OpenAI’s finance team hit a wall late last year. Contract volumes exploded from hundreds per month to over a thousand, but headcount barely budged. “In less than six months, the team went from reviewing hundreds of contracts each month to more than a thousand. And yet we’d only hired one new person,” said AI Engineer Wei An Lee. The manual approach — reading contracts line by line and retyping terms into spreadsheets — simply broke under the load.
So the finance and engineering teams built an internal tool they call the contract data agent. The principle is refreshingly straightforward: automate the soul-crushing repetition, keep human experts firmly in the driver’s seat. The agent ingests everything from clean PDFs to messy phone photos of marked-up pages, then uses retrieval-augmented prompting to parse the contracts into structured data. Crucially, it doesn’t just extract — it reasons. The system flags non-standard terms, cites relevant reference material, and even surfaces ASC 606 classification notes for reviewer confirmation.
The output lands overnight as an annotated dataset ready for validation. What once ate up entire days now arrives while the finance team sleeps. Their role shifts from data entry to professional judgment, which is exactly where it should have been all along. Engineers describe the result as “manual work already done,” not decisions replaced. Each cycle of human feedback sharpens the agent further.
The architecture has already spread beyond its original use case. Procurement, compliance, and month-end close processes now run on the same pattern: rote work automated, humans looped in for the calls that actually require expertise. Wei An summed up the stakes bluntly: “The only way we can scale as OpenAI scales is through this. Without it, you’d have to grow your team linearly in lockstep with contract volume.” That’s the quiet promise here — not flashy AGI, but a practical blueprint for keeping teams lean while handling genuine hypergrowth.
