tools.
We delve into ABBYY's approach to "purpose-built AI" in Intelligent Document Processing (IDP), emphasizing the critical need for transparency and explainability in AI-driven solutions.
But is every software company’s latest AI feature a genuine product innovation, or is there a marketing ploy at play? Should companies rely on existing AI models or build their own from scratch? And how does AI empower product managers to work more efficiently?
Get the answers to these questions and more in the fourth episode of our podcast.
There's actually a funny story about this. My academic background is in industrial sciences, specifically in computer technologies. Back in the day, I even specialized in telecom networks. As part of the course, we did a lot of programming, experimenting with technology, working with motherboards, soldering—everything. It was a very comprehensive program, and we even covered things that, at the time, were just called machine learning.
Not too long ago, I think it was last year, we had an alumni evening. Some of my former professors invited me to speak about the current state of AI because they knew I had opinions on the subject. It really hit me how long I've actually been working with AI when they handed out stickers to all the alumni that said, "Graduated before ChatGPT." I thought that was hilarious. I still have the sticker on my laptop, and I love it.
I often mention this story when people say, "AI is so new." My response is, "No, it’s older than you think. It’s older than I was in college!" To trace AI back, depending on who you ask, you’d have to go to the 1970s or even earlier to see the first forms of machines learning to perform tasks.
To trace AI back, you’d have to go to the 1970s or even earlier to see the first forms of machines learning to perform tasks.
The interesting part now is the hype and craze around generative AI, where suddenly everyone thinks they know what AI is. But that’s an important distinction—they think they know what it is. It’s fascinating to be able to talk about AI now, even though the foundational mathematics behind it is something we've been using for ages. For example, your iPhone learning when it needs to be charged—that’s AI, and we’ve had that for years.
The ability to enhance photos automatically, combine images—all of that is AI too. And these are features that have been sitting in your pocket for several years now. I think many people still underestimate just how long AI has been a part of our everyday lives.
From a product perspective, you have to admit that OpenAI made a really smart move in how they introduced their technology to the world. A lot of people don’t realize that the very first GPT model was actually released in 2018—GPT-1. But nobody knew about it, and nobody really cared. Even I had to look it up, and I’m supposed to know these things! It wasn’t until ChatGPT came along that things changed. The interesting part is that the core capabilities were already there in GPT-3.0, and GPT-3.5, but the game-changer was that OpenAI gave it an interface that was easy for everyone to use.
That simple move—giving the technology a user-friendly face—was a brilliant strategy. It allowed people to see for themselves what the technology could do, and from there, everything we’ve experienced in the last few months just exploded. This approach will definitely go down in history as a very smart way to create demand and make AI mainstream by making it accessible to everyone.
Well, maybe to clarify what Intelligent Document Processing (IDP) is, because depending on who you talk to, people either know what it is or they confuse it with something else.
Document processing is a more complex version of what a lot of people know as OCR—Optical Character Recognition. OCR is a type of computer vision. I actually have another funny story for you about explaining what we do at ABBYY and how IDP fits into the whole AI landscape.
I was in California not too long ago, and as a European citizen, you have to go through customs and border control. They’re always very interested in what brings you to the country. The officer asked me what I do for work, and I told him I’m the Senior Director for AI Strategy. His response was, “How can you be your age and a Senior Director in AI if AI has only been around for two years?”
He thought I was making things up! So I explained to him that the company I work for has been doing AI for more than 30 years—actually, for 35 years now. AI is even older than that. Similar to our earlier topic, he was baffled. I had to explain exactly what we do, not because he was curious about my occupation, but because he was amazed that AI has been around for this long. It was an interesting conversation to have at border control, but it highlights how much confusion exists under the AI umbrella. People often mistake the “old ways” for new innovations, but AI has been evolving for a long time.
When it comes to building solutions like an IDP platform, as we do at ABBYY, all the right pieces need to come together to make it something people can actually use. That’s why we focus on what we call “purpose-built AI.” It’s really our mantra for how we look at the capabilities that AI offers.
