Typical “tax” mistakes of companies: failure to register corporate taxes on time, incorrect documentation of tax benefits and deductions, incorrect assessment of tax liabilities.
What if it were possible to avoid paying millions of dollars due to tax compliance issues? We talked about the new possibilities of self-adapting IT architectures and predictive models with Feride Osmanova, a leading backend engineer at RegTech and LegalTech startup LOVAT.
“Software based on artificial intelligence (AI) is not what it was 5 years ago and provides almost 100% tax compliance,” the expert said in a conversation with us. Osmanova knows the situation “first-hand”, as she oversaw the development of a platform for automating the fulfillment of tax requirements at LOVAT Compliance.
Among other things, the cloud solution from LOVAT can predict tax rates using neural networks. A function that helps corporate users of the system to build financial planning and risk management much more efficiently.
LOVAT Compliance is a company that provides cloud RegTech services all over the world. The company’s head office is in the UK, but the SaaS services it supports can work with the legislation of different countries.
- Feride, what can AI do today if we are talking about compliance with legal requirements?
- The performance of “neural networks”, the architecture of neural networks and large language models (LLM) have undergone a great evolution and have learned a lot that they could not do five years ago. Top managers of companies that still pay large sums of money from their profits due to compliance errors probably have not yet understood the new capabilities of RegTech platforms on AI.
For example, if robotic automation of reporting for regulators is not news, then not everyone has heard about the success of predictive, that is, predictive models of neural networks. There are studies showing that the accuracy of AI algorithms in predicting which bills will pass parliamentary approval and which will not is 95%. Can you imagine what this means in terms of compliance planning?
Companies can find out long in advance what to expect from officials and legislators! In the experiment I am talking about now, “artificial intelligence” was trained on millions of archived records of tax initiatives.
The test was conducted using the work of the Korean parliament and the US parliament as an example. Such AI models take into account an amazingly large number of data sets and facts. For example, one of the platforms that predicts legislative changes, Westlaw Edge, relies on 250 types of data in making decisions and formulating forecasts: the presence of objections (practice shows that the initiatives that are not controversial are the first to pass parliament), the number of co-authors of bills, the majority in parliament, economic indicators, and much more. This is amazing! One example: AI can even take into account in its assessment how the wording changes (“climate change” vs. “global warming”). Other well-known projects of this kind include startups PredictGov, ScoposLab.
If we talk about what we have been doing and are doing at LOVAT, then RegTech has a number of interesting experiments on forecasting companies’ tax tactics using AI and LLM – large language models (ANN). The predictive accuracy in analyzing corporate tax data and key financial indicators was 92%. Since we now have a fairly good understanding of the factors influencing legislation and the factors that affect the behavior of companies in the context of regulators’ actions, we can successfully predict the occurrence of obligations and even model – this is the next step in the development of such software – a successful line of behavior on the part of the company.
- And how do users of your predictive service, which was created within the LOVAT cloud platform, do they?
- When studying user experience, we found that the use of tools for automating tax reporting reduces the number of errors by 70%. If we are talking about the errors that various RegTech AIs are now making in predicting tax rates, then now it is about 15%. Our indicators in LOVAT are approximately at this level. We have done a lot to ensure that there are as few errors as possible. We have implemented a retry mechanism and logging of unsuccessful operations.
This has helped us reduce the number of unsuccessful requests to the tax authorities’ API by 30%. The lack of alternatives to government services, and on the other hand, the inability to influence their operation, is a big problem for any developer who creates applications based on government APIs. We have integrated RabbitMQ, a queue manager, into the architecture, increasing fault tolerance and efficiency in working with big data. Machine learning of AI, which operates on the LOVAT platform, is carried out using 1,200 legal acts of the US, EU, and UK. We have also applied legal-aware design when developing the architecture.
Special microservice development patterns also help to increase efficiency. They contribute to the accuracy, maintainability of the code, and the stability of the system in using data. And yet, despite all the errors, the neural network also makes. Another thing is that without it, there would be many more of these errors. And now it is impossible to process increasing volumes of data manually or even in a semi-manual mode.
