OpenAI launches Presence for enterprise agents

OpenAI is introducing Presence, a product that lets companies deploy AI agents capable of handling queries, using internal systems, carrying out approved actions and handing cases off to a person when necessary.

The proposal starts with a specific problem: proving that an agent works is no longer enough. To use one for important tasks, a company needs to control what it can do, check whether it responds correctly and update its behavior when its products, policies or customers change.

An agent with defined limits

Each Presence deployment is designed for a specific task, such as resolving billing issues, processing insurance claims or handling internal IT support requests.

The agent receives only the knowledge and system access it needs for that job. The company decides:

• Which actions it can carry out on its own.
• When it must request authorization.
• Which situations require human intervention.
• How the quality of its responses is evaluated.

For example, when handling a billing query, Presence can understand the issue, verify the customer’s identity, check their data, apply the company’s rules and carry out an authorized action. If the case falls outside the defined limits, it must escalate it.

This matters because an error in a sales conversation can be annoying, but an error when modifying a bank account, approving a claim or handling an internal issue can have real consequences.

The system is also reviewed after launch

Presence includes simulations and evaluation tools to test the agent before putting it in front of users. These tests cover common requests, unusual cases and higher-risk situations.

The checks determine whether the agent reached the correct result, followed the company’s policies, used its tools properly and requested human help when appropriate. So-called guardrails can block or redirect an interaction that moves outside the established limits.

Monitoring continues after launch. Real conversations, escalations and other quality signals show where the system is failing or where people need more help. With the Presence add-on, Codex can investigate those signals and propose changes.

The company’s team tests each modification against the active version and decides whether to approve it. This allows the agent to adapt without automatically changing its behavior in production and without supervision.

Results from OpenAI’s phone support

OpenAI says Presence powers its English-language phone support channel, available at 1-888-GPT-0090. The system handles open-ended requests, verifies callers, checks account context and carries out approved actions.

According to the company, within a few weeks it reached or exceeded the internal standards used to evaluate the quality of first-line human support. It currently resolves 75% of incoming queries without human help.

In addition, the Codex-based improvement cycle reduced transfers to human agents by 15 percentage points in 10 days, according to OpenAI. These figures apply to that specific channel, not to all Presence deployments.

Early enterprise use cases

Several companies are exploring the same platform for different uses:

• BBVA is analyzing AI phone support for everyday banking operations in Mexico.
• SoftBank is testing natural conversations in Japanese.
• IAG is studying support during periods of high demand, such as extreme weather events.

The system currently works across voice and chat experiences, including customer service, outbound sales and high-risk internal processes. Companies can retain shared policies, evaluations and escalation rules while adapting the agent to each channel or workflow.

It is not a self-service tool

Presence is available to eligible companies through a limited general availability program. Deployments are led by OpenAI engineers who specialize in working within each customer’s operations, together with some global systems integrators.

For now, it cannot be activated as a self-service product. OpenAI is involved in connecting the company’s systems, defining permissions, testing and putting the deployment into production.

For you, the development points to a clear trend: enterprise agents are no longer sold only as chatbots that answer questions. The challenge is integrating them with real processes, giving them measurable limits and keeping them under control as the environment changes. The question to watch is whether this guided deployment model can scale without turning every implementation into a long and costly project.