NEW

OpenAI releases enterprise agents with human control

OpenAI Presence is designed to help large organizations put AI agents in customer service and internal flows. The service is available to selected business customers and cannot be purchased as a standard self-service service.

Publicerad 24 July 2026, 09.44

AI-generated illustration of voice and chat flows passing company rules, approved actions and a point of human handover.

OpenAI has launched Presence, an enterprise product for AI agents in voice and chat services. An AI agent is a system that not only responds, but can also use approved tools and perform defined steps. OpenAI mentions customer service, insurance matters and internal IT support as examples.

Each entry begins with a specific task. The company chooses which knowledge and which systems the agent can use, which actions it can perform and when a human must approve or take over. That differentiates Presence from a regular chatbot that only leaves text.

Presence contains rules, security barriers, simulations and tests to check whether the agent follows the company's way of working. After launch, real conversations and handovers can show where the system is failing. Codex can then suggest changes that the team tests and approves before rolling them out.

OpenAI states that its own English-language phone support solves 75 percent of incoming cases without human help. It is the company's own measurement and there is no independent evidence that the same results apply to other organizations, languages ​​or workflows.

The service is available to select business customers through a program with limited general availability. OpenAI engineers and some system integrators are leading the deployment. Presence is not available as a self-service service, and OpenAI does not publish a standard price or a list of available countries.

For Swedish businesses, the important question is therefore not whether an agent can speak naturally, but who determines authorizations, reviews results and takes responsibility when something goes wrong. A cautious pilot project should start with a narrow task, innocuous test data, and clear handoff to a human.

Därför spelar det roll

Business agents go from typing responses to using systems and executing actions. It can save time, but an error can also affect a customer, payment or internal process. Presence shows that permissions, testing and human escalation need to be built in before launch.

Det här kan du göra

  1. Choose a narrow task and write down exactly what the agent can and cannot do.
  2. Test common questions, borderline cases and errors with harmless data before letting real customers in.
  3. Require human approval for sensitive or irreversible actions and track each handover.