Over the last few years, we’ve seen a lot of startups and other companies just pushing technology into products, hoping something sticks. This has created a sentiment in the market where people are searching for the usefulness of generative AI within the broader spectrum of enterprise applications. That’s where we’ve positioned ourselves—as a company that purposefully builds AI, using the right type of AI, whether it’s machine learning, neural networks, or even generative AI and language models. The goal is always to solve the customer’s problem.
We’ve been doing this for many years. We started with OCR, literally turning a page of text into something a computer can understand. That has now evolved into Intelligent Document Processing, which means we can understand the entire content of a document. We can recognize who is being referenced, whether there are multiple parties in a contract, the amount on an invoice, whether a tax form is complete, and more. All of these processes can now happen automatically, thanks to IDP.
If you think about where IDP is useful, you’ll see it in many places. For example, when you send a document to your bank, there’s likely an IDP platform behind it. There’s no one manually entering that data into the system anymore—it’s all automated. This plays into the larger trend of improving customer experiences. No one wants to wait for a loan to be approved or for onboarding as a new customer. Whether it’s scanning a passport, ID card, or driver’s license, all of this is powered by IDP behind the scenes. This data, whether from physical or electronic documents, is transformed into information that can be fully automated into processes.
So, in a nutshell, that’s what IDP is. And to your question about where AI plays a role—it's everywhere, in every single step. The key point is that it encompasses the full spectrum of AI capabilities, not just generative AI. One issue I’ve seen is people trying to solve large problems with just one technology, like generative AI, and then wondering why they aren’t successful.
I would definitely say that this shift came from a need within the organization, or rather, a need that emerged in the market. People suddenly wanted to know more. At the start of my career, I was told very clearly: don't talk about technology, only talk about outcomes and solutions. Back then, people were only focused on the results you could deliver as a vendor. How you solved the problem? That was for the tech guy in the corner, and he might have cared, but he had no real input.
Then, almost overnight—especially in terms of how fast the enterprise industry moves—everyone, from the CEO to the CFO, suddenly wanted to know, "Hey, how does this technically work?" As a tech guy, I was thrilled! All the things we’d been doing before AI was cool, we could now talk about and call cool. It was a shift where the market wanted to know the technical details, and we had to respond to that.
So, I was able to dust off a lot of technical things, put a shiny new polish on them, and say, "Look, this is what we have." People were amazed, like, “Wow, we didn’t know.” Well, they didn’t care before, but now they did. That’s one of the big changes.
On the other hand, if you look at the evolution of artificial intelligence, particularly in terms of computing power and the availability of data, these are the two driving factors that are enabling us to explore AI’s full capabilities. This combination of more powerful infrastructure and vast amounts of available information is what’s truly making a difference.
Let me explain that a bit more. One of the things people are confused about is the creation of new capabilities, like what we see with GPT. We're continuing to pump more processing power and data into AI technology that hasn’t fundamentally changed since 2017, when someone at Google wrote a pivotal paper. In just a short amount of time, we’ve been increasing our infrastructure and data, which has led to discovering that, when we do this, AI does things we didn’t expect. So now, we’re on a journey of exploration.
This also means that along the way, we’re uncovering new, useful applications for our Intelligent Document Processing (IDP) platform. We don’t need AI to think on its own; we just need it to look at documents more accurately, process them faster, and make setup easier for customers. That’s what they care about.
Is this purely marketing? No, not at all. There are definitely organizations out there that take whatever is new and throw it into their products without really thinking about how it delivers value for customers. But at ABBYY, we’re focused on showcasing everything we've already accomplished. As a technology company, it's really exciting to do that. While we still need to deliver outcomes for customers, now we also have to talk about the technology behind those outcomes. And as we spot interesting capabilities from all the amazing research happening right now, we can incorporate that into our products, improving performance, speed, and capabilities to make our customers' lives even easier.
That’s the exciting part—the speed at which new things are being discovered in the tech world right now is incredible to be part of.
