Kepler, a financial research startup, has built a highly reliable AI system that passes rigorous audits and regulations by placing the AI model Claude within a strict control system rather than using it as a standalone tool.
Imagine you are the head of a global investment bank managing tens of billions of dollars in client assets. Every morning, your desk is piled high with thousands of pages of financial statements and thick investment reports issued by dozens of companies. To analyze this vast amount of data, you’ve introduced the latest AI assistant, touted as the smartest in the world.
Early in the morning, you instruct the AI: “Summarize the real debt ratio for Company A for the fourth quarter of last year.” In just three seconds, the AI produces a report complete with neat tables and graphs. You firmly believe those numbers and hit the investment button for a massive amount of money.
But a few days later, it’s revealed that those numbers were ‘plausible lies’—hallucinations created because the AI misunderstood the context. The company suffered astronomical losses, and financial authorities immediately launched an investigation. To the investigator’s sharp question, “Why did you make such a fatal decision?”, could you possibly answer, “Because the AI analyzed it that way”?
Absolutely not. This is precisely why Wall Street and the global financial sector have been hesitant to open their wallets despite the incredible capabilities of Large Language Models (LLMs—AI technology that learns from massive amounts of text to understand and generate sentences like a human). In the financial industry, where a single small calculation error can lead to fatal financial loss, handing over decision-making power to an AI with an opaque judgment process is akin to ‘dangerous gambling’—driving down a highway at 200 km/h while blindfolded.
However, a groundbreaking case has recently emerged that calms this deep-seated anxiety in the iron-walled financial sector. It is the story of Kepler, a research-based startup focusing on building a ‘Verifiable AI’ platform for trusted financial services [Source Title]. How exactly did they capture the hearts of the conservative and demanding financial industry?
Why It Matters
You might have seen chatbot AI used for fun in daily life occasionally say something absurd, like King Sejong throwing a MacBook. This is technically called ‘Hallucination.’ While this is just a funny anecdote in casual conversation or light writing, in the financial world—where not even a 1-cent error is allowed—it’s a major accident that can determine the survival of a company.
Financial service companies operate in a regulatory environment that is stricter and more stifling than one might imagine. In particular, to satisfy the guidelines and regulations of powerful financial authorities like Brazil’s Central Bank (Banco Central) or the Securities and Exchange Commission (CVM), every financial judgment and decision made by a company must have clear grounds and be capable of back-tracing at any time [Source Title].
Simply saying “the final calculation is correct” is far from enough. ‘Auditability’—the ability to prove “what data was used, what calculation formula was applied, and why this result came out,” much like writing out the steps of a solution in a math exam—is the lifeblood of the industry.
In this thin-ice financial environment that must meet the world’s most meticulous regulatory and audit requirements, Kepler has completed its own unique and secure AI platform using Anthropic’s AI model ‘Claude’ as its core brain [Source Title]. Their success holds significant meaning for both the IT and financial sectors, as it provides the most realistic and exemplary answer to the fundamental question: “Can the financial industry fully trust and use fickle AI?”
Easy Understanding: Putting an Off-Road Vehicle on Train Tracks
So, how did Kepler transform a brilliant but occasionally mistake-prone AI into a ‘trusted financial expert who never lies’? The amazing secret lies not in focusing on further growing the AI’s intelligence, but in changing the very philosophy of how to handle AI by 180 degrees.
The Kepler team, through in-depth case studies with Anthropic (the developers of Claude), drew and shared a crucial conclusion: “In the financial field, an AI model must never be allowed to become the entire system itself on its own” [Source Title]. In other words, they realized that unpredictable AI should not be given infinite freedom.
To solve this fatal problem, Kepler built a solid ‘Deterministic Infrastructure’ that tightly surrounds the artificial intelligence. Simply put, this infrastructure acts as a powerful ‘shield of trust and verification (Layer)’ that prevents the AI from getting distracted or acting unexpectedly [Source Title].
The difficult concept of ‘deterministic infrastructure’ that experts talk about can be explained with this analogy: Typical AI technology is like a high-performance ‘off-road vehicle’ that can run freely in any direction—mountains or fields—without a set destination. It’s fast and powerful, but there’s always a risk of it falling off a cliff if control slips even slightly. On the other hand, the deterministic infrastructure built by Kepler is like taking the rubber tires off this high-performance vehicle and placing it on solid ‘steel rails’ with a fixed destination. The engine power of the AI (its excellent language processing and document analysis capabilities) is still used, but the directions it can move and where it must stop are perfectly restricted by strict, deterministic rules created by humans.
Easy Understanding: Teaching a Smart Part-Timer How to Say ‘I Don’t Know’
Kepler went a step further than just restricting the AI’s path. Instead of blindly throwing “analyze all 100 pages of this thick financial statement and give a conclusion,” they broke the work down into very small pieces and assigned only ‘precisely defined tasks’ one by one.
Furthermore, they provided the AI in advance with complex specialized knowledge and glossaries from the systematized financial field, minimizing the area of uncertainty where the AI has to judge for itself. The most impressive and core part here is that they established ‘hard boundaries’ between problems the AI should ‘Resolve’ on its own and problems it deems beyond its capability and must ‘Escalate’ to human experts [Source Title].
