# What Jarvis said and why it matters
# What FDEs do inside companies
Engagements start with a two-day visit where FDEs ask business leaders to ignore AI and describe the biggest levers in their business. The team then targets whichever lever matters most, not the most AI-friendly use case.
# Concrete examples
- Clinical trials: a system helps draft trial documents while keeping a human in charge of approvals.
# Why pilots fail Jarvis named two frequent mistakes:
- Choosing a use case because it looks like a fit for AI rather than because it moves a meaningful business metric.
- Letting a successful pilot remain a demo inside one department instead of moving it into production and other units.
He contrasts this with a semiconductor customer that had about 35 live use cases after roughly 18 months because it built a central team to scale projects and placed small engineering groups in each business unit.
# How success is measured FDEs are judged on whether a project reaches production and moves real metrics, not on usage revenue or adoption incentives. Jarvis says OpenAI has no financial incentive tied to usage. When OpenAI's embeddings were too slow for a Klarna search service, he advised using an open-source model instead, stressing that OpenAI should act as a temporary solution when appropriate.
# Safety, pace, and testing frameworks Jarvis noted OpenAI has shown it will pause work when safety frameworks have reached their limits. He said the company paused a main reinforcement learning run "in September this year" and that OpenAI published a post on 18 August describing a two-week pause in reinforcement learning training on its latest models. He also said FDEs help test whether safety frameworks that work in labs hold up in messy, real-world company environments.
# Market context Other large providers use similar deployment models: AWS has invested about $1bn in comparable on-site engineering, and Microsoft launched a $2.5bn deployment business in July.
# Bottom line