Benefit Street Partners gets a first-pass read on a 300-page credit agreement in about 20 minutes.

A Portfolio Manager Who Builds His Own Tools

Benefit Street Partners is a credit-focused investment manager. Anant Kumar, a portfolio manager at the firm, built two internal tools using Codex, a coding assistant from OpenAI. GenAI credit agreement analysis is the second of them, and it runs in the firm’s own hosted portal.

“I’ve gone from being just an analyst and a portfolio manager to being a developer as well now. I can build things,” Kumar said.

What GenAI Credit Agreement Analysis Does

Credit agreements often run to 300 pages or more. The tool takes an uploaded agreement and produces a first-pass analysis, flagging clauses that could shift value away from lenders. That pass takes about 20 minutes. A person then spends an hour or more reviewing the output. The firm later refactored the tool with GPT-6 Astra.

The other tool tracks loan valuations. Several lenders can hold the same loan and put different values on it. Business development companies disclose those valuations in filings to the US Securities and Exchange Commission. Therefore the tracker pulls the filings and compares them, so the team can see how its own marks sit against other holders’. Kumar built the first version in about a day, and total development time came to roughly 10 to 12 hours including later additions.

The Result

Committee follow-up questions used to take days or weeks. Now they turn around within a day, and are often answered during the meeting itself. However, the firm is not using that speed to do more deals. Instead it adds review capacity and rigour inside the same timeframe.

“The quality of my analysis has become so much better and so much faster,” Kumar said.

GenAI credit agreement analysis of this kind sits inside the work rather than alongside it. The person who knows what a costly clause looks like is also the person who built the tool. Other financial firms have compressed analysis the same way: GoTyme dropped root cause analysis to 10 minutes.

Why It Matters

  • When the person with the domain knowledge also builds the tool, the specification problem largely disappears. For narrow, judgement-heavy analysis, that matters more than the choice of model.
  • A 10 to 12 hour build changes the question. At that cost, the test is no longer whether a tool justifies a project, but whether someone has an afternoon.
  • First-pass output still needed an hour or more of human review. Therefore budget for the review, because the saving sits in the first pass and not in the sign-off.
  • Speed can be taken as extra capacity or as extra volume. Decide which before deploying, since the two choices produce very different numbers at year end.