AI Products
OpenAI Introduces Presence for Enterprise AI Agents
OpenAI has introduced Presence, a limited-availability enterprise product for deploying voice and chat agents with policies, evaluations, approved actions, and human escalation.
What OpenAI announced
OpenAI announced Presence on July 22, 2026 as a deployed product for eligible enterprise customers that want to operate AI agents in voice and chat workflows. OpenAI describes it as a product for agents that can answer questions, resolve issues, use connected company systems, take approved actions, and hand a conversation to a person when the situation requires it. It is not a self-serve product at launch: deployments are led by OpenAI Forward Deployed Engineers and selected systems integrators.
The announcement is notable because it focuses less on a single model and more on the operating system around an agent. Presence combines model reasoning with company policies, standard operating procedures, guardrails, approved actions, simulations, evaluations, and escalation rules. The goal is to make a specific workflow reliable enough for production while allowing the organization to retain control over what an agent can access and do.
A narrow job is the starting point
OpenAI's description emphasizes starting with a defined job, such as a billing request, an insurance claim, or an employee IT-service issue. For each job, the agent should receive only the knowledge and system access that are necessary. The company decides its permission boundaries: which actions may be performed automatically, which need an approval step, and which cases must be passed to a human.
That design has a practical implication for buyers. An agent that can converse naturally is not automatically ready to modify an account, apply a policy, or make a high-impact decision. Teams should separate the language interface from the action layer, document every permitted action, and test the cases in which the system must decline, ask for more information, or escalate. A sensible pilot has measurable limits: a single queue, a narrow set of intents, representative data, and a clear owner for exceptions.
Evaluation continues after launch
Presence also presents agent deployment as a continuing quality process. Before release, teams can use simulations and graders to test routine requests, edge cases, and higher-risk scenarios. The stated checks include whether the agent reached the right outcome, followed policy, used tools correctly, and escalated when appropriate. Guardrails can intervene when an interaction moves outside the defined boundaries.
After release, production sessions, escalations, and quality signals can expose where the workflow needs adjustment. OpenAI says Codex can investigate those signals through a Presence plugin and propose changes for teams to test and approve against the version in production. That feedback loop may be valuable, but it also makes change management essential: keep versioned policies and prompts, evaluate proposed updates on held-out cases, and require accountable human approval before expanding access or autonomy.
What the published examples do and do not establish
OpenAI says Presence powers its English-language phone-support channel and reports that the system resolves 75% of inbound issues without human assistance; it also says a Codex-powered improvement loop reduced human handoffs by 15 percentage points in 10 days. BBVA, SoftBank, and IAG are described as exploring or testing related customer-agent uses. These examples illustrate the intended direction, but they are not a universal performance guarantee.
Organizations should validate their own language coverage, data residency, identity verification, audit logging, reliability targets, and recovery process. They should also measure the cost of errors, not merely containment or handle time. The useful question is whether a bounded workflow can be made safer, clearer, and more useful for customers and employees than the current process—not whether an agent can complete a convincing demonstration.
Availability and practical next steps
According to OpenAI, Presence is available through a limited general-availability program for eligible enterprise customers, rather than as a product anyone can activate directly. Interested organizations should confirm eligibility, commercial terms, regional support, integration scope, and the responsibility split between their own team, OpenAI, and any systems integrator.
For a responsible evaluation, select one high-volume but reversible workflow, map its data and permissions, define escalation rules before connecting tools, and establish a review cadence for real conversations. Keep a way to pause actions quickly, preserve audit evidence, and give users a straightforward route to a human. These controls turn an agent demonstration into an operational service.