Every company I talk to that has tried to put an AI agent in front of real customers has run into the same wall. The demo works. The pilot works. Then it meets an actual customer with an actual weird problem, and the thing either makes something up, gets stuck, or quietly does the wrong action to a real account. OpenAI's answer to that wall is a product called Presence, and it is worth looking at closely because of what it admits.

What Presence actually is

OpenAI describes Presence as an enterprise AI agent platform for deploying voice and chat agents in customer-facing and internal workflows. The Decoder's write-up on August 2, 2026 adds the useful detail: OpenAI's existing Workspace Agents are essentially customizable GPTs built for internal use, while Presence is aimed at production deployments — the agents that talk to your customers, not just your staff.

The part I find most telling is the delivery model. When a use case goes past what Presence handles out of the box, OpenAI sends in Forward Deployed Engineers who work directly with the customer to pick workflows, connect existing systems, set guardrails, and run testing through to launch. Presence is currently limited to qualifying enterprise customers. It is not an open product you can sign up for.

Read that again: the company with the best models on the planet is shipping humans alongside the software.

Why that detail matters more than the launch

There is a story the agent industry has been telling for two years, which is that you drop in an agent, point it at your systems, and it figures the rest out. Presence is OpenAI quietly conceding that the last mile is integration work, permissions work, and testing work — and that it is hard enough to need engineers on site.

I believe that, because it matches what actually happens. The model is rarely the bottleneck. The bottleneck is that your CRM has three fields that mean the same thing, your refund policy lives in a person's head, and nobody has decided what the agent is allowed to do without asking. None of that is fixed by a smarter model. If you are still sorting out the categories here, I wrote a plain version of agents versus chatbots and one on what AI customer support realistically covers.

What OpenAI has not said

Two things are missing, and I am not going to paper over them.

First, compliance. The Decoder points out that how Presence handles specific regulatory requirements like the EU AI Act is unclear. OpenAI mentions trust mechanisms; it has not published legal detail. For anyone deploying a voice agent to EU customers, that is not a footnote, that is the whole procurement conversation.

Second, pricing and availability. "Qualifying enterprise customers" means invite-only enterprise sales. There is no public price, no self-serve tier, and no way for a small business to evaluate it. So if you run a 12-person company and you were hoping this was your answer, it is not — not this year.

Anything beyond that, including claims about accuracy or resolution rates, is vendor material until independent numbers exist. I have not seen independent numbers.

The risk nobody puts in the launch post

Production agents hold credentials. To resolve a customer issue, an agent needs to read an account, maybe issue a refund, maybe change a booking. That means real access to real systems, running unattended, driven by text that customers type.

We already have public examples of agent deployments going sideways. OpenAI itself has been through the rogue agent incident this year, and the pattern in those stories is never "the model was dumb." It is "the model had more access than anyone had actually inventoried."

So if a Presence-style deployment is on your roadmap, the boring work is the important work. Write down exactly which systems the agent can touch and which actions require a human. Give it scoped, revocable credentials rather than the shared key that already exists in your environment. Then go audit what your agents can already reach before you add another one — most teams are surprised by that list.

Disclosure while I am on this subject: I build Agent Master Key, a local credential broker for exactly this scoped-access problem, so treat me as an interested party on the credentials point. The advice stands on its own — you do not need my tool to write down what your agent is allowed to do.

Who this is actually for

Presence is for companies with a call centre, a support queue measured in thousands of tickets, an existing systems integrator budget, and a compliance team. For those companies it is a serious offering, and the Forward Deployed Engineers model is a real advantage over buying a platform and figuring it out alone.

For everyone smaller, the useful takeaway is the shape of the thing, not the thing itself. OpenAI is telling you what production agents require: defined workflows, connected systems, explicit guidelines, and a real testing phase before launch. You can do all four of those with cheaper tools. Most failed agent projects I have seen skipped all four with an expensive one.

My verdict

Presence is a credible enterprise product and an honest signal. The honesty is in the staffing: OpenAI is admitting that shipping agents into production is a services problem, not just a model problem.

If you are enterprise, take the meeting, and make the EU AI Act question the first thing you ask, because that answer is not public yet. If you are not enterprise, ignore the product and copy the process. Scope the workflows, connect the systems properly, define what the agent may not do, test it against your ugliest real tickets. That is the part that decides whether an agent works, and it was always the part nobody wanted to do.

Sources