How Benefit Street Partners uses Codex to answer investment questions faster

OpenAI describes how Benefit Street Partners, a credit-focused investment manager, uses Codex to shorten parts of its analysis workflow. This is not a story about AI making investment decisions. The reported benefit is more capacity to prepare, synthesize, and check information within the same deadline, while accountable people retain the final decision.

Table of contents

What the source reports

In credit underwriting, follow-up questions from an investment committee could previously take days or weeks to resolve. BSP says its analysts can now turn a round of comments in a day and sometimes answer questions during the meeting. The stated goal is greater rigor in underwriting, not simply doing more deals.

One internal tool is a dashboard using public SEC filings from business development companies. It compares how firms value the same loan. Portfolio manager Anant Kumar says the first version took one day, and later additions brought total development time to roughly 10–12 hours. Codex helped draft code and review it for errors.

A narrow use case: credit agreements

The second tool analyzes credit agreements, documents that often exceed 300 pages. It helps surface clauses that may move value away from lenders. BSP combined memos from leading law firms about clauses and risks to watch, then used that material to inform a first-pass analysis. An uploaded agreement takes about 20 minutes for that first pass, followed by an hour or more of human review. BSP says the tool was refactored with GPT-6 Astra and is live in a company-hosted portal.

Why this matters

The example shows where AI can be useful: a bounded workflow with clear inputs and expert oversight. A model can collect data, inspect long documents, compare patterns, and produce a draft. It cannot independently establish that a clause applies in every situation, replace counsel, or approve risk. Faster work also makes source traceability, result checking, and access control more important.

Five implementation lessons

Start with a bounded task: trackers, clause extraction, and recurring questions are safer pilots than automated decisions.
Control the source set: identify permitted documents, authoritative versions, and data refresh dates.
Design human review: name the reviewer, expected review time, and conditions requiring a check against original text.
Protect sensitive data: contracts and client material need classification, access control, audit logs, and retention rules.
Measure quality: compare omissions, handling time, rework, and confidence before scaling.

Practical checklist

Pilot one document type and a defined question set.
Require results to point to pages, clauses, or source records.
Do not give the model authority to send memos, alter records, or execute transactions.
Test difficult documents, exceptions, and stale data.
Review vendor controls, logging, retention, and access before production use.

Conclusion

BSP uses Codex to accelerate the analysis loop while retaining human responsibility at the decision point. That is a practical model for finance, legal, and IT teams: automate preparation, make evidence visible, and treat professional review and data governance as mandatory controls.

Controls before scale

A useful pilot should make the evidence path visible. For each answer, reviewers should be able to open the underlying filing, agreement, or approved memo and decide whether the model selected the right text. Teams should also record when the system has insufficient evidence rather than forcing a confident answer. That design makes errors easier to find and turns review feedback into a better process.

Production rollout should be staged. Start with a small group, keep original workpapers, compare AI-assisted and conventional reviews, and investigate material disagreements. Establish an owner for prompts, source updates, access changes, and incident handling. The result is not merely faster drafting; it is a controlled workflow in which speed does not erase accountability.

Source

OpenAI: Benefit Street Partners answers questions faster with Codex