Let’s explain this with an everyday situation as well. Suppose you’ve hired a genius part-timer (AI) with unparalleled mental arithmetic skills for a bank’s loan counter. This employee is a genius at calculating numbers, but doesn’t know the subtle nuances of complex financial laws or a customer’s hidden intentions. In this case, no manager would fully delegate all loan screening authority to this new employee.
Instead, the manager gives a strict manual: “You handle customer identity verification and simple number calculations perfectly (Resolve), but if you see even a slightly suspicious forged document or if a large loan approval request over $1,000 comes in, don’t judge it yourself and always pass the documents to me (Escalate).” This is exactly how Kepler handles Claude. Not letting it judge everything alone just because it’s smart. Paradoxically, making it possible for the AI to clearly draw a line and say, “I don’t know from here on. Human expert, please help,” is the biggest secret to making the financial sector trust AI 100%.
Current Situation: The Tenacity to Read Even Complex Footnotes
Equipped with such a tight leash and clear guidelines, Claude is showing off truly incredible capabilities within Kepler’s control system. What the fierce financial industry currently requires in practice is not simple news summaries or greeting writing.
The current situation can be understood most accurately by listening to the voices of experts working on the front lines, specifically regarding what practical purpose Kepler’s platform was built for. One expert recently emphasized: “Right now, every AI conversation in the financial sector is focused on ‘Capability.’ Can the model properly handle complex multi-step analysis? Can it meticulously read even the footnotes (supplementary explanations written small at the bottom of the text) tucked away in corners of documents? That is exactly why we built Kepler.” [Source Title]
Most people tend to skip them, but in financial documents where investment and loan decisions for massive funds are made, footnotes can be either a fatal poison or a golden key. For example, a corporate report might state in large letters that “operating profit increased by $10 million this year,” but a tiny footnote at the very bottom might say, “However, this is a one-time payment from the sale of factory land, unrelated to the core business.” This is an area humans easily miss when tired.
In the past, poorly trained AIs often ignored the importance of such fine details and just grasped the overall mood. However, Claude within Kepler, given a clear mission and a thorough step-by-step control system, is accurately identifying the meaning of hidden footnotes and their impact on the overall financial status, as well as complex financial reasoning that requires persistent multiple steps. In effect, they’ve reduced human error and perfectly blocked AI weaknesses.
What’s Next?
Kepler’s achievement will not end as a one-time success story of a single fintech startup with good technology. The design architecture of the ‘Verifiable AI’ platform they’ve sturdily demonstrated will bring a massive wave of change to numerous industries that have been hesitant to introduce AI.
Particularly in ‘high-risk, high-regulation’ industries—not just finance, but also medicine, law, defense, and pharmaceuticals—where a single small error can cost lives or lead directly to massive property damage, Kepler’s method will serve as an excellent textbook and blueprint. Medical AI, which must go through thorough logical verification steps to prevent misdiagnosis, or legal AI, which must control hallucinations to 0% when analyzing tens of thousands of legal precedents, will all adopt this ‘deterministic infrastructure.’
Until now, we’ve been enthusiastic only about how much more human-like and smart AI models themselves are becoming and how much more data they are learning. However, Kepler has clearly proven that true industrial innovation depends on ‘how to put that superior intelligence into a safe and controllable basket and apply it to the unstable real world.’
In the future, the competition arena for global AI companies will change completely. The core paradigm will rapidly shift from “who can make an AI that writes poems or novels like a human” to “who can make an AI system that can defend itself with perfect supporting formulas against the sharp questions of meticulous government auditors.”
MindTickleBytes AI’s Perspective
The development of technology often looks dizzying and precarious, like a high-end sports car without brakes. When you’re obsessed with speed under the name of innovation, it’s easy to miss the most important seatbelt: trust.
However, Kepler’s case shows very well that the true value and explosive power of AI bloom not in ‘infinite freedom and autonomy,’ but paradoxically in ‘sophisticated control and clear setting of limits.’ An AI that runs wild alone, claiming it will perfectly replace human judgment, can never cross the thick wall of regulation.
Instead, a transparent and verifiable tool that quietly supplements human inescapable weaknesses (lack of time, decreased physical strength, loss of concentration in front of vast data) within strict rules and fences meticulously designed by humans. That will be the most ideal and safe future AI model we will encounter in reality. The final hurdle for artificial intelligence to truly permeate our lives’ core infrastructure is, after all, not the height of its ‘intelligence,’ but the depth of its ‘trust.’
References
- How Kepler built verifiable AI for financial services with…
- New frontiers in AI and finance: Kepler built verifiable AI fon…
- Kepler Builds Verifiable AI for Finance With Claude - GogoAI News
- Trending Now: How Kepler built verifiable AI for financial services…
- How Kepler built verifiable AI for financial services… — RadarTrend
- Earlier today, Anthropic published a profile on Kepler and what we’re…
- An autonomous AI that makes all decisions on its own
- Claude operating on a deterministic infrastructure
- A language model designed to evade audits
- It has learned more internet humor.
- It strictly distinguishes between problems it should solve and those it should escalate to humans.
- It automatically ignores all regulator inspections.
- Creativity in sentence generation
- Frequency of buzzword usage
- Auditability and regulatory compliance