OpenAI's Presence: they stopped selling models and started selling the whole agent workforce

The short version

OpenAI just launched Presence, an enterprise platform for building, deploying and managing real-time voice and chat agents. It already resolves about 75% of OpenAI’s own inbound support, which is the number that made me sit up. And it puts OpenAI squarely against Salesforce, ServiceNow and Zendesk, not just other model labs.

What exactly did OpenAI launch?

OpenAI unveiled Presence, a deployed enterprise product for putting AI agents to work across customer-facing and internal workflows. Presence is not a demo or a research preview. It ships with policies, guardrails, evaluations and human escalation baked in, which are the boring, load-bearing parts that actually decide whether a company trusts an AI agent to talk to real customers.

Here’s the part that tells you how OpenAI sees this product: Presence is available through limited general availability, led by OpenAI’s Forward Deployed Engineers. It is not self-service. You do not sign up with a credit card and start building. OpenAI sends humans to sit with your team and stand it up. That is the sales motion of a company selling an outcome, not an API.

I’ve been reading AI product launches all year, and most of them are a model with a new number attached. Presence is different in kind. According to the original report from VentureBeat, Presence already resolves roughly 75% of OpenAI’s own inbound support volume. When a company dogfoods a tool on its own support desk and reports a number like that, it is telling you the thing works well enough that they trust it with their own customers first.

Why does this matter more than another model release?

For most of the last two years, OpenAI sold raw intelligence. You bought tokens, you built the product, you owned the risk. Presence flips that. OpenAI is now selling the assembled agent workforce, wiring and safety rails included. That is a completely different business, and it changes who OpenAI competes with.

OpenAI is no longer mainly competing with Anthropic and Google over who has the smartest model. With Presence, OpenAI is competing with Salesforce, ServiceNow and Zendesk over who owns your customer service stack. Those companies have spent decades embedding themselves in enterprise workflows. OpenAI is walking in with the argument that the underlying intelligence is now good enough to collapse a lot of that tooling into a handful of agents.

The timing is not an accident either. Anthropic recently launched Ode, its own consulting arm, to help enterprises actually deploy agents. So the two biggest labs both decided in the same stretch that selling the model is not enough. The margin, and the moat, is in deployment. Whoever gets embedded in the enterprise workflow wins the long game, because switching costs on a live customer-service system are brutal.

How does Presence actually work?

Presence lets a company launch and manage real-time voice agents and chatbots that handle both customer-facing and internal tasks. The real-time voice piece is the interesting bit. Text chatbots have existed forever and most of them are frustrating. Real-time voice that can actually hold a conversation, follow a policy, and hand off to a human when it hits a wall is a meaningfully harder problem, and it is where the pain lives in call centers.

The management layer is what separates Presence from a weekend hack. Policies define what an agent is allowed to say and do. Guardrails stop it from going off the rails. Evaluations let a company measure whether the agent is actually resolving problems or just deflecting them. And escalation means the agent knows when to stop and pull in a person. That escalation path is the single feature I’d interrogate hardest if I were buying, because a confident agent that never escalates is how you end up with a viral screenshot of your bot promising something insane.

OpenAI has named early customers including BBVA Mexico and SoftBank, which signals it is aiming at large, regulated, high-volume operations rather than small startups. A bank does not put an AI agent in front of customers casually. If BBVA Mexico is running this, the compliance and guardrail story is at least serious enough to survive a bank’s procurement process.

What does this mean for the rest of us?

Here’s my honest read. If you run or work in customer support, this is the clearest signal yet that the job is changing from answering tickets to supervising agents. The 75% resolution number is a benchmark, not a ceiling. OpenAI is publishing it precisely because it wants every support leader to walk into their next budget meeting and ask why their team isn’t hitting something similar.

If you’re a smaller company, the frustrating part is that Presence is gated behind Forward Deployed Engineers and limited availability. You can’t just grab it. But the pattern always plays out the same way: the white-glove enterprise version comes first, the self-service version follows once the kinks are worked out. I’d bet that within a year there’s a lighter, self-serve path, or a competitor rushes to fill that gap for smaller teams.

And if you build products on the OpenAI API, this is worth watching closely, because OpenAI just became your competitor in the application layer. A lot of startups have quietly built businesses as the deployment layer on top of OpenAI models. Presence is OpenAI saying it wants that layer too. That’s the named tradeoff of building on someone else’s platform: they can always move up the stack into your business.

Is it worth paying attention to?

Yes, but not because voice agents are new. They aren’t. It’s worth attention because a model company is now selling a managed workforce, and that reframes the entire AI-in-enterprise conversation from “which model is smartest” to “who runs your operations.” That’s a bigger shift than any single benchmark score this year.

What I don’t know yet, and what nobody outside those early accounts knows, is how well Presence holds up outside OpenAI’s own tidy support environment. Resolving 75% of your own inbound is impressive. Resolving 75% of a messy, multilingual, regulated bank’s inbound is a different mountain. I’ll believe the general case when BBVA Mexico or SoftBank publishes their own numbers.

FAQ

Is Presence available to sign up for right now? No. Presence is in limited general availability and rollout is led by OpenAI’s Forward Deployed Engineers. It is a hands-on, deployed engagement rather than a self-service product you can buy with a card today.

Who does Presence compete with? Presence competes directly with enterprise workflow and customer-service platforms like Salesforce, ServiceNow and Zendesk, plus Anthropic’s newly launched Ode consulting arm. It is aimed at owning the deployment layer, not just supplying a model.

What is the 75% number about? OpenAI reports that Presence already resolves roughly 75% of OpenAI’s own inbound support volume. It’s a dogfooding benchmark showing OpenAI trusts the product on its own customers before selling it widely.

Does an AI agent replace human support staff? Not entirely. Presence includes escalation to human agents by design, so the realistic shift is toward humans supervising agents and handling the hard, escalated cases rather than answering every routine ticket.

Should smaller companies care if they can’t access it? Yes. The enterprise-first, white-glove version usually precedes a self-service one. The 75% resolution benchmark also sets an expectation that will trickle down to tools smaller teams can actually buy.