OpenAI Presence Brings Governed Agents Into Production

OpenAI is positioning enterprise AI agents as an operations problem, not just a model problem. Its new OpenAI Presence announcement describes a managed platform for deploying voice and chat agents that can answer questions, use company systems, take approved actions, and escalate sensitive cases to people.

What Presence adds around the model

Presence packages the controls that production deployments usually have to assemble themselves: policies, standard operating procedures, guardrails, simulations, evaluation tools, permissions, and escalation rules. Each deployment is scoped to a specific job, such as billing support, insurance claims, or internal IT requests. The company decides what the agent may access, which actions require approval, and when a human must take over.

That framing matters. A capable model can produce a convincing answer, but an enterprise workflow also needs a predictable boundary around data and actions. Presence is designed to give the agent only the knowledge and system access needed for its assigned task, reducing the blast radius when a request falls outside policy.

A controlled improvement loop

OpenAI says production sessions, escalations, and quality signals reveal where an agent needs work after launch. Codex can propose updates, while teams test each change against the production version before approving a controlled rollout. This is closer to software release management than to one-time prompt tuning.

OpenAI also says Presence powers its English-language phone support channel and resolves 75% of inbound issues without human assistance. Those performance figures are company-reported, so builders should treat them as deployment claims rather than independent benchmarks.

Why builders should pay attention

Presence is available to eligible enterprise customers through a limited general-availability program, with deployments led by OpenAI engineers and selected systems integrators. It is not a self-serve product yet.

The practical lesson is clear: production agents need an explicit control plane. Teams building with n8n, APIs, or custom orchestration should model permissions, approval gates, simulations, evaluation datasets, and rollback paths as first-class workflow components. The model is only one layer. Reliability comes from everything wrapped around it.

This also connects to the wider move toward portable agent infrastructure, including Microsoft's Agent Host Protocol, where the runtime and control surface become as important as the underlying model.