On July 22, 2026, OpenAI introduced OpenAI Presence, a product for putting voice and chat agents into customer-support and internal workflows. The problem it targets is not whether a model can answer a question. It is whether an agent can perform a defined job reliably while operating within permissions, policies, tests, and human handoff rules.
Presence deployments start with a specific job, such as resolving billing questions, supporting insurance claims, or handling employee IT requests. The company gives the agent only the knowledge and system access needed for that job, then decides which actions are allowed, which require approval, and when a person must take over. Designing around a bounded workflow makes an agent easier to test and govern than connecting a general chatbot to every system at once.
The product combines policies, standard operating procedures, guardrails, approved actions, simulations, and evaluation tools. OpenAI also says production conversations, escalations, and quality signals feed a Codex-powered improvement process. Teams can test proposed changes against the production version and approve a controlled rollout. The stated model is supervised change management, rather than an agent rewriting its own operating rules.
OpenAI points to its English-language phone-support channel as an example. It says Presence can handle open-ended requests, verify callers, use account context, and take approved actions. The company reports that the agent now resolves 75% of inbound issues without human assistance, while its Codex-powered improvement loop reduced human handoffs by 15 percentage points in ten days. These are vendor-reported internal figures, not independent evidence that the same results will transfer across companies, languages, or risk levels.
The delivery model is another important constraint. OpenAI says Presence is available through a limited general availability program for eligible enterprise customers, with deployments led by OpenAI Forward Deployed Engineers and selected systems integrators. It is not yet a self-serve product. Buyers therefore need to account for system integration, permission design, voice quality, monitoring, and the cost of human escalation as part of the deployment.
The announcement shows where enterprise-agent competition is moving. Model reasoning is only one layer. Production reliability also depends on task boundaries, tool permissions, simulation cases, evaluation rules, exception handling, audit records, and a safe process for changing the agent. A useful evaluation question is whether every workflow has a clear definition of done and whether staff can see and take over the agent at the moments that matter.
