The product is the deployment, not just the agent

OpenAI has launched an enterprise agent product that arrives with people, process and a fairly long setup list.

Presence is designed for repeatable workflows where an agent may need to answer a customer, look up account information, apply a policy, change a record or hand the case to a person. The first supported channels are voice and chat.

It is not a self-serve tool. OpenAI says its forward deployed engineers, selected systems integrators or both work with each customer to define the job and put the agent into production.

That makes Presence closer to a managed operational service than another chatbot builder. The model matters, but so do the permissions, system connections, acceptance tests, escalation paths and later updates.

OpenAI is making a clear business bet: for important enterprise work, the difficult part is no longer getting an agent to produce a plausible answer. It is making the whole system dependable enough to use every day.

A narrow job comes before broad intelligence

Each Presence deployment begins with a specific task. OpenAI gives examples such as billing support, insurance claims and employee IT requests.

The agent receives the knowledge and system access needed for that task. The customer defines what it may do, when approval is required and when a person takes over. Depending on the deployment, the agent can retrieve information, update a system and complete an approved action.

This narrow framing is easy to miss amid talk about capable agents. It is also one of the more practical safeguards. An agent that can reset a password does not automatically need access to payroll, procurement or every customer record.

Presence includes simulations, evaluations, guardrails, session histories and controlled rollouts. The help documentation says deployment also includes security, privacy and legal review.

None of those steps guarantees a correct outcome. They make failures easier to look for before launch and easier to investigate afterwards.

The improvement loop still needs approval

OpenAI says production conversations and escalations reveal where an agent struggles. Codex can then investigate those signals and propose changes.

A team can test a proposed change against the live version before approving a controlled rollout. This is an important detail. The system is described as improving with use, but it is not supposed to rewrite itself in production without a review gate.

The underlying idea is ordinary software operations applied to less predictable software: observe, find a failure, change one thing, test it and keep a way back.

Presence also includes human handoffs. That sounds basic, yet the quality of the handoff often decides whether automation saves time. A useful escalation needs the right context, a clear reason and a person who is allowed to act.

OpenAI has not published a universal service level or handoff design. Its help page says the integration, routing, authentication and escalation setup are agreed for each deployment.

The strongest numbers come from OpenAI’s own support line

OpenAI says Presence powers its English-language phone support channel. According to the company, the agent now resolves 75% of inbound issues without a person and its improvement loop reduced human handoffs by 15 percentage points in ten days.

Those figures are specific and useful. They are also internal results. OpenAI has not published the sample size, issue mix, scoring method or an independent audit alongside the launch.

The named external customers are earlier in the story. BBVA is exploring voice support in Mexico. SoftBank is testing conversations in Japanese. IAG is exploring help during periods such as severe weather. Their statements describe design work and trials, not a completed roll-out with comparable performance figures.

BBVA has a wider relationship with OpenAI and is a founding partner in its deployment company. That context helps explain why the bank is close to the product, but it is not evidence that Presence will work the same way in another organisation.

The early case is promising. It is still mostly the vendor’s case.

What is confirmed, what OpenAI says and what remains open

Presence exists as a limited general-availability product for eligible enterprise customers. It supports managed voice and chat deployments, connects to approved systems, includes testing and escalation controls, and is not available as a self-serve product.

The resolution and handoff improvements are OpenAI’s claims about its own support operation. The descriptions from BBVA, SoftBank and IAG are partner statements published by OpenAI. Model Current has not independently audited those deployments.

Pricing is not public. Neither is a fixed model list. OpenAI says capacity, data handling, retention, location, security details and service commitments are defined for each customer.

There is no public evidence yet on failure rates across different industries, how much staff time a deployment needs after launch or whether the economics still work for workflows with many exceptions.

Presence is interesting because it makes the hidden work visible. The agent is one part. Someone still has to decide what it can touch, what good looks like and when it should stop.

Sources

  1. OpenAI — Introducing PresencePrimary product announcement published 22 July 2026. Source for product scope, internal support metrics, named design partners, improvement workflow and limited availability.
  2. OpenAI Help Center — PresencePrimary operating documentation for permissions, deployment steps, supported channels, data-handling variability, access and non-self-serve status.
  3. OpenAI — Presence product pagePrimary product page describing use cases, evaluation and escalation controls, design-partner statements and forward-deployed support.
  4. OpenAI — Enterprise privacyPrimary statement of general business-data commitments, including default model-training policy, access controls and encryption.
  5. BBVA — OpenAI deployment partnershipPrimary bank announcement from May 2026 giving context on BBVA’s role as a deployment-company founding partner and its wider OpenAI collaboration.