Yeah, absolutely. I think there’s definitely a push because of the expectations being created in the market. People want to see AI features in products. I often compare it to consumerization. A great metaphor for this is the release of the very first iPhone. Before then, enterprise software was gray, difficult to use, and required experts who worked with it day in and day out. Then the iPhone came along, and suddenly, people had technology in their hands that just worked, and it was easy to use.
This shift in the consumer market created a change in the enterprise market as well. Now, enterprise customers wanted to be able to do things themselves, without needing experts twice over. This is why we see AI being extremely successful in creative processes. However, in other areas, it becomes more complex to implement, even though AI has clear merits and capabilities, like integrating it into an IDP platform to reduce setup time, which is often the hardest part.
The challenge is figuring out how to create a feature powered by generative AI that not only lets you claim it’s AI-powered but also genuinely makes a difference for the customer. Unfortunately, what I’ve seen in the market is a push to include AI in products for a cool demo, but when it comes to actually using it or moving to production, it becomes impossible. I think the industry is moving past that now, though. When I talk to customers today, they ask about our AI plans, but they’re more focused on the results—what AI can actually deliver.
There’s a growing realization in the market that people are looking past the surface-level AI hype. They're saying, "Okay, show me what's under the hood. I want to make sure it not only looks impressive but also solves my business problem." That’s really the core of it. We've already seen that it’s possible, which kick-started the whole low-code, no-code transformation over the last few years.
Today, customers are saying, "Okay, show me what's under the hood. I want to make sure it not only looks impressive but also solves my business problem."
ChatGPT did a similar thing—it gave consumers a technology that was incredibly easy to use, while behind the scenes, it was complex and powerful. But this ease of use has now created an expectation that AI should "just work." People think, “If I ask a question, I should get an answer. If I want to do something with your technology, why isn’t it just working?” There’s now an expectation that products should have AI capabilities built in.
At ABBYY, we've had the benefit of working with AI for a long time, so we can easily showcase our AI capabilities without needing to immediately incorporate new technology into our products. However, we do see that AI opens up new possibilities, and that’s where the balance comes in—adding new features without always knowing exactly what they’ll deliver for the customer. That’s been the journey with generative AI for everyone: figuring out how to make sure it consistently does what we want, since it's a probabilistic type of technology.
That's why we see AI being extremely successful in creative processes. However, in other areas, it becomes more complex to deal with, even though it clearly has merits and capabilities. For example, adding AI into an IDP platform can significantly shorten the setup time, which is often the hardest part.
The question becomes: how do you create a feature powered by generative AI that not only lets you say it's AI-powered but actually makes a meaningful difference for the customer? Unfortunately, what I’ve seen in the market is a lot of companies pushing to include AI in products for the sake of having a cool demo. But when it comes time to actually use it or move it to production, that’s where things fall apart.
I think the industry is starting to move past that. When I talk to customers today, you still get the question, "What kind of AI ideas do you have? What are you planning to do with AI?" But what they really want to see is the outcome. They want to know what AI can actually deliver. More and more, there’s been a realization in the market that people are looking past that initial AI veneer. They’re saying, “Okay, show me what’s under the hood.” They want to make sure it’s not just impressive for a few moments but that it genuinely solves their business problem. And that’s really the core of it.
From our perspective in the IDP space, particularly at ABBYY, we make sure that the most important components of our platform are fully proprietary. This is essential because we want to be able to explain to our customers exactly how our technology works, so it’s not a black box. The explainability and traceability of AI are becoming increasingly important, especially from an InfoSec and compliance standpoint.
With many vendors offering capabilities as a service, like OpenAI, transparency can become a challenge for organizations. For instance, in KYC (Know Your Customer) processes, where we process passports, utility bills, driver’s licenses, and other forms of personally identifiable information (PII)—and even in healthcare, where we’re dealing with sensitive data—this transparency is critical.
