Terminal X, a startup that builds AI agents for investment firms, says its quality-checking system catches about 25% more recommendations that stray from a fund's mandate when it runs on OpenAI's GPT-6.1 Sol than on the next-best model. The details come from an OpenAI customer story, so treat the numbers as the companies' own claims.
The pitch is aimed at a real problem in finance: the judgment behind each deal often sits in the heads of a few senior people. When they leave, Terminal X says, a firm keeps the documents but loses the reasoning. The company calls that "memory debt."
How the agents get built
Terminal X sends forward-deployed engineers, described as former finance professionals from JPMorgan and Blackstone, to sit with a customer's investment teams for weeks. They study the firm's trusted memos and capture steps nobody wrote down, such as how a credit desk screens a borrower.
Those steps become hundreds of agents, split by asset class and strategy, with sub-agents for each investment stage. The firm owns them, and they can run on-premise or in the firm's own cloud. GPT-6.1 Sol handles the judgment-heavy stages, like forming a thesis and checking fit with the mandate. The cheaper GPT-6 Luna covers high-volume steps.
What the agents learn is stored as a versioned "Memory" of how one desk makes one kind of decision. Every production query is auditable, with the steps, sources, model and reasoning on record. The agents read English, Korean, Japanese and Chinese documents.
The Eval Agent, and what it claims
The checker is called the Eval Agent. It reads each output the way a VP reads an analyst's work, scoring pass or fail against a 54-criterion rubric. It asks whether numbers, comps and quotes came from the right sources, and whether the recommendation ties to the house view. Each fail gets a one-line reason and a fix, which becomes a new Memory or test case.
On GPT-6.1 Sol, the company says the Eval Agent ran at roughly 30% lower cost per query than the next-best model. One private equity customer reported $2.4 million in annualized savings from automated due diligence and memo work.
The headline figure, 88.6% matching senior investor judgment, appears in the post's title. The text we could read does not explain how it was measured, which model was the comparison, or how many reviews were involved. We also don't know which customer saved the $2.4 million.
What's missing from the story
The OpenAI post is the only source here. We did not get comment from Terminal X's chief executive on how its agents learn from a customer's decisions, and no customer has discussed the results with us. No independent finance technologist or AI researcher has yet weighed in on whether AI grading AI is sound, or whether "memory debt" is a real market need. Until that happens, the claims above stand as assertions.
What to watch
The useful test is independent: do firms that adopt this see fewer bad calls, not just cheaper memos? If you work in finance, would you trust an AI to grade another AI's homework?
Sources
Reported by the WattsUpNext desk from the sources linked below. Spot an error? Tell us at corrections@wattsupnext.com.
The WattsUpNext Brief
Get stories like this in your inbox.
One email each weekday, only the topics you choose. Real news, sourced, no fluff. Unsubscribe in one click.
