Kepler's Verifiable AI Aims to Solve Finance's Trust Problem

NEW YORK, NY – May 21, 2026 – While artificial intelligence has rapidly transformed numerous industries, the world of high-stakes finance has remained a cautious holdout. The risk of a single error—a "hallucinated" number or a flawed assumption—is too great when billions of dollars are on the line. Now, New York-based Kepler is gaining attention for an architecture that may finally bridge this trust gap, a development recently spotlighted by AI leader Anthropic.

Kepler’s platform for financial research is built on a simple yet powerful thesis: language models alone are not enough. By combining the interpretive power of AI with the rigid reliability of deterministic code, the company has created what it calls "verifiable AI," designed to meet the exacting standards of investment committees and regulatory audits.

The High Cost of 'Close Enough'

For years, the financial sector has grappled with a significant AI dilemma. On one hand, traditional financial software is often rigid and slow, unable to adapt to the fluid, complex questions analysts need to ask. On the other hand, large language models (LLMs), despite their impressive flexibility, have a critical flaw for finance: they can be unreliable.

Without a robust system of checks and balances, LLMs can invent figures, misinterpret context, or quietly drop critical constraints during a multi-step analysis. As Kepler CEO Vinoo Ganesh noted, "One wrong assumption early in a financial analysis breaks everything downstream." This inherent unpredictability has made widespread adoption a non-starter in an industry where every number must be defensible.

This challenge has left financial institutions in a difficult position, watching other sectors leverage AI for massive efficiency gains while they remain constrained by the need for absolute accuracy and auditability. Regulatory bodies like the Securities and Exchange Commission (SEC) and the Financial Industry Regulatory Authority (FINRA) have amplified this caution, emphasizing that firms are responsible for the outputs of their AI systems and must ensure they are explainable, fair, and auditable. The "black box" problem, where even the creators of an AI cannot fully explain its reasoning, is simply unacceptable in a regulated financial context.

An Architecture of Trust

Kepler's solution, detailed in a recent customer profile by Anthropic, tackles this problem by re-architecting the workflow. Instead of asking a single AI model to do everything, Kepler splits the work between components that play to their strengths.

The process begins when an analyst poses a question in natural language. An LLM, specifically Anthropic's Claude, interprets the query and decomposes it into a logical plan. This is where the LLM's flexibility shines. However, when it comes to the critical tasks of data retrieval and calculation, the LLM steps aside.

At this point, deterministic, verifiable code takes over. This code retrieves the precise data from Kepler’s indexed library—which includes over 26 million SEC filings and 50 million other public documents—and performs the necessary calculations. Mediating this entire process is Kepler’s proprietary financial ontology, a sophisticated dictionary that maps the language analysts use, such as "EBITDA" or "segment revenue," to the exact line items in the underlying source documents.

The result is a system where the AI model never invents a number. Every figure in a final report is generated by explicit, reproducible code. The LLM's role is confined to interpretation and generating the final narrative around these verified figures. This specialized approach yields dramatic results; Kepler reports its models can identify the correct line item in a 10-K filing with 94 percent accuracy, compared to the 38 to 46 percent accuracy of frontier models operating alone on the same task.

Reliability Over Raw Power

In the race for AI supremacy, benchmark scores often dominate headlines. Yet, Kepler’s selection of Anthropic’s Claude model was driven by a different, more practical metric: consistency.

"On our workloads, Claude was the model that consistently held the plan together," said Ganesh. "Other models would start strong and then quietly drop a constraint by step five." This subtle but critical failure mode is precisely the kind of error that can invalidate an entire financial analysis. For Kepler, a model’s ability to adhere to a multi-step logical plan without deviation was more important than any other performance benchmark.

Dr. John McRaven, Kepler's CTO, framed this approach as moving beyond simple prompt engineering. “In finance, the model can’t be the whole system,” he stated. “We treat it as one stage in a pipeline whose job is to hand the model exactly what it needs to succeed at exactly that stage.” This philosophy of "content engineering"—optimizing the entire system around the AI call—ensures that the model's creative and interpretive strengths are leveraged while its potential for error is contained.

The Augmented Analyst in a Regulated World

For the buy-side analysts at the private equity firms, hedge funds, and investment banks using the platform, this architecture translates into tangible daily benefits that directly address regulatory pressures. The platform's core feature is its end-to-end auditability.

Every figure presented in a Kepler-generated answer is interactive. An analyst can click on a number and instantly trace it back to the source filing, page, and line item. This "click-through" verification is a game-changer, transforming the arduous process of sourcing data for an investment committee memo or a compliance review from hours of manual work into a single click.

Furthermore, because all calculations are explicit and run by deterministic code, the answers are reproducible. Running the same query twice yields the exact same result, eliminating the variability that can plague purely LLM-driven systems and providing a stable foundation for analysis. This directly answers the call from regulators for explainable and auditable AI (XAI). Instead of trying to peer inside a "black box," an auditor can review the explicit formulas and trace the data provenance, creating a clear and defensible audit trail.

Built by a team of ex-Palantir engineers accustomed to creating data systems for demanding organizations, and backed by founders from OpenAI and Meta AI Research, Kepler was developed in under three months based on interviews with over 140 financial firms. The company is already planning to extend its verifiable architecture into adjacent markets, with private credit as its next target, signaling a belief that this model for trustworthy AI can become the standard for regulated industries.