OpenAI's Forward Deployed Engineering (FDE) team exists to close the "deployment gap": the difference between what models look capable of in research and what businesses can depend on in production. FDEs work alongside customers to identify their hardest, real-world problems, build solutions that can be operated reliably, and feed concrete needs back to Research so models can improve.
Colin Jarvis founded the team in early 2025. The team's guiding principle is: “Start with the outcome, do what it takes to succeed, then follow with scale.” In practice that means shipping a working solution quickly, learning what holds up in production, and then simplifying and scaling what works.
How they work: embed, learn, measure
FDEs typically embed with a customer’s team, learn the domain and workflows, and co-design evaluations to measure whether an AI delivers real value. Where current models suffice, FDEs work with Product and the customer to make solutions repeatable and production-ready. Where models fall short, FDEs define the capability gap for Research and collaborate to close it. The aim is for every engagement to leave a better starting point for the next customer—if we make ourselves unnecessary for a class of problem, the team has succeeded.
Examples and measurable outcomes
-
John Deere: See & Spray is a precision weed-control system that uses cameras and machine learning to spray weeds rather than whole fields, reducing chemical use by up to 70%. The challenge was scaling personalized, actionable support for farmers—help with configuration, seasonal planning, and adapting practices as conditions change. John Deere deployed OpenAI’s APIs to provide natural-language answers, tailored setup guidance, diagnostics for underperforming equipment, and ongoing, relevant guidance across the customer journey to improve efficiency and sustainability.
-
Large semiconductor customer (mid‑2025): In mid-2025 about 15 FDEs were working with a major semiconductor customer on more than 10 use cases, aiming to make verification AI-native. Verification can account for ~70% of a project’s effort in that industry, so the opportunity was large. FDEs embedded with verification teams, learned chip design, built evals with engineers to test AI-generated tests and failure diagnosis, and created custom harnesses. Over the first 3–6 months all solutions reached production and early adoption. Measured results included:
- ~48% time saved via AI-directed test generation
- 22% median performance improvement through kernel optimization
- Weeks of debugging compressed into hours
-
Life sciences: FDEs with scientific backgrounds worked with researchers on pre-clinical drug-discovery tasks such as target validation and prioritization, gene-expression analysis, and chemical synthesis-route reconstruction. These engagements demonstrated value but initially required too much custom integration to replicate across teams. Rebuilding around Codex reduced roughly 90% of custom code by turning core functionality into centralized primitives, skills and plugins, making it easier to share capabilities. For example, an August 2026 deployment that initially took two months was extended to a second team in one day.
-
Thrive Holdings: OpenAI FDEs and researchers partnered with engineers and accountants to build Tax AI for Crete’s accounting firms. During the pilot Tax AI processed 7,000 tax returns and increased throughput by about 50%. Automating rental-property returns was one of the hardest tasks—reaching 90% precision and recall required six weeks of engineering and practitioner review—but produced reusable patterns that eased expansion to other complex tax schedules.
-
Voice customer-service work: With a large telecommunications customer, voice accounted for around 90% of their customer-service volume. Initial model performance on key evals was about 60%, indicating capability limits for function calling and instruction following in voice scenarios. FDE built the platform and defined requirements with the customer, Research worked on model improvements, and over six months performance rose from ~60% to ~80% and then to 90%+, enabling production deployment. The improved voice model now supports that customer and other managed deployments through Presence.
Why the work is hard
Problems are often open-ended, lacking established ground truth or playbooks. Teams must invent new systems and do the integration work few want to do. Hiring talented FDEs while raising the bar for what counts as "good enough" requires continuous experimentation and practical application. FDE teams are built from both generalists and domain experts: enough domain knowledge to make sound decisions, and enough curiosity to challenge assumptions.
Making yourselves obsolete — two pathways
There are two ways the FDEs can make themselves obsolete. One is product-driven: as a product (or model) improves, much of the custom implementation can be replaced and the value preserved. The other is capability-driven: by working on problems that current models cannot yet solve, FDEs surface concrete gaps for Research; when Research closes those gaps, future models can handle harder tasks without bespoke engineering. Both outcomes extend the return from any single engagement.
Final note
Colin Jarvis says he joined OpenAI to make AI useful in everyday work. The practical approach is to build simply where possible, ship quickly, learn from real users, and then scale what works. Each engagement should improve the next—whether by making a product easier to deploy, nudging model capabilities forward, or reducing custom work for future customers. If you have a hard problem to solve, the FDE team offers partnership; if you want to help solve such problems, they’re hiring.
