OpenAI Presence: What the New Enterprise AI Agent Platform Actually Does (July 2026)

TL;DR. OpenAI launched Presence on July 22, 2026 — a managed enterprise platform for deploying AI agents in production. It combines policy configuration, built-in guardrails, simulation testing, and a Codex-powered loop that reviews live interactions and recommends improvements. OpenAI uses it for its own English-language phone support: the agent resolves 75% of calls without a human. Presence is currently available to enterprise customers through OpenAI's Forward Deployed Engineers — not self-serve.

There's a well-documented gap between AI agent demos and AI agents that work reliably at scale. OpenAI Presence is the company's managed attempt to close that gap for enterprise customers.

Launched on July 22, 2026, Presence is not a new model. It's a deployment and governance layer that sits on top of OpenAI's existing models — GPT-5.6 and the broader model stack — and handles the production-readiness work that most enterprise teams have been struggling to do themselves.

What Presence actually does

The core challenge with production AI agents isn't getting one to work in a demo. It's getting it to work consistently across thousands of real interactions, with real customers, under conditions the demo never covered. A customer who calls at 11pm with a question that doesn't match any training scenario. An employee who submits a request that falls between two service categories. An edge case no one thought to test.

Presence addresses this with a set of interlocking components:

Policies and standard operating procedures. Organizations configure what the agent is allowed to know, what systems it can access, what actions it can take autonomously, and what requires a human to authorize. This is not just a list of rules — it's a structured policy layer that the agent checks against at each decision point.

Guardrails. When an interaction moves outside the company's defined boundaries — a topic the agent isn't authorized to address, a question that touches sensitive data, a request that requires escalation — the guardrails intervene before the agent proceeds. OpenAI says guardrails operate "at machine-speed," meaning they run in real time during the interaction rather than as a post-hoc review.

Simulation and evaluation before launch. Before an agent goes live, organizations can test it against a library of real-world scenarios and edge cases. The evaluation checks whether the agent reaches the correct outcome, follows policy, uses tools properly, and escalates when appropriate. The goal is catching failure modes in simulation rather than discovering them in production.

Codex-powered improvement after launch. This is the piece that differentiates Presence from a one-time deployment. Once an agent is live, Codex — OpenAI's coding agent — reviews production sessions and escalations, identifies patterns in where the agent fell short, and proposes specific changes to improve behavior. Staff review and approve every proposed change before it goes live. The agent adapts as company policies shift or as new edge cases emerge, without requiring the engineering team to rewrite it from scratch.

The numbers OpenAI is citing

OpenAI's primary proof point is its own English-language phone support line, which runs on Presence. According to OpenAI:

• The agent resolves 75% of inbound calls without human intervention.
• The Codex-powered improvement process reduced human handoffs by 15 percentage points within 10 days of initial launch.

These numbers come from OpenAI's own internal deployment — they are not independently verified third-party data. Treat them as directional: OpenAI is confident enough in the performance to use them publicly, and the 15-point handoff reduction in 10 days is a specific enough claim to be meaningful, but it's also the best-case scenario from the vendor's own use. Real-world performance at other organizations will vary.

What the numbers tell you, even conservatively: a well-configured Presence deployment can resolve the majority of routine support requests without a human in the loop, and the improvement loop genuinely works fast enough to matter.

Who can use Presence today

Presence is in limited general availability, which means it's available — but not to everyone, and not as a self-serve product.

Access requires an existing OpenAI enterprise account with a dedicated account team. Deployments are led by OpenAI Forward Deployed Engineers — a team that works directly with enterprise clients to implement complex configurations — along with select systems integration partners. There is no sign-up page, no pricing listed publicly, and no indication of when or whether Presence will become self-serve.

This access model is deliberate. Production AI agents interacting with customers carry real reputational and operational risk for the deploying organization. OpenAI's position is that deployments of this type should be scoped, tested, and monitored by engineers who specialize in them — at least for now.

The practical consequence: Presence is not an option for most small and mid-sized businesses today. If you don't already have an enterprise-tier relationship with OpenAI, you'll be watching this from the sidelines until OpenAI broadens access.

What this means if you're not a qualifying enterprise

Even if Presence isn't available to you, the launch clarifies where the market is heading and what the production reliability bar looks like:

Guardrails are no longer optional. Every AI agent operating in a professional context should have defined policies on what it can and cannot do, and guardrails that enforce those policies in real time. If you're building with the API, this means investing time in defining scope and escalation triggers explicitly — not assuming the model will infer them.

The improvement loop matters more than the initial configuration. The Codex-in-the-loop design reflects a real insight: the best initial configuration is not the final configuration. Agents that work well in production are ones that get updated regularly based on what actually happened. If you're running your own agents, plan for ongoing review of failure cases, not just a launch and leave.

The reliability gap is closing — but slowly. Earlier in 2026, Zuckerberg publicly acknowledged that Meta's agent efforts had stalled. OpenAI Presence is a serious engineering response to the reliability problem, and it's backed by the resources to run it correctly. For most organizations, the practical path to reliable agents still runs through careful scoping, limited autonomy, and human review points — not a fully autonomous agent running unsupervised.

The enterprise path is becoming defined. Before Presence, large organizations deploying AI agents had to build or buy their own governance and improvement infrastructure. Presence packages that into a managed offering. Expect Google (Vertex AI Agent Builder) and Anthropic (Claude Cowork for enterprise) to continue developing comparable managed offerings in response.

What to do right now

If you're an enterprise with an existing OpenAI relationship: contact your account team and ask about Presence. The limited GA means access is expanding, and early adopters are getting support directly from OpenAI engineers.

If you're building with the OpenAI API or another provider's API: the design principles Presence uses — policies, guardrails, simulation testing, and an improvement loop — apply to any agent deployment. You don't need Presence to implement them; you need discipline about defining scope, testing edge cases, and treating launch as the beginning of a review cycle, not the end.

If you're using ChatGPT Work or Claude Cowork for day-to-day professional tasks: nothing changes today. Presence operates at a level above what you're doing in a chat interface — it's infrastructure for organizations deploying agents to serve their customers or internal operations at scale, not a feature in your personal AI subscription.

---

Sources

• Help Net Security: OpenAI Presence connects AI agents to enterprise data with built-in guardrails (July 22, 2026)
• CX Today: OpenAI Launches Presence Amid AI Agent Safety Concerns (July 2026)
• The Decoder: OpenAI Presence wants to make AI agents production-ready for businesses (August 2, 2026)
• VentureBeat: OpenAI unveils Presence, a new platform that lets enterprises launch and manage realtime voice agents and chatbots (July 2026)
• OpenAI: Introducing OpenAI Presence (July 22, 2026)