OpenAI is abandoning its self-serve roots for a new enterprise product called Presence, a managed service delivered by the company’s own Forward Deployed Engineers. Announced July 22, it’s a clear admission that connecting a powerful model to a messy enterprise is a labor problem, not just a software one. The company is betting that sending its own people into the trenches—scoping business outcomes, navigating legal review, and managing staged rollouts—is the only way to stop the failure rate Gartner predicts will claim over 40% of agentic AI projects by 2027.
The offering treats each deployment as a project, starting with a single, tightly-scoped job like a billing dispute. The customer writes the rules for escalation and sign-off, and a feedback loop called Codex analyzes production sessions to suggest guardrail tweaks. It’s a pragmatic response to the reality that governance and operational discipline, not model capability, are what kill enterprise AI initiatives. OpenAI’s own phone support line, 1-888-GPT-0090, is the proof point, and the company claims the agent there now resolves 75% of inbound issues without a human.
But the model’s constraints are as interesting as its promise. Access isn’t gated by a credit card but by “delivery capacity,” which is a consulting bottleneck, not a compute one. Borrowing the Forward Deployed Engineer title from Palantir signals a move into high-touch, people-heavy implementation work. That creates an awkward overlap with the systems integrators OpenAI also needs to scale, an arrangement that works at low volume and gets complicated fast. It also blurs the lines of accountability when the model vendor is also the team plugging it into your core systems.
Early customers are kicking the tires rather than running at scale. BBVA is exploring voice support in Mexico, SoftBank is testing Japanese-language conversations, and IAG is looking at high-demand event support. Daniel Ordaz from BBVA describes the bank as a “design partner” helping shape the product, which is honest language for a product that’s still being finished in the field. The managed model is a sensible fix for a real problem—it’s also a bet that OpenAI can master a services business before the integrators it’s competing with figure out the technology.
