From an agent that demos to an agent that works
OpenAI introduced Presence as a product for deploying voice and chat agents in enterprise workflows. The promise goes beyond answering: an agent can consult systems, execute approved actions, and hand the conversation to a person when an exception appears.
Each implementation starts with a specific job. Instead of granting general company-wide access, Presence limits knowledge, permissions, and actions to the role the agent must perform.
The real product is the control system
Presence combines procedures, policies, guardrails, simulations, evaluations, and escalation rules. The model reasons, but the organization defines the perimeter in which it may act.
OpenAI says its own English-language phone channel already uses this architecture and reports operational results. Those figures are provider claims, not a transferable guarantee; the design pattern is still relevant.
Improving without letting go of the wheel
Real sessions and escalated cases reveal where the agent fails. Codex can investigate those signals and propose changes, but the team tests every version and approves deployment. Continuous improvement does not mean autonomous updates without review.
Separating observation, proposal, evaluation, and approval makes operations part of the product. A useful agent does not end on launch day; it needs a permanent validation loop.
What any company can borrow
You do not need a global platform to adopt the principle. Start with one valuable action, minimum access, a clear approval rule, and a small set of test cases.
The right question is not only which model to use. It is who controls the process, how an incorrect response is detected, and what happens when AI should not decide.
- Define a concrete job before choosing the technology.
- Separate automatic actions, approved actions, and human handoffs.
- Use real incidents and exceptions to improve the next version.