OpenAI’s Colin Jarvis: Enterprise AI Stalled on Deployment, Not Models
OpenAI's Colin Jarvis told HumanX that most enterprise AI problems revolve around deployment and trust, not model capability. His engineers aim to make themselves unnecessary.
September 23, 2026 - 10:10 pm
Image: OpenAI’s Colin Jarvis at HumanX in Amsterdam, 23 September 2026. Credit: ALX MEDIA / HumanX
Jarvis, head of forward deployed engineering at OpenAI (learn more), noted that about 80% of companies struggling with enterprise AI face deployment issues, rather than awaiting better models. In the remaining 20%, customers cite limitations in model functionality. These gaps are becoming more specialized and less prevalent in everyday office tasks.
What Forward Deployed Engineers (FDEs) Do
Initially, FDEs were primarily software engineers due to the need to build on raw APIs from scratch. However, tools like Codex have made 50% of projects now involve custom work, down from 90%. This shift has led to a greater emphasis on domain expertise among the team, including former investment bankers, scientists, and chip verification engineers.
Every engagement starts with a two-day visit where the FDE team asks business leaders to identify key levers in their operations, regardless of AI. They then focus their efforts on optimizing these areas.
Case Study: A semiconductor company struggled with bugs from overnight jobs, consuming 80% of engineering time. OpenAI developed a system that pinpoints root causes, proposes patches, and applies them after human review. This solution spread throughout the company, resulting in an estimated annual savings of $40m to $50m.
Other projects include:
- An agent that generates car part designs from natural language instructions.
- An AI agent that assists in drafting clinical trial documents while a human remains in control.
Common Pitfalls in AI Pilots
Jarvis identified two primary mistakes:
- Selecting use cases based on perceived AI fit rather than business importance.
- Achieving success in one department but failing to scale it across the organization.
Successful companies, according to Jarvis, measure success by production usage rather than proofs of concept. One semiconductor customer has approximately 35 live use cases after about 18 months, with a dedicated central team and small groups in each business unit.
Incentives and Pauses
FDEs have no financial incentives tied to usage; their performance is measured by whether projects reach production and impact key metrics. In one instance, Jarvis advised Klarna to use an open-source model instead of OpenAI’s embeddings when it was too slow for their search service.
"From OpenAI's side, we should always be temporary," Jarvis stated.
The Pace of AI Development
Regarding Sam Altman's presence at the UN and growing pressure to slow down AI, Jarvis noted that OpenAI has paused its main reinforcement learning run "in September this year." He referenced OpenAI’s own post on the pause from August 18th, which details a two-week halt in reinforcement learning training for their latest models.