OpenAI Presence Turns Agent Deployment into an Enterprise Product
OpenAI introduced Presence, a limited-availability enterprise product for deploying voice and chat agents with policies, guardrails, evaluations, approved actions, and human escalation.
Most companies no longer need to be convinced that AI agents can do useful work. The harder question is operational: how do you connect an agent to real systems, give it enough authority to resolve issues, keep it inside policy, and improve it after launch without turning production into an experiment?
On July 22, 2026, OpenAI introduced OpenAI Presence, an enterprise product for deploying AI agents across customer and internal workflows.1 Presence is available today for voice and chat agents, but only through a limited general availability program for eligible enterprise customers.1 It is not a self-serve product. OpenAI says deployments are led by OpenAI Forward Deployed Engineers and selected global systems integrators.1
That delivery model is the interesting part. Presence is not another SDK that a team installs and wires together alone. OpenAI is packaging agent deployment as a managed production system: workflow selection, knowledge and system connections, permissions, policies, simulations, evaluations, guardrails, approved actions, escalation rules, and ongoing improvement.1
What Presence actually does
OpenAI describes Presence agents as systems that can answer questions, resolve issues, use company systems, take approved actions, and escalate to people when needed.1 Each deployment starts with a specific job, such as billing support, insurance claims, or employee IT service requests.1 The agent receives only the knowledge and system access required for that job, while the company defines what it can do, when approval is required, and when a human should take over.1
The product page lists concrete use cases: customer support, demand generation, claims, human resources, IT help desk, and procurement.2 The common pattern is not "chatbot answers FAQ." It is "agent follows a business process": verify a user, read account context, apply a policy, update a system, route exceptions, and hand off higher-risk work.2
OpenAI also says Presence includes simulations and graders before launch. Those tests check whether an agent reaches the right outcome, follows policy, uses tools correctly, and escalates at the right time.1 Guardrails can intervene if an interaction moves outside company boundaries.1
The Codex loop inside the product
The most notable developer-tool angle is the improvement loop. After launch, production sessions, escalations, and quality signals show where an agent is failing or drifting. OpenAI says Codex, using the Presence plugin, can investigate those signals and suggest updates.1 Teams can then test proposed changes against the production version and approve a controlled rollout.1
That makes Presence less like a static bot deployment and more like a monitored software system. The agent has behavior. The behavior is evaluated. Gaps become change proposals. Those changes are reviewed before rollout.
OpenAI reports that Presence powers its own English-language phone support channel at 1-888-GPT-0090. According to the launch post, it handles open-ended requests, verifies callers, uses account context, and takes approved actions.1 OpenAI says it now resolves 75% of inbound issues without human assistance, and that its Codex-powered improvement loop reduced human handoffs by 15 percentage points in 10 days.1 Those are OpenAI's own production metrics, not independent benchmarks, but they are specific and tied to a live support use case.
Why this matters
Presence is a signal that enterprise AI is moving from model access toward operational deployment. The hard parts are no longer just prompt quality or model selection. They are permissions, policy enforcement, evaluation, escalation, auditability, and controlled change management.
That also changes the role of developer teams. Building production agents now looks closer to building internal platforms than launching isolated assistants. The work includes tool integration, data boundaries, workflow design, test scenarios, rollout control, and observability around failures.
OpenAI's own positioning is careful here. Presence is limited availability, focused on voice and chat agents, and not yet self-serve.1 That says something important: high-value agent deployment still needs hands-on integration work. The product story is not "agents are easy now." It is that agent reliability has become important enough to justify a dedicated enterprise deployment product.
For teams watching the agentic AI market, that is the fresh signal. The race is no longer only about who has the strongest model or the slickest assistant. It is about who can make agents trustworthy enough to run inside messy, high-stakes business workflows.