OpenAI Presence brings managed AI agents to enterprise workflows

OpenAI has introduced OpenAI Presence enterprise agents, a managed offering for putting AI agents into customer-facing and internal workflows. Announced July 22, the product is available in limited general availability to eligible enterprises, but it is not self-service and OpenAI has not disclosed pricing, implementation costs or service terms.

Presence is OpenAI’s attempt to sell the operational layer around an agent, not merely model access. The company says the product supports real-time voice and chat work such as customer support, outbound sales and higher-risk internal requests. Deployments are led by OpenAI Forward Deployed Engineers, with select systems integrators able to support customers as deployments expand.

What are OpenAI Presence enterprise agents?

Presence is a separately scoped and managed product for production workflows, rather than an agent made through the ChatGPT or Slack Workspace Agents interface. OpenAI describes it as combining its models with company policies, guardrails, approved actions, simulations, evaluation tools and rules for transferring a case to a person.

The premise is narrow: an agent starts with a defined job, such as handling a billing issue, an insurance claim or an employee IT request. OpenAI says the customer determines the knowledge and system access needed for that job, what actions the agent may take, when it needs approval and when a human must take over. Exact integrations and capabilities are confirmed during technical scoping.

How the managed deployment model works

OpenAI says its engineers and deployment partners work with customers to define outcomes and initial workflows, connect systems, encode permissions and escalation paths, complete security, privacy and legal reviews, then test and stage a rollout. Its help documentation says feeding documents to an agent alone does not make it ready for production.

Before release, OpenAI says teams can test agents against routine requests, edge cases and higher-risk scenarios. The company describes simulations and graders that assess whether an agent reached the intended outcome, followed policy, used tools correctly and escalated appropriately. Guardrails can intervene if an interaction falls outside the customer’s defined boundaries.

After launch, OpenAI says production sessions, escalations and quality signals can show teams where an agent needs attention as business policies, products or user behavior change. Codex can propose updates based on those signals; teams test proposed changes against the production version and approve a controlled rollout. OpenAI’s documentation also lists monitoring and rollback among controls that may be included depending on the deployment.

What is available, and what remains unclear?

During limited GA, Presence supports voice and chat. Access depends on workflow fit, implementation readiness and available delivery capacity, according to OpenAI’s help documentation. The company says data access, logging, masking of sensitive information, retention, storage and access controls are reviewed for each deployment, rather than standardized in the public product description.

Commercial details are similarly customer-specific: OpenAI says exact features, model configuration, channels, capacity, data handling, pricing and service commitments are set per deployment. That leaves buyers without a published price card or common service baseline.

OpenAI says Presence powers its English-language phone-support line and resolves 75% of inbound issues without human assistance. It also says a Codex-powered improvement process reduced handoffs by 15 percentage points in 10 days. Those are company-reported results, with no independent validation in the available material. BBVA is exploring banking voice support in Mexico, while SoftBank is testing Japanese-language conversations and IAG is exploring support during high-demand events.

This story draws on original reporting from The Decoder.