OpenAI's head of forward deployed engineering, Colin Jarvis, revealed that about 80% of enterprise AI issues stem from difficulties in deployment rather than limitations of AI models themselves. Jarvis highlighted that many organizations lack the know-how to roll out, govern, and establish trust in AI solutions.
- 80% of enterprise AI problems occur during deployment, not model performance.
- OpenAI embeds domain experts to customize AI solutions tailored to business needs.
- Successful AI adoption requires measuring production use, not just proofs of concept.
What happened
Colin Jarvis, leading OpenAI’s forward deployed engineering team, shared insights on the primary challenges companies face when adopting enterprise AI. According to Jarvis, most problems arise in the deployment phase, with organizations unsure how to implement, govern, and trust AI effectively. This insight comes from OpenAI's hands-on experience embedding engineers within client companies to help tailor and launch AI systems.
Jarvis noted that only about 20% of client issues stem from the AI models themselves, typically involving narrow, highly specialized tasks like semiconductor design. The majority of projects start with a direct engagement to identify key business levers, leading to practical, custom-built tools that integrate AI into existing workflows. Examples include automated software bug fixing and iterative design tools for manufacturing.
Why it matters
This perspective shifts the focus from developing ever-more-advanced AI models to solving the operational and governance complexities of AI adoption in business environments. By emphasizing deployment challenges, OpenAI highlights that AI’s real value is unlocked only when organizations build the infrastructure and trust needed for scale.
Furthermore, the approach of embedding forward deployed engineers with domain expertise ensures solutions address concrete business problems rather than theoretical AI capabilities, helping customers realize significant cost savings and productivity gains. Jarvis underscored that measuring success by actual production use and tangible impact is vital to avoid static pilot projects that never expand.
What to watch next
Enterprises and AI vendors will likely continue investing heavily in deployment support and cross-functional AI integration teams. This trend is mirrored by major players like AWS and Microsoft, who are also dedicating billions to embed engineers and scale AI solutions within client environments.
On the governance and safety front, OpenAI’s deployment teams serve as real-world testers for internal safety frameworks. As regulation debates advance globally, monitoring how companies maintain control and trust in deployed AI systems will be crucial. OpenAI’s recent pause in reinforcement learning training signals a cautious approach as technological and ethical boundaries continue to evolve.