OpenAI says its forward-deployed engineering team works with customers to productionize AI workflows, reduce custom integrations, and surface model gaps for research.
OpenAI’s forward-deployed engineering team works directly with customers to move AI systems from demonstrations into operational use, according to a company account published October 22, 2026. Colin Jarvis, who leads the team, said it was founded in early 2025 to address the distance between what models appear able to do and what organizations can reliably deploy.
One example involved a large semiconductor customer. In mid-2025, about 15 OpenAI engineers worked across more than 10 use cases, including generating verification tests, debugging chip designs and assisting with architecture work. OpenAI reports that the deployments reached production and early adoption within three to six months. The company says AI-directed test generation saved about 48% of time, while kernel optimization improved median performance by 22%; some debugging work that took weeks was reduced to hours.
The team then used Codex to standardize parts of those systems. OpenAI says custom applications were converted into plugins and skills using shared Codex components, removing roughly 90% of custom code from each project. In one August 2026 example, a deployment that took two months to establish was extended to a second team in one day.
The account also describes work with a telecommunications customer seeking to automate voice-based customer service. OpenAI says the model initially scored around 60% on the relevant evaluations. Over six months of joint work between the customer, the engineering team and OpenAI Research, performance rose to 80% and then above 90%, which the company says met the production bar.
These cases illustrate the team’s stated approach: build around a specific customer outcome, identify where existing models or products fall short, and turn successful custom work into reusable capabilities. OpenAI also cites other deployments, including a tax-processing pilot that handled 7,000 returns and increased throughput by about 50%. The company presents the role as temporary by design: as products improve and workflows become reusable, less bespoke engineering should be needed for similar customer problems.