OpenAI Presence: Enterprise Agents With Guardrails

In simple terms, OpenAI Presence is a managed AI operator for company phone calls and chats. It can qualify leads, answer customer questions, book appointments, update connected systems, and transfer difficult cases to a person. It is not a new AI model, and it is not only a cold-calling product.

Cold outbound sales is one possible use. The broader purpose is to give an AI agent one defined job, the information required to do it, and strict rules for what it may do without an employee.

What OpenAI Presence actually includes

OpenAI describes Presence as a deployed product for eligible enterprise customers. Each implementation begins with one specific job, such as resolving a billing question, supporting an insurance claim, or handling an internal IT request.

The agent receives only the knowledge and system access needed for that job. The company defines:

• what the agent may answer;
• which actions it may take;
• which actions require approval;
• when guardrails should intervene;
• when the conversation must move to a person;
• how quality will be measured after launch.

Presence currently supports real-time voice and chat experiences. OpenAI says it combines policies, standard operating procedures, approved actions, simulations, graders, guardrails, escalation rules, and a Codex-powered process for proposing improvements.

The important product difference

Presence is closer to an operating system for one production agent than a blank prompt box.

This is the same architectural lesson that applies to smaller projects: the model is only one component. A production agent also needs identity, permissions, observability, fallbacks, and someone accountable for changes.

What would a company actually use it for?

The strongest use cases have many similar conversations, a clear procedure, connected business systems, and an obvious point at which a person should take over.

Typical examples include:

• qualifying incoming leads before a sales manager calls them;
• following up with warm leads by phone or chat;
• running controlled outbound campaigns where consent and contact rules allow it;
• answering common customer-support questions;
• booking consultations or appointments;
• updating a CRM after a conversation;
• transferring valuable, unusual, or sensitive cases to an employee.

Sales and lead qualification

An AI operator can call or message a lead, ask several qualifying questions, record the answers in a CRM, offer an appointment, and notify a sales manager when the lead matches the company’s criteria.

For cold outreach, the important part is not simply making more calls. The system also needs consent rules, contact policies, approved claims, CRM logging, and strict handoff conditions. Presence can support this type of workflow, but it should not be treated as an unrestricted automatic dialler.

Customer support

An agent can understand a request, verify the customer, retrieve account context, apply policy, take an approved action, and escalate exceptions. The workflow is more valuable than an agent that only drafts an answer.

Internal service desks

IT and employee-support requests often have repeatable procedures and known escalation paths. They can be a safer first deployment than an open-ended customer agent, provided access remains scoped to the employee and request.

What the launch does not mean

Presence is not available as a normal API feature that any developer can switch on. As of July 29, 2026:

• access is limited to eligible enterprise customers;
• deployments are led by OpenAI Forward Deployed Engineers and selected systems integrators;
• the product is not self-service;
• voice customers can still build with OpenAI models through the API, but that is not the same managed package;
• published customer results come from OpenAI and its design partners, not an independent comparative study.

OpenAI reports that Presence handles its English-language phone support line and resolves 75% of inbound issues without human assistance. That is a useful example of the intended operating model, but it should not be used as a universal forecast. Resolution rate depends on request mix, policy scope, integrations, data quality, and how “resolved” is defined.

Questions to ask before buying or building

Use this checklist whether you evaluate Presence or build a similar system yourself:

1. What single job will the first agent own?
2. Which systems can it read from, and which can it change?
3. Which actions always require human approval?
4. What evidence proves that an answer or action is correct?
5. What triggers an immediate handoff?
6. Can every tool call and policy decision be reconstructed later?
7. How are regressions tested before an update reaches production?
8. Who can stop the agent and revoke its access?
9. What happens when a connected system is unavailable?
10. Which metric matters: containment, resolution, revenue, time saved, or risk avoided?

If these answers are unclear, the project is not ready for autonomous actions.

How FollowAI can help

FollowAI helps companies turn this type of agent into a working business process. We can:

• identify whether Presence, an OpenAI API voice agent, or a simpler automation is the right fit;
• connect phone, chat, WhatsApp, email, CRM, calendars, and company knowledge;
• design sales or support scripts, permissions, guardrails, and human handoffs;
• build and test a controlled pilot before increasing volume;
• train sales, support, and operations teams to use and supervise the system;
• maintain and improve the integration after launch.

Where enterprise access to OpenAI Presence is available and the use case fits, our team can support a Presence-based implementation. Where it is not available, we can build a comparable voice or chat workflow with OpenAI APIs and the company’s existing tools.

The best first step is one narrow scenario—for example, qualifying inbound leads or answering one repeatable support request—rather than replacing an entire department at once.

Who should wait

Presence is unlikely to be the right first step when request volume is low, procedures are undocumented, data is fragmented, or the business cannot define an acceptable outcome.

Start with a controlled workflow or an answer-drafting assistant instead. A reliable system can first retrieve information, prepare a recommendation, and ask a person to approve the final action. Our guide to RAG systems explains the knowledge layer, while the OpenAI–Hugging Face incident review shows why permissions and isolation matter.

Bottom line

OpenAI Presence makes the enterprise agent market more concrete. The product is not selling “intelligence” alone; it is selling the policies, evaluations, deployment work, escalation paths, and improvement process required to operate an agent safely.

That is the useful lesson even for teams that never buy it: define one job, minimize access, test realistic failures, keep a human fallback, and treat every production change as a controlled release.