"The problem Tinapse wants to solve touches a point that all financial companies inevitably go through. The discourse on financial AI is moving beyond 'which model is smarter' to 'which model is more trustworthy and won't cause incidents.'"
Kang Min-seung, CEO of the financial artificial intelligence (AI) startup Tinapse, recently described the latest financial AI trends this way. Tinapse, which showcased the excellence of its financial counseling agent in collaboration with KB금융 last month, received the Financial Services Commission (FSC) chairman's award with the highest score among 44 teams. Within six months of its founding, Tinapse raised 4.5 billion won from Mirae Asset and Kakao Ventures, among others, and was the only team from Asia to place in the top ranks at a startup competition hosted by Amazon Web Services (AWS) and NVIDIA.
The Financial Services Commission (FSC) said in May that it would support financial companies' use of AI by easing network separation rules. It added that it is also considering a full lift of the rules for financial firms with advanced security and AI capabilities.
Kang said the FSC's decision was important for AI innovation in Korea and likened Tinapse's solution to a "filter." While each company's AI chatbots supply purified water, adding one more filter can remove any potentially fatal contaminants.
Tinapse is conducting proofs of concept (PoC) for collaboration with several commercial banks and large companies, including KB Kookmin Bank. Kang said, "Banks already recognize the potential for problems in AI answers, and demand is surging. In finance, even if it is a few seconds slower, answers must be accurate." The following is a Q&A with Kang.
―What do you think was the decisive reason you were chosen as the best collaboration case among 44 teams?
"I think we scored high because of the shared recognition that 'an incident could happen to us too.' While explaining the background of the solution, I mentioned a real case at a commercial bank. A chatbot's wrong answer led to an incorrect system process, and after the incident, an employee manually corrected the ledger to resolve it. If the customer had filed a complaint with the Financial Supervisory Service (FSS), it could have led to sanctions.
AI adoption is spreading quickly in banking, but preparation for incorrect answers that AI can produce is insufficient. In finance, even a single wrong character in a code can lead to a major incident."
―How does Tinapse's solution prevent AI from giving wrong answers?
"To keep wrong answers from going out, real-time control is necessary. In areas like finance where one wrong answer becomes an incident, analyzing answers (logs) after the fact, as AI has done so far, is too late. Tinapse's solution compares potentially erroneous answers one more time before sending them with an answer key built on the grammar of finance, that is, ground truth.
However, that standard is not made for general use. You can only know what kinds of errors actually occur in financial work by going on-site, and that is why we have been building finance-specialized ground truth through collaboration with KB."
―Does real-time answer verification slow response speed?
"If we use only a detection model, we keep it to the 200–300 ms (millisecond; one-thousandth of a second) range. But answers that require reasoning rise to the seconds level. Think of using ChatGPT in thinking mode. It's the same structure as a person pausing briefly before a question that requires judgment."
―Why do you leave an audit trail for all of AI's actions and judgments?
"In finance, regulations and incident liability often lead to legal disputes. If you leave an audit trail for AI answers, it is stored in a form with legal effect and becomes the basis for retracing why an answer came out that way and proving responsibility when an incident occurs. As with reviewing a dashcam after a car crash, you can trace the records to find the cause. For using AI in regulated industries, explainability is as important as performance. Even Google is being sued over AI incidents, and our goal is not a smarter AI but trust that makes AI usable for work."
―You scored higher than Gemini or Claude on certain answers.
"Gemini and Claude are far superior in perception and reasoning. But limited to financial knowledge, answers that go through our model come out satisfactorily for the financial sector. Thanks to experience in finance before founding the company, we understood the grammar of finance and the perspective of financial companies and were able to build a model that gives finance-specialized answers."
―There are many PoCs in finance, but few lead to actual adoption. What was different about the collaboration with KB?
"In banking, no matter how important it is, an AI department that does not directly lead to revenue has difficulty gaining clout. But KB금융 has more than 100 professionals at its AI center and is investing heavily in advanced AI. It actively supported collaboration with Tinapse, and we were able to get advice from departments not directly related, such as loans and deposits and pensions, which I think helped create a strong solution."
―Implementing AI itself is an expense, and using a trust verification model adds a second expense. Why is it still indispensable?
"AI increases work speed and efficiency. But the ripple effect of an incident grows accordingly. Setting Tinapse aside, introducing AI without trust verification is like migrating IT systems without a firewall or antivirus. Suppose AI gives an answer that could violate the Financial Consumer Protection Act, an explanation that leads to misselling, or a socially or politically sensitive remark. Considering the brand damage and severe penalties from supervisors, using AI without a safety net is actually the more expensive choice."
―What do you mean when you say the reliability verification model built through Tinapse's collaboration with KB could become a global standard?
"The reliability verification market is just opening. Even among those, no verification model specialized for finance has yet taken root anywhere in the world. Korea is already a market that does not use bills and coins, where finance and IT combined early, and with the Basic AI Act it has a regulatory framework for high-impact areas such as finance and healthcare. There is no better test bed for advancing verification models.
Because of this background, in finance, Korea can produce answers ahead of the United States or Japan. A model validated here can become the global standard. With Korea's financial market and innovation technology, and the traditional finance that supports them, it is entirely achievable."