Benefit Street Partners uses Codex to answer committee questions in a day

OpenAI published a customer story about Benefit Street Partners (BSP), a credit-focused investment manager. Portfolio manager Anant Kumar says answering an investment committee's follow-up questions used to take days or weeks. Analysts can now turn a round of comments in a day, and often answer during the meeting itself.

Kumar built two tools. The first is a tracker that uses public SEC filings from business development companies (BDCs) to compare how different firms value the same loan, so the team can see how its marks stack up. He used Codex for code review. The first version took a day, and he puts the total at 10 to 12 hours including later refinements.

The second tool analyzes credit agreements, which are often 300 pages or longer. It's built on memos from leading law firms about clauses and risks to watch for. An uploaded agreement takes about 20 minutes to produce a first-pass analysis, followed by an hour or more of human review. BSP has refactored it with GPT-6 Astra, and it now runs in the company's own portal. OpenAI says BSP uses the tools to add review rigor in the same timeframe, not to do more deals.

My take

The part I find most useful is not the speed. It's that a domain expert built his own tools. Kumar says he went from analyst and portfolio manager to "a developer as well." The person who knows what a dangerous clause looks like is the person shaping the tool. That usually beats a developer guessing at requirements.

I also like that the story keeps a human in the loop. Twenty minutes of machine work, then an hour or more of human review, is a sensible split. Kumar himself says the final investment decision has to be made by a human. If you build something like this for legal or finance work, copy that design: treat the output as a first pass, never a verdict.

Now the scepticism. This is a vendor's own case study, so the numbers are Kumar's estimates, not an independent measurement. It doesn't say how errors are caught, how accurate the analyzer is, or how the old process compares in cost. "Code review" is also a modest claim. It's Codex checking the code, not proof the tools are correct. And the 10 to 12 hours doesn't count the knowledge he had already gathered, like those law firm memos.

My suggestion if you're building something similar: start with one narrow, boring workflow, like a tracker built on public data, and write down how you'll verify the output before anyone relies on it. I'd want to see error rates before trusting any of these figures.

Sources

• OpenAI News

This post was written with AI assistance from the sources above.