OpenAI launched Presence on July 22, a managed product that deploys AI agents inside enterprise companies for customer service and internal operations tasks. It is not available to buy online. Every deployment is led by OpenAI's own engineers and a small set of selected global partners.
The scope is deliberately narrow at the start. Each engagement begins with one specific job: resolving a billing dispute, handling an insurance claim, processing an IT service request. The agent gets access only to the systems and information that job requires. The customer writes the rules governing what it can do independently and when a human must step in.
OpenAI's own support line is the clearest proof point it offers. The company says the agent now resolves 75% of inbound calls without human help, and a continuous improvement loop cut the rate of calls handed to humans by 15 percentage points in ten days. Those numbers are OpenAI's own, measured against OpenAI's own criteria, so they are not independent. But the transparency itself is useful, because most vendors in this space do not publish anything comparable.
The three named early customers are BBVA, which is exploring voice support for everyday banking in Mexico; SoftBank, testing Japanese-language conversations; and IAG, the Australian insurer, which is looking at customer support during severe weather events. All three are described as exploring or testing, not running at scale. That is an honest framing for a product that is still in limited availability.
The strategic move here is larger than a single product launch. In May 2026, OpenAI had already set up a separate entity called the OpenAI Deployment Company, backed by more than $4 billion from 19 investors including McKinsey, Bain & Company, and Capgemini, specifically to embed engineers inside enterprise customers. Presence is the product those engineers deliver. The acquisition of a 150-person applied AI engineering firm called Tomoro gave the Deployment Company its initial field team.
This puts OpenAI in competition with the very consultancies and system integrators it is also partnering with to scale. That tension is manageable while volumes are small. It becomes more complicated as the business grows and OpenAI's engineers are competing for the same contracts as the partners it relies on for distribution.
The model itself is borrowed from Palantir, which spent more than a decade proving that complex software deployments in large organisations require engineers embedded on-site, not just a licence and a manual. OpenAI is copying that playbook, and there is good reason to think it is the right one for this moment. Gartner, after polling more than 3,400 organisations, predicts that over 40% of AI agent projects will be cancelled by end of 2027, with the failures driven by poor governance, unclear business value, and integration problems, not by the technology.
Presence is structured to address each of those failure modes directly: one task at a time, policy written before launch, simulation tests before any customer sees the agent, escalation paths that hand humans structured context rather than a dead transcript, and controlled rollouts with rollback options.
Two things are worth watching carefully if you are evaluating this for your own organisation. First, pricing is not published. Cost is set per deployment, which is standard for enterprise services but means you cannot compare it against what you currently spend on a contact centre or a service desk without going through a sales conversation. Second, the specific AI model powering Presence is not named and can change as the deployment evolves. That flexibility makes engineering sense, but any contract should specify what benchmarks OpenAI is held to when the model configuration changes under you.
The accountability question is the sharpest one. When the company selling you the agent is also the company deploying it and writing the policy rules, the lines between model failure and implementation failure need to be in the contract, not assumed.