On 22 July, OpenAI announced Presence, a managed service that deploys AI agents inside large organisations to handle well-defined operational work: billing disputes, insurance claims, IT requests. The most interesting part of the launch is not the model behind it but the delivery format: every deployment is led by OpenAI's own Forward Deployed Engineers or by a small group of approved systems integrators. The company that popularised self-serve AI has concluded that enterprise agents are not a download; they are a project.
The design choices deserve attention from any company evaluating agents. Each Presence agent receives only the knowledge and the system access required for its specific role. The customer defines the operating rules: what the agent may do on its own, which actions need approval, and when a case escalates to a human. A companion tool reviews production sessions and proposes improvements that the customer tests before they reach users. OpenAI describes a six-stage lifecycle: scoping, security and legal review, simulation, testing, staged rollout and post-launch iteration.
The reference numbers come from OpenAI's own support operation: its English-language line now resolves 75% of incoming issues without human assistance, and handoffs to humans fell fifteen percentage points within ten days of deployment. Early design partners include BBVA (voice banking in Mexico), SoftBank and the airline group IAG. Pricing is not public; it is set per customer, and availability is limited for now.
Three practical lessons for businesses weighing AI agents.
First, agents work when they are narrow. The winning pattern is a scoped role with limited access and clear escalation, not a universal assistant holding the keys to every system.
Second, deployment discipline beats model choice. The six-stage checklist above is exactly what a buyer should demand from any provider, at any project size. If a vendor cannot explain its simulation and rollout plan, that silence is the answer.
Third, the market has priced in complexity. If deploying agents were trivial, OpenAI would not attach engineers to every contract. Expect services alongside software, and treat promises of plug-and-play autonomy for critical processes with healthy scepticism.
There is also a structural signal here: frontier labs are moving up the stack, from selling tokens to selling outcomes. That brings more vendor accountability, but also higher switching costs. Contracts should preserve data portability and demand transparent metrics from day one.
The encouraging news for mid-sized companies is that none of this methodology requires a nine-figure budget. Scoped access, human escalation, staged rollouts and measured resolution rates scale down perfectly well. What matters is working with a partner who actually applies them.