What FDE means at OpenAI

Colin Jarvis, who heads Forward Deployed Engineering at OpenAI, laid out how the team works: embed with a customer, get a hard problem into production, fold what's reusable into products like Codex, and send the problems the models can't yet solve to Research. He says he founded the team in early 2025 to close what OpenAI calls the deployment gap, the space between what models seem able to do and what people can reliably use.

This matters because it's a candid description of how AI actually reaches work. Not a model release and a press demo, but teams of engineers learning a customer's business, writing evaluations with its staff, and grinding until something survives real conditions.

We think this is one of the more useful things OpenAI has published about enterprise AI, precisely because it admits how much work sits between a benchmark and a job done. The motto Jarvis puts on a slide, start with the outcome and scale later, is the right instinct in an industry that, as he concedes, is prone to hype.

The examples carry real numbers. With a large semiconductor customer, about 15 engineers worked on more than ten verification use cases in 2025; OpenAI reports roughly 48% time saved on directed test generation and debugging that shrank from weeks to hours. When Codex matured, the team says it removed about 90% of the custom code per project, and extending one deployment to a second team took a day after the first took two months. With Thrive Holdings, a tax tool processed 7,000 returns in its pilot and lifted throughput about 50%. These are OpenAI's own figures about its own customers, unnamed in the semiconductor and telecom cases, so we treat them as claims, not audited results.

The telecom story is the one I keep coming back to. The model scored around 60% on the customer's key evaluations for voice support. OpenAI told the customer, brought in Research, and over six months the scores rose past 90%. The resulting voice model now serves other customers too. That's a genuinely productive loop. It is also a structural advantage that's easy to miss: a frontier lab placing engineers inside major companies, learning their hardest workflows, and turning that into model improvements that competitors without the same access can't easily match.

There are costs to name. One customer now handles more than half its calls with AI, and the telecom work aimed at voice, which made up around 90% of that company's service volume. Efficiency for the company can mean fewer jobs for the people who used to answer those calls, and customers who never chose to talk to a machine. Accountants and chip engineers were partners here; call-center staff are mostly absent from the story.

Our read: this is how useful AI gets built, and OpenAI deserves credit for describing the unglamorous part. But the model it describes, where the vendor's engineers sit inside your business and feed lessons back to the lab, should make buyers think carefully about dependence, data and who captures the value.

The post also cites John Deere using OpenAI's APIs to give farmers personalized guidance on its See & Spray weed-control system, which the post says can cut chemical use by up to 70%. OpenAI Presence, its enterprise voice and chat agent product, is described as growing out of these field deployments, tracing back through Swarm and the Agents API.