OpenAI Presence: Inside the Enterprise Platform for Deploying and Managing AI Agents

OpenAI has introduced Presence, an enterprise platform built to help organizations deploy, manage, and improve AI agents across customer support, sales, IT, and other business operations. It represents the company's deepest move yet into corporate software, shifting OpenAI from a vendor that sells access to models toward one that supplies an end-to-end system for running agents in live production environments.

What OpenAI Presence Actually Provides

At its core, Presence gives companies a shared foundation for agent operations: policies, guardrails, permissions, and evaluation tooling that apply consistently across both voice and chat channels. Rather than treating each agent as a bespoke build, the platform standardizes the controls that govern how agents behave once they're talking to real customers or real employees.

Every deployment starts narrow. A company defines one specific job — resolving a billing dispute, working through an insurance claim, handling an employee IT request — and the agent is granted only the knowledge and system access that particular job requires. Nothing more. That scoping decision is the platform's central design choice, and it addresses the most common enterprise objection to agentic AI: an agent with unbounded access to internal systems is a liability, not an asset.

From there, the organization writes its own rules. Teams decide what actions the agent is permitted to take on its own, which actions require a human to sign off first, and what conditions should trigger an escalation to a person. Agents can answer questions, work inside company systems, carry out approved actions, and hand off when the situation calls for it.

Voice and Chat Under One Set of Controls

Because Presence covers both voice and chat, a company doesn't need separate governance regimes for a phone line and a support widget. The same permissions, the same approval thresholds, and the same escalation logic carry across customer-facing and internal workflows alike.

The Codex Improvement Loop

What separates Presence from a straightforward deployment tool is what happens after launch. The platform runs a Codex-powered improvement loop that treats live traffic as diagnostic material.

The cycle works like this:

• Production sessions and escalations expose gaps. Every handoff to a human is a signal that the agent hit a limit.
• Codex proposes updates based on those gaps.
• Teams test the proposed changes against the live version before anything ships.
• A rollout proceeds only after approval.

That final step matters. The loop is not autonomous self-modification — a human approves the change, and the comparison against the current live agent gives teams evidence before they commit.

The Numbers OpenAI Is Citing

OpenAI is using its own operation as the proof point. Presence powers the company's English-language phone support line at 1-888-GPT-0090, where it resolves 75% of inbound issues without any human involvement. The Codex improvement loop cut human handoffs by 15 percentage points over a span of 10 days.

That second figure is arguably the more interesting one. A 75% resolution rate describes a system's current performance; a 15-point drop in handoffs across ten days describes how quickly the system closes its own gaps. For enterprise buyers weighing deployment cost against time-to-value, the rate of improvement is often the number that decides the purchase.

Early Enterprise Customers

Several large organizations are already building on the platform, and the initial use cases span very different operating conditions:

Each of these stresses a different capability. BBVA's case puts the platform in a regulated financial environment where permissions and approval gates carry real consequence. SoftBank's tests push conversational quality in Japanese, well beyond the English deployment OpenAI runs internally. IAG's interest centers on volume surges — the moments when a support organization is overwhelmed and escalation paths get saturated.

How Companies Can Access Presence

Presence is available to eligible enterprise customers through a limited general availability program. It is not a self-serve product, and there's no path to spin one up independently.

Deployments are led by OpenAI's Forward Deployed Engineers alongside a set of selected global systems integrators. That delivery model tells you something about the product's current maturity: this is white-glove implementation work, not a sign-up flow. It also puts a natural ceiling on how fast OpenAI can scale the offering, since growth is bound by engineering capacity and partner readiness rather than server provisioning.

Where Presence Fits in OpenAI's Enterprise Strategy

Presence builds on groundwork OpenAI laid earlier. In February, the company introduced Frontier, its platform for building and managing AI agents. Presence extends that lineage from construction into operation — governing, evaluating, and refining agents that are already handling live workloads.

The launch also aligns with the direction CFO Sarah Friar has been describing. Friar framed 2026 as a year focused on "practical adoption," language that signals a pivot from capability demonstrations toward measurable deployment inside real businesses.

The revenue picture supports that framing. Enterprise customers made up roughly 40% of OpenAI's business as of January, and Friar expected that share to approach 50% by year's end. Products like Presence are how that number moves — recurring, operationally embedded software rather than model access sold by the token.

From Model Provider to Infrastructure Vendor

The strategic repositioning is the real story. Selling a model puts a company at the top of a stack that customers assemble themselves, competing on benchmarks and price. Selling deployed, managed AI infrastructure puts the vendor inside the customer's operations, where switching costs are high and the relationship is measured in resolution rates rather than API calls.

With Presence, OpenAI is making a claim on that second position: not just the provider of intelligence, but the operator of the systems that put it to work.