The rapid evolution of this landscape has brought increased scrutiny from customers, prospects, and the market itself. There's a growing focus on whether AI is compliant with new and emerging legislation, and equally important, whether it's ethical. These are the kinds of questions we’re getting more and more frequently.
That’s why it’s crucial for us to ensure that we can explain and demonstrate how our technology works. We also emphasize the importance of showcasing the open-source capabilities we've investigated, optimized, and adapted with our own training data. This allows us to maintain full end-to-end traceability, so our customers have a clear understanding of the finished product they’re working with, from top to bottom.
I think Stanford University actually did a kind of matrix evaluating the transparency of different models available, which is quite interesting to look at. They rated these models on various categories, such as how transparent they are about the training data used, the technology behind the model, and even energy consumption—ESG is a big deal these days. People are becoming increasingly conscious of the fact that, for all their cool capabilities, these models are incredibly power-hungry. Companies like Meta and OpenAI have discussed this, and many organizations have to meet ESG goals.
To get back to the question, "Is OpenAI too much of a black box?"—for a lot of organizations, particularly those in government, healthcare, or highly regulated industries, the answer is yes. There's very little transparency about the training data they use, which can be a big concern when you're dealing with things like KYC (Know Your Customer) processes, where sensitive PII (Personally Identifiable Information) is involved. This applies even more to sectors like healthcare, where people's most sensitive data is at stake.
There has been some progress, though. OpenAI, for example, has addressed concerns about shielding your data from being used for their model training. They quickly resolved this issue when they began offering enterprise-level capabilities. Also, the option to run OpenAI on Azure helps alleviate some concerns, but overall, many people still feel there’s not enough transparency to use it directly—especially in heavily regulated sectors. Just last week, I spoke to a company in the Netherlands that had to drop a competing vendor because they were using OpenAI, and it simply wasn’t transparent enough to meet their compliance requirements.
I think the issue is that we now have an abundance of models to choose from. If you're a product manager or in the market, you're faced with a forest of models to evaluate, each with its own strengths and weaknesses. It’s critical to carefully consider what exactly you need the AI to do. You need to think about the specific function you need it to fulfill and choose the right AI capability to solve that problem.
I was talking to a group of data scientists who were going to ask the business for 30,000 documents to annotate in order to identify a specific number, intending to build a model for it. This was during the height of the trend where everyone was building their own models. I told them, "I can solve this in five minutes with a simple regex." They were shocked because they hadn't considered that solution. It’s a perfect example of how everyone got tunnel vision on AI, thinking everything had to be solved by a complex model. But now people are realizing that sometimes older, simpler solutions might actually be better for certain tasks than the latest AI models.
It's a fascinating evolution. When we talk about compliance and transparency, there's definitely a recognition in the market now that not every vendor or model is open enough for certain applications. The focus is shifting to finding the right technology for the specific problem you’re trying to solve. That's the most important factor.
For me, AI has definitely changed a lot, especially in terms of my personal productivity. I love using all the new capabilities available and try to automate as much as I can. I’ve even experimented with creating a kind of "duplicate" of myself, so when people have questions, they can chat with my AI version instead of asking me directly. I humorously named it "MaxGPT." I really enjoy the possibilities that AI offers, and I think they’re endless.
From a product management perspective—my previous role—AI has also had a significant impact. It has made certain problems easier to solve, but at the same time, it has introduced new complexities. For product managers today, it's a fun challenge to navigate the vast menu of possibilities that AI brings. There are so many things that are now possible, or easier to implement than before, without needing to build everything from scratch. You can simply integrate these capabilities into a product and make them useful. But the key challenge is that "make it useful" part—what does it really change for the customer? How does it enhance the product? If it’s not going to be used, then it can become quite costly and difficult to manage.
A great example of this relates to both this and the previous question. Not too long ago, OpenAI announced they are going to deprecate GPT-3.5 and 3.5 Turbo. These were the first models that many companies started integrating into their products to offer AI-powered features. If you think about it, these models were introduced just a year or a year and a half ago. Now, in less than a year, these companies have to switch to the next model.
