Times Radio - CEO of the Week interview with Ulf Persson, ABBYY CEO (audio, subtitles)

Well, you're right, you're spot on. It's intelligent process automation, a very dry way of explaining how you can find, classify, and extract information from vast amounts of data. It could be in a business document, could be an invoice, could be a driver's license, could be a pay slip, whatever, and you put it into a process where it needs to go. That could be a decision-making process, it could be a payment process, it could be a claims management process or whatever it is. But we expect effectively we are releasing information from, or data from, we're setting it free. Exactly.

It's a good point. I mean, AI for us is a contributing technology or set of technologies on how we do this. Okay? And AI has been around for a long time, for decades now. It's not been as topical as it is right now, and we'll get back to that, but we've used AI for probably 20 years to extract and classify information.

I wouldn't think so, I wouldn't think so. I think AI has been around, but what most people refer to when they talk about AI these days is what we called generative AI, where you have large amounts of data being processed by ever increasingly difficult algorithms and models to extract and understand information, and then create information. So you can create text, you can create music, you can create pictures, you can create movies, as we've all seen. That's what people are generally referring to. But as I said, parts of the AI tech stack have been around for a long time.

I think it has been to a lesser extent consumer facing, so we may not have seen it as consumers. But behind, under the bonnet of some of the tools that we're using, we've definitely had AI. And as I said, finding information, extracting information from business documents has been done with the help of AI for quite a long time.

No, I think you're absolutely right. There is a general concern, and I think that concern is more about what we call the general purpose models. And then you have more specifically oriented models that are trained on data that is very specific for, it could be legal documents, it could be sports, it could be something else, where you can much easier control what comes in, and that way you can also control what goes out. In the general models, it is indeed difficult. You have all heard of the word 'hallucination' – that the model will throw out something that doesn't exist, and that is in turn very difficult to audit. There we have a big challenge.

I think what is happening right before our eyes is we have a technology that is developing at breakneck speed, and we have a regulatory and audit framework that has not caught up in the same vein. And if you think about it, it's quite natural. The regulators are typically governmental, and governments don't move as quickly as the entrepreneurial world. And we're seeing an increasing pace of development in AI, and we're not seeing that increasing pace of development in regulatory frameworks. The United States is taking a slightly different, less hands-on approach, and then you have China who have said that in 20 years we're going to be the world's most dominant power in AI, and that's probably also through a very carefully crafted regulatory framework. But the interesting thing is that AI is global. We can have a European framework, and you can have a US framework, and you can have an Asian framework, but the companies you mentioned, they all work in all markets, and the data doesn't know whether it's from China or it is from Sweden or it is from the UK or it is from America. At the end of the day, what we need is a very global set of policies and a global charter of how we deal with AI as companies that are on the production side but also on the consumer side.

Interestingly, we were talking about that in the waiting room, and no, we haven't. We haven't really. We all see the problems with social media. You tend to be in an echo chamber. If you believe a certain set of beliefs, you listen to and you follow people who have the same beliefs, and you are then through the algorithms you are recommended people of the same view, so you don't get the breadth of information and you don't get challenged. Same with Spotify, although I think that's probably better now. But you listen to David Bowie, and then you get Bowie-like songs, so you get people who are in that orbit rather than something that's completely different that could teach you to look elsewhere.

No, no, you can. You can do that. Okay.

As I said, the focus of what we're doing is really extracting information from documents and business documents and making that available in further business processes. But the interesting thing is, with the advent of AI, what used to be fairly backend-oriented business is now much more on the front end. So if you are a consumer looking at a bank loan application, or you're looking at a claims process, an insurance claims process, instead of compiling all the documents and taking photos, printing them out, and sending them in an envelope to your insurance company or broker, or you go to the bank for a bank loan and putting together all the documents you have there, you now do it with your phone. So effectively, you have your phone, you scan your pay slip, you scan your passport or ID card, you scan something that shows your address where you live, and that gets into your loan process, and the relevant information is extracted and used. And better still, if on the back end you have an AI model that can control all the information that you have given them, but also run that against other data that is linked to your person or your address or your social security number or your credit history, you can make that a very easy decision. Rather than having days or weeks going into this, you can do it in an hour. And when the answer is yes, if you're applying for that bank loan or that bank account on your phone, which you can do, you can just download an app and have a bank account within an hour.

Maybe you wouldn't, but I think there are ways to always improve the models. There's the importance of the regulatory framework, how you can audit and understand the actual underlying models, how you understand the data that goes into it. So all of that has to happen, and to a large extent it's already happening. But I think if we can do that for 80, 90% of the cases, that leaves time and availability for the remaining 10 to 20%, which may be a little bit more difficult, and you can have human beings deal with that. That must be much better than having human beings deal with everything and not having the time to dig deeper into the more complex cases.

You have to use AI and automation and technology where it helps you, but it shouldn't really make the service worse. It should make the service better.

That is true, but you imagine 10 years ago you couldn't do that, or five years – it was much worse. Where are we in 20 years? Where are we in 5 years? I think that the quality of the models is going to develop. I think that we will sometime, maybe in five or 10 years, we won't even see a difference between whether we talk to a person or we talk to a robot. One thing that I wanted to point out – there's something that we call at Abby, and I know others do it as well, it's called AI with a purpose. It's not AI for the sake of AI, it's AI for the sake of improving people's lives, or improving the outcome of what you're trying to do. The application of AI must make sense from the user's perspective, and you were questioning some cases where it might not, right? But it must also make financial sense. So both for the user and for the provider of the service, it should make financial sense. And thirdly, it must be safe. So the results must be of good quality, they should be auditable and consistent and unbiased. Most of our large customers, they are now asking us, when we build our own models, and saying, 'Where do you get the data from? Can we audit it? How do we make sure that there's no bias?' So the consumers, the users of the output from those models, are very aware of that. And I think that's a great sign.

I love pudding.

I will try some pudding with you.