The Pulse
Halluminate Raises $30 Million to Train AI for Financial Work
Halluminate raised $30 million in a Series A led by Oak HC/FT to build financial-work benchmarks and AI training environments. Its Westworld Finance Diligence Bench tests agents on 88 tasks modeled on a full company acquisition review.
AI.info Team · Written with AI ·
“When the agent starts getting into long horizon work, testing work and specialization will really be key.”
Matt Streisfeld, general partner at Oak HC/FT
Oak HC/FT leads Halluminate’s $30 million round
Halluminate announced its $30 million Series A on October 1, 2026, with Oak HC/FT leading the round. Y Combinator, Orange Collective, FT Partners and Heavybit also participated, bringing the company’s total capital raised to $38.5 million, according to the company.
The San Francisco startup builds benchmarks and reinforcement-learning environments for AI systems doing knowledge work. Its first focus is financial services, where tasks such as building models, reviewing transaction documents and preparing client presentations demand more than answering isolated questions. Halluminate says it is also expanding into adjacent fields.
Streisfeld’s comment, reported by Fortune, points to the investment thesis: as agents take on longer workflows, investors and model developers need ways to test how they perform across a chain of tasks, not just on a single prompt. Halluminate aims to identify where agents fail and build simulated training environments around those gaps.
An 88-task test follows a deal from analysis to close
Halluminate’s Westworld Finance Diligence Bench, released in August, covers 88 problems drawn from anonymized data from real private transactions. The company says finance professionals wrote and reviewed the tasks, which span areas including financial modeling, underwriting, quality-of-earnings analysis, diligence review, pitch materials and closing work.
Rather than confining an agent to a question-and-answer interface, the benchmark places it in a dynamic desktop environment. A task may require an agent to work across files, office applications, a data room, email and chat; some trajectories run to hundreds of steps. Each problem receives a weighted score combining agentic rubric grading with deterministic checks, according to Halluminate’s benchmark description.
One example asks an agent to redline a statement of work using a 160-file data room, 21 emails spread across nine threads and four meeting notes. The agent must track changing deal terms while preserving provisions that have not changed. Fortune reported that seven frontier models took on the benchmark and that the highest average score was 51%.
Revenue claims and a narrow customer focus
Halluminate says it went from no revenue to a mid-eight-figure revenue run rate in the 10 months before the announcement while remaining strongly profitable. It also says it built and scaled reinforcement-learning environments with four of the five leading closed-source U.S. AI labs. Fortune reported that CEO Jerry Wu described those labs as paying customers and said the revenue figure was annualized from quarterly work already delivered and paid for.
Wu told Fortune the company is concentrating on frontier model labs rather than broadly pursuing enterprise customers for now. That approach gives Halluminate a direct link between its evaluation work and the labs training agents, but it also ties its near-term business to a small set of buyers.
The benchmark measures a difficult slice of office work
Westworld’s whole-process format addresses a real weakness in narrower tests: a model can perform well on a discrete finance question yet lose track of instructions across documents and steps. In the sample contract task, the challenge is not just editing a file; it is reconciling the latest terms across a changing record and producing the right final document.
The 51% result is a score on Halluminate’s benchmark, not a general measure of how well AI performs in finance jobs. The company created the test and its scoring system, and the published tasks represent a specific slice of deal work. For now, the round backs a clear bet: that realistic, specialized tests and training environments can help labs find and address failures before agents take on longer financial workflows.