Ukrainian firm explores AI agents for accounting and legal workflows

And today, I propose a fairly conversational format, because a lot is being said about artificial intelligence, and I would like to make this as interactive as possible, to analyze some cases where you use AI, what requests and needs for implementation you have, and to answer them, in order to kick off the topic a bit, while I will share some information today that is more about trends and how artificial intelligence is currently being used. I will show a small use case, for example, for accountants, as well as for lawyers, which can simplify your life. But I would like to get the maximum number of questions and information so that we can analyze use cases based on your examples and deliver maximum value to you, so that after the webinar, you can use elements of artificial intelligence a bit better or more effectively. What has changed in terms of AI usage over the last six months to a year is that, with the super-fast growth of models, they have become quite highly autonomous. And working with agents, which seemed irrelevant for most businesses, for example, a year ago. Now it is a quite simple story. And most importantly, simple agents with the help of something like Claude Code or work in Claude or ChatGPT has become quite fast and easy to do. That is, during a simple discussion with a model , you can say: "Remember this sequence of actions." And in this way, we can use these skills that we give the model for the future, but this is a level of personal efficiency. When working, for example, with accounting or legal systems, there is now an opportunity to connect models directly inside the artificial intelligence, well, the AI, right, the chats, directly to ERP systems or accounting systems, and thus move away from working with an interface in the classic way, and perform tasks in an interactive mode. That is, this could be, for example, an accountant's assistant for working with counterparties. This could be a legal assistant that will highlight the contractual framework. Moreover, because models are now directly integrated with some Ukrainian databases, they can quite clearly identify legal violations and process information. Also, what has changed over the last six months is that document recognition and the analytical capabilities for working with these documents have become much higher. And, for example, a year ago, working with scanned copies was quite difficult. Now it is standard practice. And by simply uploading a photo or image into the system, it quite clearly understands what kind of document it is and how to work with it. While I'm opening one of the demos, please write in the chat or feel free to speak up and share your own experience with using artificial intelligence, or what interesting tasks you've encountered. So, let's get to know each other a little. Yes, someone share something. So, uh, Iryna, right? Meaning, not a single artificial intelligence model provides a one-hundred-percent answer, or rather, a one-hundred-percent guarantee for recognition. And, unfortunately, that is a fact that we cannot do anything about. So, this is the last example I would like to bring up; it's figuratively like a car with autopilot. The car drives perfectly, it changes lanes automatically, everything is fine , but the system always asks the driver to keep their hands on the wheel and control the car in an emergency. So, even with our roads, for example, it often happens that a car on autopilot just confuses lanes or drifts onto the shoulder. And this is the norm for artificial intelligence. Because, figuratively, if we have 10 fields recognized at 95%, then for a document to be 100%correctly recognized, it won't be 95%, it will be a combinatorial set. Well, all these failures essentially add up, and the probability of 100%recognition will be, say, 80-60%. But for each individual field, it will be 95%. To combat this, you need to change the process and understand that the responsibility for completing any task still lies with the human. And a classic solution option, uh, for this task is that the system pre-fills the information and then passes it to a human for verification. And verifying 10 pre-filled fields, instead of typing them out by hand, is much faster than typing them by hand and then verifying. That is the first such option. The second option is cross-checks. We are currently doing a project for customs brokers, for example. And in principle, all information regarding vehicles, carriers, and carrier registration numbers is available in additional sources. And so, if we recognized a license plate or a registration number, it can be verified. A license plate can be checked in the Interior Ministry database, a registration number can be checked in YouControl, and in this way, we can distinguish whether the information is correct or incorrect. So , summing up, this format of additional cross-checks, post-processing of data, and human verification is what gives those missing percentages for the solutions. But here, unfortunately, there is no silver bullet. And there will always be a percentage that artificial intelligence processed the information incorrectly. Uh, Oksana, a bit later I will show a live demo of how to connect artificial intelligence to 1C. So, I have noted your question. Ah, Oleg, could you provide more details about the Claude Small Business package and the analysis? I didn't quite understand it. the question. So, Claude packages are divided into three groups. There is the free personal version, and the one for $ 20. The corporate Claude and the developer version. And in principle, they all have more or less the same capabilities . The only thing is that Claude can also be divided into two global blocks. One is the browser version that we use most often, and the desktop version, Claude Computer Use, which can work not only with data uploaded to the Claude system but also with data you grant access to on your PC and your operating system. Very roughly speaking, with your operating system. It can perform certain actions, like going into a browser, searching, opening a new account, and doing some additional tasks. And what does that give you? You can even schedule tasks to be performed by it instead of you. For example, checking email, reading all attached internal documents, and prioritizing that information. Checking the registry of court rulings and updating the registry of exclusions. And there are actually many such tasks that are routine and performed daily. And in this way, you can supplement this information. Okay, one of my demos has opened. It's a short video. It's an integration of artificial intelligence, one of the variants of integrating AI with 1C. And, can you see my screen, I hope? Yes. The essence of the agent is that it helps provide answers to counterparty inquiries on various issues. Reconciliation statements, clarification of invoices, and many other basic questions. And counterparties, especially in retail and service companies, ask these often. And these tasks take up a huge amount of time to resolve. They are all small, but you get distracted for 15 minutes, lose 30 minutes, and that's how time is lost. What does it do? Any counterparty verified by email can send a request to a specialized email address. The AI, understanding this request, classifies it and verifies that it has enough data to perform a specific action. If we look closely, the person wrote a request for a reconciliation statement but didn't specify the date the statement should cover. The agent understood this and asked the person via email, and within the same communication channel, the person provided the details, and the system launched an agent that in 1C generated this document and its Uh, yes, I’ll answer that now. It creates this reconciliation statement and emails the ready document to the person. In this way, in principle, no time is wasted on communication. And only when the system cannot provide an answer does it forward this document directly to the accounting department, where a real person responds to it. And to integrate it, you need to perform several actions . This means connecting an ODATA block and setting up a custom agent. What is unique about integrating all AI models for ERP systems? They are all quite heavily business-oriented. For some, sending a reconciliation statement in such a format is okay. For others, it’s a more complex story because some have multiple legal entities, some don't, are they in one database or are they branched database systems? And response protocols, who we can respond to, who we cannot respond to. And to implement all this, it is necessary to clearly define it. Uh, okay, Oleg, I will, well, I'll take a look a bit later. Alright. Uh, you need to clearly define the business rules. And in fact, when configuring, the main task of the integrator is not to form the technological solution itself, but to go through and define the queries we want to get answers to. And plus, to write out the rules of the game for the artificial intelligence. Why is this important? Because, for example, this is a real case. That is, a year ago I talked about this more like tales from colleagues. Now it's a real case. We had one agent asking for: "You need access, well, access to more extended higher access to the system, because it believed it could not download the document, but it was its hallucination, which had to be dealt with somehow." And therefore, in principle, the main task during configuration is precisely the business-analytical limitation of the model. And upon its implementation, and in order to implement it, it has become technically easier than analytically. And here you need to turn to specialists, because any such integration requires two factors directly. This is the configuration of, say, 1C or the system for granting access to data objects within the system. And separately an analyst who an analyst slash prompt engineer who can directly configure the m