OpenAI now answers its own support phone line with an AI agent, and is selling the system that does it

Most companies that tried to put an AI agent in front of customers discovered the same thing: the model was never the hard part. Keeping it reliable as products, policies and customer behaviour change is the hard part. OpenAI has now packaged its own answer to that as a product called Presence, and the most persuasive detail is that the company points it at itself.

Presence powers OpenAI’s English-language phone support line. According to OpenAI, it verifies callers, pulls account context, takes approved actions and handles open-ended requests, and now resolves 75 percent of inbound issues without a human. The company says it met or exceeded the internal benchmarks used to grade human frontline support within weeks, and that an improvement loop driven by Codex cut human handoffs by 15 percentage points in ten days.

The structure is deliberately narrow. Each deployment starts with one specific job, such as resolving billing issues or handling IT service requests, and the agent gets only the knowledge and system access that job requires. The company sets what the agent may do, when it needs approval and when a person takes over. Before launch, teams run simulations against edge cases and higher-risk scenarios, with graders checking whether the agent reached the right outcome and followed policy. After launch, real sessions and escalations feed back in, and Codex proposes changes that humans test and approve.

Named design partners are still exploring rather than deployed. BBVA is looking at voice support for everyday banking in Mexico, SoftBank is testing Japanese-language conversations, and insurer IAG wants support during severe weather events. Presence is available only to eligible enterprise customers through a limited programme, led by OpenAI’s own deployed engineers. There is no self-serve version.

What is behind this. This is an admission dressed as a launch. Three years of enterprise AI have shown that handing a company API access to a good model does not get you a working agent, and OpenAI, Anthropic and Microsoft have all responded by building service organisations that send engineers on site. Presence is that lesson turned into a product: the value is not the model, it is the policies, the simulations, the escalation rules and the review loop wrapped around it. Notice how much of the design is about restriction rather than capability. Narrow job, minimum access, approval gates, a human backstop. That is roughly the opposite of the fully autonomous agent story, and it is what actually shipping looks like right now. One thing OpenAI does not address: how deployments handle specific compliance requirements such as the EU AI Act. The announcement mentions trust mechanisms but no legal detail.

What this means for you: if you run a small business, you cannot buy this, and that is worth knowing before someone sells you an imitation. The design principles are free though, and they transfer to whatever agent tooling you can actually get: give it one job, give it only the access that job needs, decide in advance what needs your approval, and make sure there is an obvious route to a human. As a customer, the 75 percent figure is the one to keep in mind next time you are on hold. The agent picking up will increasingly be able to do things, not just answer questions, which is better when it works and more consequential when it does not.

Sources

• OpenAI: Introducing OpenAI Presence
• The Decoder: OpenAI Presence wants to make AI agents production-ready for businesses