OpenAI introduced Presence on July 22, 2026 as an enterprise product for deploying AI agents in customer and internal workflows. It combines voice or chat agents with company policies, limited system access, approved actions, testing, guardrails, and human escalation. Presence is not a new ChatGPT button for ordinary users, and it is not currently a self-service product. The practical change is that some support bots may move beyond answering questions and start resolving carefully defined requests.
What is OpenAI Presence?
OpenAI Presence is a managed product for eligible enterprises that want to run AI agents in real business processes. An agent is different from a basic chatbot because it can follow a multi-step workflow, use connected tools, and take an allowed action on the user’s behalf.
For example, a normal support bot may explain a refund policy. An agent may verify the customer, look up the order, check the policy, submit an approved refund, and escalate the case when it falls outside its authority.
OpenAI says Presence currently supports real-time voice and chat experiences for jobs such as customer support, outbound sales, insurance claims, and employee IT requests. Each deployment begins with a specific job rather than an instruction to “handle everything.” This limited scope matters because an agent that can change records or approve actions needs tighter controls than a question-answering bot.
How Presence is designed to work
The model is only one part of the product. Presence adds an operational layer around the model so a company can define what the agent knows, what it can do, when it must ask for approval, and when a person should take over.
- Choose one job: the company starts with a specific workflow, such as resolving a billing question.
- Limit knowledge and access: the agent receives only the company information and system access required for that job.
- Set policies and approved actions: the company defines the rules, boundaries, and actions the agent may take.
- Test before launch: simulations and evaluations check normal requests, edge cases, policy compliance, and tool use.
- Escalate when necessary: higher-risk or unsupported requests move to a human.
- Review real failures: production sessions and escalations reveal gaps; proposed changes can be tested before controlled rollout.
This structure reflects a broader lesson in the current AI-agent market: the hardest problem is no longer producing a fluent reply. It is controlling access, measuring whether the outcome is correct, and maintaining the workflow when products or policies change.
Worked example: an AI agent handles a billing dispute
OpenAI uses billing support as an example in its announcement. The following walkthrough is an illustration of that described workflow, not a personal test or a transcript from a real customer.
| Stage | Agent action | Control that should exist |
|---|---|---|
| Understand | Identifies whether the customer is asking about a duplicate charge, a refund, or another issue. | If the intent remains unclear after a limited number of attempts, transfer to a person. |
| Verify | Uses the company’s approved identity-verification process. | Do not reveal account information before verification succeeds. |
| Retrieve | Looks up only the account and transaction information needed for the request. | Use minimum necessary access and record which system actions occurred. |
| Apply policy | Checks the current billing rules and determines the permitted next step. | Escalate exceptions instead of inventing a policy or making an unsupported promise. |
| Act or escalate | Takes an approved reversible action, requests authorization, or hands the case to a person. | Sensitive, irreversible, or high-value actions should require human oversight. |
A useful agent does not need unlimited authority. It needs enough access to complete one defined job, plus a reliable exit when the request becomes uncertain or risky.
Presence versus a chatbot and a custom AI agent
| Option | Typical role | Who builds and maintains it? | Main limitation |
|---|---|---|---|
| Basic support chatbot | Answers common questions and directs users to help articles. | The business or its software provider. | Usually cannot complete a multi-step action across company systems. |
| Custom API-built agent | Uses models and tools in a workflow designed for a particular organization. | The organization’s technical team or development partner. | The team must assemble, test, secure, monitor, and update the system. |
| OpenAI Presence | Runs defined voice or chat workflows with policies, evaluations, approved actions, and escalation. | OpenAI Forward Deployed Engineers and selected systems integrators work with the enterprise. | Limited availability; it is not currently a self-service product. |
What the announcement means for ordinary customers
Most readers will not buy Presence directly. They may encounter it when contacting a large company by phone or chat. A well-designed deployment could reduce the frustrating cycle of repeating a problem, opening several help pages, and waiting for a person to perform a routine action.
However, a natural voice does not prove that the answer is correct or that an action is authorized. Customers should still receive clear confirmation before a charge, cancellation, refund, claim update, or account change. They should also have a practical way to reach a person.
If you meet an AI support agent, ask four simple questions:
- Am I speaking with an AI system?
- What action are you about to take?
- Will I receive written confirmation?
- Can you transfer me to a person if this is not resolved?
What Presence does not mean
It is not a consumer ChatGPT feature. OpenAI says Presence is available to eligible enterprise customers through limited general availability and is not self-service.
It is not unrestricted autonomy. The company using it chooses the knowledge, access, actions, policies, approvals, and escalation rules for each workflow.
It is not proof that every support interaction can be automated. OpenAI reports that Presence powers its English-language phone support and resolves 75% of inbound issues without human assistance. It also reports a 15-percentage-point reduction in human handoffs over ten days after its improvement process. These are vendor-reported results from one deployment, not a guarantee for every company, language, or workflow.
It does not remove the need for people. Human intervention remains important when the agent exceeds failure limits or reaches a sensitive, irreversible, or high-impact decision.
A practical decision framework for businesses
A small business should not read this launch as a reason to replace its entire support process. Presence is not currently self-service, and many small teams may be better served by improving their help pages, contact forms, and simple automations first.
A larger organization should begin with one narrow, high-volume workflow where the correct outcome can be clearly tested. Good candidates have documented policies, well-defined actions, measurable success, and an obvious human escalation path. Poor candidates involve unclear responsibility, constantly changing rules, or decisions where a mistake could seriously affect a person.
This is the same reason I recommend comparing AI tools by workflow rather than headline. My earlier analysis of recent AI model releases and what they mean for users explains why a stronger model does not automatically produce a better real-world system.
Five safeguards worth checking
- Minimum access: the agent should see only the data required for its job.
- Action limits: sensitive or irreversible actions should require approval.
- Human escalation: the user must have a clear route to a person.
- Pre-launch testing: evaluations should cover normal requests, edge cases, and policy failures.
- Ongoing review: real failures should lead to tested changes rather than silent, uncontrolled updates.
OpenAI’s own practical guide to building agents recommends layered guardrails and human intervention, especially after repeated failures or before high-risk actions. Those principles matter regardless of which company supplies the model.
Frequently asked questions
What is OpenAI Presence used for?
It is used by eligible enterprises to deploy voice and chat AI agents for defined customer or internal workflows, such as billing support, insurance claims, sales conversations, or employee IT requests.
Is OpenAI Presence available to everyone?
No. At launch, OpenAI says it is available to eligible enterprise customers through a limited general availability program. It is not a self-service product.
Is OpenAI Presence the same as ChatGPT?
No. ChatGPT is a general AI assistant. Presence is an enterprise deployment product that combines models with company systems, policies, evaluations, guardrails, approved actions, and escalation rules.
Will AI customer service agents replace human support?
The announcement does not establish that conclusion. Presence is designed to escalate cases when required, and high-risk or uncertain situations still need human oversight. The likely near-term pattern is automation of defined requests with people handling exceptions and sensitive decisions.
Practical takeaway: OpenAI Presence is important because it packages the less glamorous parts of AI agents—permissions, policies, testing, approved actions, monitoring, and escalation—around voice and chat models. For customers, the real test will not be how human the agent sounds. It will be whether the system completes the correct action, explains what happened, protects private information, and reaches a person when it should.
Official sources: OpenAI’s Presence announcement and OpenAI’s practical guide to building AI agents.
Join the conversation
Post a Comment