OpenAI describes Forward Deployed Engineering (FDE) as a team that works alongside customers to convert model capabilities into reliable, deployable systems. The group was founded in early 2025 to close what OpenAI calls the “deployment gap”—the difference between what models appear able to do and what organizations can use consistently in real settings.
Mission and working method
Colin Jarvis, Head of Forward Deployed Engineering at OpenAI, says FDEs start engagements by defining the outcome, solve the problem to a production standard, then focus on scaling. The team embeds with customers, learns domain specifics and builds evals and harnesses that prove whether a model can handle real variables and constraints.
Customer cases and measured results
OpenAI published several customer examples to illustrate the approach. With John Deere, FDEs helped apply APIs to make guidance for See & Spray—its precision weed-control system—more personalized, with the goal of helping farmers configure equipment and plan seasonal work. John Deere’s See & Spray itself can reduce chemical use by up to 70%, OpenAI notes.
In the semiconductor example, mid-2025 teams of roughly 15 FDEs worked on verification use cases. OpenAI reports metrics from those deployments: about 48% time saved through AI-based directed test generation and a 22% median performance improvement from kernel optimisation. Early deployments reached production in three to six months.
For Thrive Holdings’ Tax AI pilot, OpenAI reported that the system processed 7,000 tax returns and increased throughput by roughly 50%. Rental-property returns required six weeks of engineering and practitioner review to reach 90% precision and recall; those efforts produced reusable patterns for other tax schedules.
Scaling, productization and research feedback
FDE work sometimes leads to product changes. OpenAI says moving multiple custom applications onto a common Codex-based harness removed roughly 90% of custom code per project and enabled reuse through plugins and skills. Presence, OpenAI’s enterprise voice and chat product, grew from deployments and now supports managed deployments where one customer handles more than half of its calls with AI.
Where models initially fell short, FDEs worked with customers and Research to define gaps. In one voice deployment, key eval scores rose from about 60% to more than 90% over six months as Research improved function calling and instruction-following behavior; the upgraded model now supports that and other customer-service deployments.
OpenAI presents these outcomes as iterative: build usable solutions, learn what holds up in practice, and remove custom work where a product or model can take over. As one concrete marker, an initial deployment that took two months was extended to a second team in one day by August 2026, illustrating the reuse achieved through FDE-led standardization.
Original source: OpenAI News