But sometimes even one error in fulfilling regulatory requirements is enough to ruin a company? In the case of a human error, the problem can be quickly localized. But AI will have time to reproduce incorrect decisions a thousand times before it is stopped…
There are approaches to solving this problem. In 2016, the UK became the first country (FCA Regulatory Sandbox) to create the so-called “regulatory sandbox” in fintech. This is a special regulatory regime supported by government agencies so that new software in the field of compliance can be tested with minimal risks.
Now such “regulatory sandboxes” have been created in Singapore (MAS FinTech Regulatory Sandbox), the USA, Russia. You join such a program. You test the software. At the same time, the regulator does not punish you and is ready for mistakes. In return, you carefully report to the regulator.
- And how do “regulatory sandboxes” work? Who can file, for example, taxes in such a mode?
- Both technology companies developing tax software on AI and companies that have decided to switch to such IT solutions, but are afraid of the consequences, can take part in such. Usually, to participate in the “sandbox”, the company must meet a number of requirements: no criminal liability, no bankruptcy cases. An application is submitted, and if you meet all the requirements of the program, then the regulatory authority approves this application.
Experts often recommend that organizations that are going to participate in the “regulatory sandbox” have a clear plan for exiting the experiment in case of failure. Since you must fulfill all the requirements of the relevant program in any outcome. I must say that there are no special “tax sandboxes” anywhere yet. Usually, the functionality of test work with the fulfillment of certain tax requirements is present in fintech sandboxes.
- What did you discover for yourself as a back-end developer of RegTech applications when you were creating a neural network solution for automating tax compliance?
- First of all, the discovery for me was the enormous benefit of queue managers, the same RabbitMQ in projects with similar loads for working with big data (we had about 100 thousand transactions per month). Otherwise, I got the most out of Python, its excellent frameworks for working with neural network architectures and Postgres for maintaining databases. If we talk about frameworks for Python, then the project used a lot of integration with third-party APIs.
Commissions and taxes were calculated using information from there. The FastAPI and Django libraries turned out to be very useful for this.
- Are there periodic disputes about the responsibility of RegTech platform developers for errors made by neural networks? How relevant is this issue now, in your opinion?
- Yes, the issue of subsidiary liability of AI developers in the field of compliance is periodically raised. But unlike medicine, unmanned transport, regulators look at the issue more unambiguously. If you use software that you have chosen yourself and at the same time there is a discrepancy with mandatory requirements, then you did not take measures to ensure that the requirement is met. You are to blame.
Despite this problem, I think that RegTech tools will continue to be used more and more in business because without them, I repeat, there are not fewer, but more errors. Yes, the staff of specialists needed to work with requirements, the number of which in the tax system and in general is only increasing, is unavailable not only to medium-sized businesses, but also to large companies.
- What do companies get as a result of effective AI tools that predict tax rates?
- There are many advantages for businesses. It is possible to assess the impact of changes on the economy and business as a whole in advance. It is better to build your own strategy in these conditions. It is possible to achieve proactive compliance with regulatory requirements.
That is, you are not only trying to keep up with all the requirements that are changing before your eyes, but you have a little time to quickly fix something if there is a problem. If we delve into the details of strategic planning, then with innovative tools, the establishment of companies can more flexibly and thoughtfully build an investment strategy, manage cash flows, and prevent cash gaps. The budgeting process becomes an order of magnitude more efficient.
In an era when companies compete not in who has automated business processes, including tax compliance processes, but in who has achieved the deepest degree of automation, Ferida Osmanova and other practitioners of neural network architectures turn out to be exactly the people who make their companies winners in the competitive struggle. What does the future hold for regulators? Nobody knows. Any new discovery in the field of “artificial intelligence” can change the landscape of this industry in the most unexpected way. And the only constant in the changing world of technology is expertise.