For enterprise customers, this can be a huge challenge. Many of them, even if they’re using SaaS platforms, have strict rules and regulations around updates. The larger the company, the stricter the regulations. Some can only update their systems once a year, and even then, only within a very narrow window. This makes it much harder for product managers to keep up with the rapid innovation, while also dealing with the volatility that comes with it. While the pace of innovation is exciting, it also brings instability, which isn’t ideal when you're trying to build a stable product that customers can rely on. So, it's definitely something product managers need to navigate carefully.
No, I haven’t made MaxGPT public yet. It is something I’ve built for myself, and I’ve loaded it with a lot of things I’ve written, along with transcriptions of conversations from my podcast, The AI Pulse Podcast. So, it speaks like me, which can be quite amusing—it sometimes feels like I’m talking to myself. I’ve also equipped it with knowledge on topics I care about, such as articles, news items, product brochures, and documentation, making it fairly capable.
However, one of the challenges is validating the answers to ensure accuracy. But it’s definitely a useful tool. Once it becomes more reliable, I plan to make it available internally, so if people have questions they’d usually come to me for, MaxGPT can act as a first line of response.
That’s essentially what MaxGPT is. As for my day-to-day processes, AI—and more specifically, automation—plays a huge role. I think calling it just AI is a bit misleading because it’s really automation in general that’s everywhere, with AI being just one part of that. Looking ahead, with agentic AI and AI agents on the rise, we’re going to see even more possibilities, particularly in personal and ad-hoc automation. The idea of simply telling your computer what you want to automate and letting AI agents figure out how to do it is really exciting.
So, MaxGPT is still in beta, but it’s coming along. I hope to make it available at some point.
Well, all technology tends to follow an S curve, right? It starts with figuring out what it’s useful for, then it becomes very exciting, and eventually, it settles into being quite boring. The reason why everyone thinks ChatGPT, LLMs, and generative AI represent the entirety of AI is because all the earlier forms of AI have already moved into the "boring" phase. We’ve gotten so used to them that they’re no longer exciting.
To put that into perspective, I always come back to the iPhone as an example. When smartphones first came out, like the Nokia Communicator, people wondered what they would even use it for. You’d see things like Kelly Clarkson texting Nelly on Excel in her video clip, and it just seemed impractical. It wasn’t until the iPhone that smartphones became exciting. But today, people aren’t lining up at Apple stores when the iPhone 16 or 16 Pro is released. As a big Apple fan myself, I only upgrade my iPhone every two years now instead of every year. The upgrades are nice, but they don’t seem revolutionary anymore. The iPhone has moved along the S curve—from why do we need this? to this is exciting to this is just another phone, even though the innovations are still significant.
All technology tends to follow an S curve. It starts with figuring out what it’s useful for, then it becomes very exciting, and eventually, it settles into being quite boring.
The same thing is already happening with ChatGPT, LLMs, and generative AI. This doesn’t mean they’re going away. It just means they’ll become part of our everyday lives, embedded into productivity tools and technologies like AI assistants, and we’ll eventually take them for granted. At some point, people will say, This is normal. I’m used to it. And then, the next big thing will come along, and we’ll ask, What does this even do?
Looking ahead to 2025, I think AI agents will be the next big S curve. There’s already a lot of interest and confusion around them. I also think the tech market is pushing this because they see generative AI moving into the "boring" phase faster than expected. So, there’s a new category being created to start the S curve cycle all over again and keep the momentum going.
Will AI continue to have an impact? Absolutely. We’ll keep seeing incredible innovations, especially as the market demands products that solve specific problems. We’ll have this duality where companies like OpenAI, Anthropic (with Claude), Meta, and others push the boundaries of AI exploration, and that cutting-edge technology will trickle down into practical, everyday applications that people use in their daily lives.
That's a great question. Actually, I like to think that I've solved my number one challenge at work, which is the blank slate problem. I believe many of us, across various roles and responsibilities, face the challenge of starting something from scratch. Whether it's describing a new feature or crafting new messaging for a product, the blank slate problem can be daunting.
