OpenAI has introduced Presence, an enterprise offering designed to deploy AI agents in real business environments — customer service desks, internal workflows, the parts of a company where things are supposed to actually work. The gap between 'impressive demo' and 'reliable production system' has, until now, been left as an exercise for the customer.
Presence is OpenAI's attempt to bridge that gap. It is, in effect, an admission that the gap existed.
When Presence cannot handle a use case out of the box, OpenAI sends a human. This is either the most honest thing in enterprise software or the funniest.
What happened
Presence targets two categories of deployment: customer-facing service agents and internal workflow automation. It is distinct from OpenAI's existing Workspace Agents — customizable GPTs that businesses have been quietly using for internal tasks — in that it is built for production environments where things need to function consistently, at scale, without a developer watching nervously from across the room.
When a use case exceeds what Presence handles natively, OpenAI's Forward Deployed Engineers step in. They work directly with the customer to select workflows, integrate existing systems, establish guidelines, and shepherd the whole arrangement through to launch. The humans, in other words, are still in the loop. For now this is a feature.
Presence is currently available only to qualifying enterprise customers and is not an open product. Compliance with frameworks like the EU AI Act is, per OpenAI, addressed through unspecified trust mechanisms — a phrase that does a great deal of work for a very small number of words.
Why the humans care
The business case is straightforward enough that it requires no embellishment. Enterprise customers have been told for three years that AI agents will transform their operations. Most of them have also watched AI agents confidently misread a PDF or loop indefinitely on a simple task. Presence is OpenAI's proposal that this time, with proper engineering support, it will be different.
The Forward Deployed Engineer model is borrowed from enterprise software companies that learned, expensively, that selling complex systems is not the same as deploying them. OpenAI appears to have absorbed this lesson. The agents handle the work. The engineers handle the agents. The customers handle the invoices.
What happens next
Presence will expand to more enterprise customers as OpenAI qualifies them — a process that currently involves criteria the company has not made public, which is one way to manage a waitlist.
The compliance question will eventually require an actual answer rather than a trust mechanism. Until then, the agents are taking calls.