OpenAI turns enterprise agents into a product with Presence

OpenAI introduced Presence on July 22 as a product for enterprises that want agents in real workflows, not only demos. The pitch is straightforward: the hard part is no longer proving an agent can answer once, but making it reliable enough to act inside customer support, sales and internal operations.

What Presence includes #

Presence combines model reasoning with policies, guardrails, approved actions, simulations, evaluations and escalation rules. A deployment starts with a narrow job, such as a billing issue, insurance claim or employee IT request. The agent gets only the knowledge and system access needed for that job, while the company decides what actions require approval and when a person takes over.

OpenAI says Presence already powers its English-language phone support channel and resolves 75% of inbound issues without human assistance. The company also named BBVA, SoftBank and IAG as enterprise design partners.

Why it matters #

This is OpenAI moving beyond the model race into managed agent operations. The interesting part is the Codex-powered improvement loop: after launch, production sessions and escalations can surface gaps, and Codex proposes changes for teams to test and approve. That is the shape enterprise agents probably need: less magic, more controlled iteration.