With Presence, OpenAI industrializes its customer support agents

OpenAI Presence automates customer support via voice and chat. The system relies on Codex to adjust its rules after each session.

The problem with AI agents in the enterprise is no longer proving that they work, but making them reliable enough for high-stakes tasks and keeping them up to date as rules and practices change. This is the niche of OpenAI Presence, a product for deploying support agents, via both voice and chat, capable of processing a request within company systems and handing it off to a human if necessary.

Each deployment starts with a specific task, such as a billing complaint or an internal IT request, with the agent receiving only the knowledge and access relevant to that scope. The company sets the rules: what the agent can do, and when it must request validation or hand over control to a human. Once in service, real sessions reveal gaps, and Codex suggests adjustments that teams test and then approve.

Deployment is not self-service: it is led by OpenAI engineers dispatched to the client's site and by integrators, as part of a limited access program.

OpenAI says it relies on it for its own English-language phone support, which it credits with 75% of requests resolved without human intervention, with the improvement loop via Codex having, according to the company, reduced transfers to a human by fifteen percentage points in ten days. Design partners cited include BBVA for banking in Mexico, SoftBank for interactions in Japanese, and IAG for surges in demand during extreme weather events.