OpenAI just stopped selling API keys. Now it's selling engineers.

OpenAI launched "Presence" on July 22.

No API dashboard. No self-serve onboarding.

A Forward Deployed Engineer shows up at your office and deploys the agent for you.

This is the Palantir playbook applied to AI inference.

The hard part was never the model. It was integration, permissions, policy enforcement, and change management.

Gartner says 40%+ of agentic AI projects will be cancelled by 2027 — not because the models fail, but because enterprises can't operationalize them.

OpenAI's response: embed engineers in your org, scope the workflow, write the guardrails, test before launch, iterate after.

Their own support line now resolves 75% of inbound issues without humans.

BBVA, SoftBank, and IAG are design partners — none running at scale yet.

The trade-off: delivery capacity is the bottleneck. Engineers cleared into your core systems don't scale like tokens.

When your model vendor is also your implementation partner, accountability for a policy misapplied in production needs to be in the contract.

Audit your AI procurement process. The era of buying an API key and hoping your team figures it out is over. Budget for the humans who make agents production-ready — whether they come from OpenAI, an SI, or your own workforce.