OpenAI Presence aims to industrialize customer-service agents

OpenAI is launching Presence, a managed product for deploying voice and chat agents in enterprise processes. An agent can answer, verify a customer, inspect a system, apply a procedure, perform an approved action and transfer to a person when policy or risk demands it.

Presence is neither a new ChatGPT plan nor simply an SDK. OpenAI presents it as a combination of models, integrations, permissions, simulations, evaluations, monitoring and delivery support from its engineers or selected integrators. It targets high-volume workflows where a mistake can affect a bill, insurance claim or IT account.

The short answer

An agent starts with a narrow job

OpenAI describes billing resolution, insurance claims and internal IT support. Each agent receives only the knowledge and access required for its job. The company sets procedures, permitted actions and situations where human approval is mandatory.

Specialization matters. A general-purpose agent connected to every system would accumulate excessive privileges and failure modes. A narrow mission supports concrete test scenarios, limited APIs and measurable business outcomes.

A strong first scope has repetitive requests, explicit rules and reversible outputs. Exceptions should be rare enough that escalation does not overwhelm the human team.

Evaluation becomes an operations function

Before launch, Presence uses simulations and graders to check outcomes, policy adherence, tool use and escalation. Guardrails can intervene when an interaction leaves approved boundaries.

After launch, conversations, handoffs and quality signals reveal new cases. Codex investigates gaps and proposes changes. OpenAI says teams can compare a candidate against production behavior before approving a controlled rollout.

The important part is not automated editing but a repeatable cycle: observe, propose, test, approve, deploy and roll back. Without a stable evaluation set, improving one request type can silently damage another.

OpenAI's own support channel is the showcase

Presence powers OpenAI's English-language phone support. The company says it resolves 75% of inbound issues without human assistance and that a Codex-powered improvement loop reduced handoffs by 15 percentage points in ten days.

Those are vendor-reported figures and may not describe average request complexity. A customer needs its own metrics: correct resolution, repeat contact within seven days, satisfaction, erroneous refunds, duration and abandonment.

Automation rate cannot stand alone. An agent can lower handoffs by refusing early, discouraging users or taking the wrong action. Error cost belongs in the calculation.

Voice and chat share policy, not identical risk

Presence supports both channels. Voice adds latency, accent, noise and interruption constraints. Chat is easier to review but may encourage users to submit more documents or sensitive data.

A shared policy should produce consistent decisions while testing remains channel-specific. Identity verification acceptable in chat may not work by voice. The agent must identify itself clearly and offer human access without a punitive journey.

Contracts matter more than the demonstration

The help center states that models, channels, volume, data processing, location, retention, pricing and service commitments are defined per deployment. A general product page cannot establish compliance.

Approved architecture must document accessible data, logging, sensitive-data masking, trace access and retention. Technical identities need least privilege, and irreversible actions should require approval.

OpenAI says Business and API data is not used to train models by default, but customers still need to verify the Presence-specific agreement and subprocessors.

Which organizations need Presence?

The offer fits high volumes, repeatable processes and organizations able to maintain procedures and evaluation sets. A small company with a few varied requests may gain more from a strong knowledge base and routing than a complex managed deployment.

A large support center can justify the integration when systems expose suitable APIs, responsibilities are clear and a team truly owns the improvement cycle.

Presence shows that the model is no longer the hardest part of a production agent. Permissions, evaluations, escalations, contracts and controlled change determine whether fluid conversation becomes reliable service.