OpenAI Presence: enterprise agents need a control system, not just a model
OpenAI Presence deploys voice and chat agents into company workflows. Its main claim is narrow permissions, testing, escalation, and controlled improvement.
- [01]OpenAI — Introducing OpenAI Presence2026-07-27
On July 22, OpenAI announced Presence, an enterprise product for deploying voice and chat agents into customer support, sales, insurance claims, and employee IT service. It is not being offered as a conventional self-serve agent builder.
Presence is in a limited general-availability program for eligible enterprise customers, deployed by OpenAI forward-deployed engineers and selected systems integrators. Today it is closer to a packaged deployment and operations service than an API feature.
Each agent starts with one job
OpenAI's method is not to expose every company system to a general agent. A deployment begins with a specific job: resolve a billing issue, support a claim, or process an employee request. The agent receives only the information and system access needed for that job.
The company defines what the agent can do, when approval is required, and when a person must take over. The evaluation is not whether it sounded good, but whether it reached the right outcome, followed policy, used tools correctly, and escalated when necessary.
Codex sits inside a controlled improvement loop
Presence examines production sessions, escalations, and quality signals to find gaps. OpenAI says Codex proposes changes that teams test against the production version before approving a controlled rollout. The agent does not silently change its own behavior.
Presence runs OpenAI's English phone support channel. OpenAI says it resolves 75% of inbound issues without human assistance and reduced handoffs by 15 percentage points in 10 days. Those are company metrics from a specific channel, not independently audited general results.
Solo builders can copy the method
Small teams may not buy this product today, but the method scales down: define an agent as a worker with one job, give each tool a separate permission, measure failed outcomes and handoffs, test changes against recorded scenarios, and make explicit human approval the final step for high-risk actions.
My takeaway is that Presence is not mainly competing with another model. It competes with the fragmented agent-operations stack companies assemble themselves. OpenAI is packaging policy, evaluation, permission, escalation, and continuous improvement around model access.