Thanks to tools like MaxGPT, among others, I’ve been able to address this issue. Now, when I have those early, rudimentary ideas forming in my mind, I can get the stimulation I need to think more deeply about them and approach them from different perspectives. Having the ability to interact with something that challenges and refines your thoughts helps you figure out exactly how you want to describe or present something.
So, that’s definitely been a game-changer for me—solving the blank slate problem, which I think many of us struggle with, has become much easier.
I would definitely say user adoption is key. If you can make your product something that users not only want to use but are also willing to pay for, that’s like the North Star of success metrics in my opinion. I've seen many cool companies doing really innovative things, but often people don’t fully understand the problem they’re solving. Sometimes, people even ask, This is amazing, but what does it actually do for me? So, for me, user adoption would be my North Star.
I think that's something that continuously evolves. As more customers have moved to the cloud, the feedback loop has become much more integrated into the product itself. With a product-led growth (PLG) approach, customers can interact directly with the product, helping themselves, providing feedback, and even receiving support through the product interface.
If you consider all the ways we can now analyze data from product usage and specific features, the entire feedback mechanism can be automated to a higher degree. Especially with AI tools in our toolbox, the user journey can become more exciting and personalized. For example, you could send specific information or activate certain features for particular customers based on what they’re trying to do.
Thanks to the processing power we now have, we can interpret user signals in real time. This means we can help users succeed within the product in real time and even anticipate potential issues before they occur. That’s definitely the level we’re aiming for at ABBYY as well.
I think the build versus buy decision is becoming even more important these days. "Build" can mean a lot of different things now. You can build by taking something readily available in the open-source or research community and developing from there, or if it’s critical to your organization, you might consider buying instead.
The advantage today for many product companies is that there are plenty of offerings across various categories. The key consideration in the build versus buy decision should always be: What is the end result I’m trying to achieve, and how quickly do I need to achieve it? That’s the metric I use to guide the decision-making process.
It’s a complex process, and it's often challenging to clearly communicate the full picture to the organization. You need to ensure that everyone understands that while you might buy a solution, it will still be a part of your product. In today’s world, where so many things are available as APIs or services, a lot of products are simply a combination of various bought solutions that have been integrated into an end-to-end offering. And that’s exactly what the customer wants—they don’t want to deal with the complexity of stitching together multiple solutions; they just want a product that solves their problem.
Ultimately, the decision depends on what you’re trying to achieve, and the answer might differ for each specific case. From my personal perspective, as a product organization, if something exists that fits your needs, buy it. There’s no point in reinventing the wheel.
If something exists that fits your needs, buy it. There’s no point in reinventing the wheel.
From an enterprise customer’s perspective, I would also recommend buying unless you need to build something highly specific for your organization. There’s so much available out there, and people often underestimate the amount of work involved in building something from scratch. For example, building a model that reads documents might sound simple, but once you start, you quickly realize how challenging it really is.
I think this is a really important point, and its significance will only continue to grow for every product out there. What we’re seeing in the market is that customers increasingly enjoy having the flexibility to pick and choose components for their end-to-end solutions from different vendors. For example, they can get storage from Azure, AWS, Google, or even another provider. These services have become somewhat interchangeable because they all offer similar integrations.
Integration capabilities are crucial, especially as we continue to process more and more data from various sources. There’s really no way around it. The ability to easily connect multiple data sources and, increasingly, multiple AI capabilities, and then move that data to the next step in the process or the next system where it needs to be interpreted, is absolutely essential.
In this episode, we delved into the evolving role of AI in product development and its impact on the software industry. We explored how companies are leveraging AI to innovate, and how it’s transforming the way product managers work.
If you want to suggest questions and topics for our next episodes, feel free to share your ideas below!
You’ve just read an interview from our podcast, where we speak with product leaders who share their experiences. Follow us on Spotify or YouTube for more episodes.
.png)
.jpg)
.png){kind=logo}
