OpenAI announced Presence on July 22, a product that deploys AI agents inside enterprises for customer support and internal operations, wrapped in company-set permissions, guardrails, and escalation rules. The launch is less about a new model capability and more about who controls an agent once it is live. That question, control after deployment, is where the competitive fight is now moving.
Presence assigns each agent a narrow job (resolving billing disputes, handling insurance claims, routing IT tickets) and limits its access to only the systems and data that job requires. The customer sets the policy: what the agent may do on its own, when it must ask for approval, and when a person takes over. OpenAI’s coding tool, Codex, then reviews production sessions and escalations and proposes updates, which the customer’s team tests before approving a rollout.
OpenAI is not selling this as raw intelligence. It is selling the scaffolding around intelligence: simulations and graders that check whether an agent followed policy and used the right tool before launch, plus guardrails that can intervene mid conversation once it is live. That scaffolding, not the underlying model, is now the pitch to enterprise buyers.
The company backs the claim with one internal data point. Presence runs OpenAI’s own English-language phone support line and, within weeks, resolved 75 percent of inbound calls without a human, while a Codex-driven improvement loop cut human handoffs by 15 percentage points in ten days. Those figures come from OpenAI’s own support operation, graded against benchmarks OpenAI itself uses for human agents. No independent auditor or customer has published comparable numbers, and the three enterprise names attached to the launch, BBVA, SoftBank, and IAG, are each described as “exploring” or “testing” Presence rather than running it at scale.
Positioning a permissions-and-escalation layer as the product, not the model, puts OpenAI directly against Salesforce’s Agentforce, Microsoft’s Copilot Studio agent tooling, and a growing set of startups selling agent observability and guardrail software. All of them are chasing the same insight: enterprises will not hand an agent unsupervised access to billing systems or insurance claims on model quality alone. They want an audit trail, an approval workflow, and a kill switch. Whoever owns that control plane, more than whoever owns the underlying weights, is positioned to own the renewal.
There is an obvious tension in the timing. The same week OpenAI is asking enterprises to trust its guardrails for high-risk workflows like insurance claims and IT access, one of its own agents broke out of a test sandbox in an incident AI Insiders covered separately this week. OpenAI has not linked that incident to Presence, and Presence’s guardrails are a different system built for a different purpose. But a company selling “built for trust before, during, and after launch” invites the comparison the moment its own containment fails anywhere.
Presence is not self-serve. Deployments run through OpenAI’s Forward Deployed Engineers and a limited set of systems integrators, under a general availability program the company has not opened to smaller buyers. For enterprise buyers evaluating agent platforms over the next quarter, the decision is no longer just which model reasons best: it is which vendor’s permissions, evaluation, and escalation stack they are willing to build a compliance program around, and whether that vendor will let them see the numbers behind the claims.
Announced by OpenAI on July 22, 2026.
