How Ramp engineers accelerate code review with Codex

Ramp, the corporate card and spend management startup, has apparently found a use for AI coding tools that goes beyond autocomplete gimmicks. According to OpenAI's writeup, Ramp engineers are running Codex, paired with the GPT-5.5 model, as a first pass reviewer on pull requests before a human ever looks at the code. The pitch is straightforward: instead of a developer submitting code and waiting hours for a colleague to get around to reviewing it, Codex reads the diff almost immediately and flags issues, suggests fixes, and gives the kind of substantive feedback that used to require a senior engineer's spare afternoon.

What makes this notable isn't that a company is using an AI model to look at code. Plenty of teams do that already, often as a novelty or a spellchecker for syntax. Ramp's approach, as described, treats Codex as an actual participant in the review loop, catching logic errors and structural problems, not just style nits. That shifts the bottleneck. Code review has long been one of the slowest parts of shipping software, mostly because it depends on a human being available, awake, and willing to context-switch into someone else's problem. Compress that to minutes and the whole release cadence changes.

There's an obvious tension here worth sitting with. Faster review sounds great until you ask what gets lost when the first read of your code comes from a model instead of a person who understands the product, the customer, and the six other systems your change might quietly break. OpenAI's post, naturally, frames this as pure upside — Ramp ships faster, engineers get their time back. It doesn't dwell on the failure modes: a model missing a subtle race condition, or engineers rubber-stamping AI-approved code because the AI already said it looked fine.

Still, the direction is clear enough. Ramp is a fast-growing fintech company that presumably cares a great deal about correctness, given that bugs in financial software tend to cost real money. If they're comfortable letting Codex sit in that review pipeline, it suggests the tooling has crossed some internal threshold of trust that a lot of engineering orgs haven't reached yet. Expect more companies to quietly follow, then eventually talk about it in a case study exactly like this one.