Updated July 26, 2026. This article was prepared by the AI Dawn editorial team based on OpenAI’s official announcement and help center. The stated Presence results are OpenAI’s data, not an independent benchmark test.
OpenAI Presence is a managed enterprise platform for building, launching, and continuously improving AI agents in high-volume, high-stakes workflows. An agent can speak by voice or in chat, access approved systems, perform authorized actions, and hand the conversation off to a human. Presence is available to a limited group of enterprise customers and is not yet sold as a standalone service with public pricing.
The key difference between Presence and a standard chatbot is not the model, but the operational layer around it: rules, access permissions, simulations, quality evaluations, action history, controlled version releases, rollback, and human escalation. OpenAI says its English-language phone support powered by Presence resolves 75% of incoming requests on its own, and a Codex-based improvement cycle reduced transfers to an operator by 15 percentage points in 10 days.
In brief: Large companies with repetitive request volumes, integrations, and formal rules should consider Presence. Small businesses that need a single FAQ bot or simple automation will likely find it more practical to start with an API, an off-the-shelf platform, or a local pilot.
Contents
- What Is OpenAI Presence
- Why the Topic Is Heating Up
- What Presence Can Do
- How Agent Management Works
- Presence, ChatGPT Workspace Agents, and API
- Who the Platform Is Best For
- Known Limitations
- How Implementation Works
- What Metrics to Set Before a Pilot
- Company Readiness Checklist
- FAQ
- Bottom Line
What Is OpenAI Presence
OpenAI introduced Presence on July 22, 2026, as a product for production enterprise workflows. In OpenAI’s official announcement , the platform is described as a combination of models, policies, constraints, allowed actions, simulations, evaluations, and a Codex-based improvement process.
Presence is not a new general-purpose LLM and does not replace ChatGPT. It is a managed deployment that OpenAI Forward Deployed Engineers and selected partners configure for a specific job: invoice processing, insurance claims, IT service desk requests, outbound sales, or customer support.
Each agent has a limited operating scope:
- one or more preapproved tasks;
- only the necessary knowledge and systems;
- read and write permissions for data;
- rules that require approval;
- conditions for handing off to a human;
- pre-release tests and post-launch analysis.
This design addresses a core production-agent problem: a good answer in a demo does not mean the system can safely handle thousands of different requests, update the CRM without duplicates, or stop correctly on exceptions.
Why the Topic Is Heating Up
Before 2026, the market often sold the “enterprise agent” as a prompt builder, a knowledge base, and a few API integrations. Presence shifts the conversation to operations: verifiable behavior, action logs, version control, and improvement based on real signals.
This matters for three reasons.
- OpenAI is moving deeper into the integration layer. The company is selling not just the model and API, but process implementation together with engineers and systems integrators.
- Voice and chat are combined into one operational layer. Rules, evaluations, and escalations can be kept consistent across channels while the specific workflow changes.
- Agent improvement becomes a managed cycle. Codex reviews failed sessions and suggests changes, but the team tests them and approves the release.
For businesses, this means comparing solutions only by speech quality or demo answers is no longer enough. Procurement now raises questions about rollback, logging, access boundaries, and proof of results.
What Presence Can Do
According to OpenAI’s help documentation, a Presence agent is configured around five groups of actions.
| Capability | What the Agent Does | What the Company Must Define |
|---|---|---|
| Policy Compliance | applies SOPs and internal policies | the current version of the rules and the owner |
| Knowledge Work | searches for data in approved sources | source list and permissions |
| System Actions | reads and updates records through API and tools | approved operations and limits |
| Voice and Chat | holds a conversation in a supported channel | identity verification and routing flow |
| Escalation | passes the task to a human with context | triggers, queue, and SLA |
Presence supports voice and chat workflows during its limited availability phase. Specific contact centers, authentication methods, routing, and the model are selected during the technical assessment.
The platform can combine a real-time model for conversation with a deeper reasoning model or tool execution model. OpenAI does not publish one fixed model set for all deployments.
How Agent Management Works
A production agent goes through a four-layer cycle.
1. Policy
The company defines what the agent must do, what it is forbidden to do, and where approval is required. The policy must be enforceable: the phrase “act reasonably” does not replace a refund limit, a list of supporting documents, or an identity verification rule.
2. Pre-Launch Validation
Simulations and evaluations test common requests, edge cases, and risky scenarios. The focus is not the elegance of the wording, but the outcome: did the agent apply the rule correctly, call the tool, update the right record, and hand the conversation to an operator?
3. Post-Launch Monitoring
Session, action, evaluation, and escalation histories show where the agent makes mistakes. This data is then used as input for the next version.
4. Controlled Improvement
Codex with the Presence plugin analyzes signals and suggests changes. The team compares the new version with the current one, runs tests, approves the rollout, and rolls back the release if needed.
AI Dawn’s practical takeaway: an agent becomes a production system only when it has a version, a test suite, an owner, a change log, and a shutdown procedure.
Presence, ChatGPT Workspace Agents, and API
OpenAI draws a clear line between Presence and ChatGPT workspace agents. This comparison helps you avoid buying a heavier product than you actually need.
| Approach | How it’s built | What it’s best for | Main trade-off |
|---|---|---|---|
| OpenAI Presence | with OpenAI or a partner | high-volume, controlled workflows with integrations | no self-service and no public pricing |
| ChatGPT Workspace Agents | inside a supported ChatGPT or Slack workspace | personal and team workflows | not a managed production deployment of Presence |
| OpenAI API | custom development | product teams with engineering resources | the customer owns all operational responsibility |
| Ready-made industry platform | through a solution provider | fast launch of a standard use case | less flexibility and dependence on the vendor’s product |
Presence does not replace the API: OpenAI separately says it will keep access to voice models through the API. So the choice is not about which model is smarter, but who will own integration, testing, security, and ongoing improvement.
Who the platform is for
Presence makes sense if several conditions are true at the same time:
- the volume of requests is high and repetitive;
- the process creates measurable business value;
- there are formal rules and a process owner;
- the agent needs to work across multiple systems;
- an incorrect action carries a meaningful cost;
- the company is prepared for security, privacy, and legal review;
- there is a team that will handle exceptions after launch.
Strong starting use cases:
| Use case | Why it fits | Human gateway |
|---|---|---|
| Billing support | rules and data are usually well defined | disputed refund or suspected fraud |
| Insurance claim | many repetitive steps and documents | financial or legal risk in the decision |
| IT service desk | standardized tickets and access requests | privilege escalation |
| Voice support | high volume and measurable handoffs | an emotional or unusual conversation |
| Outbound sales | repetitive qualification | aligning on terms and commitments |
Presence is unlikely to be the first choice for low volume, an uncontrolled process without an owner, work without a stable API, or a scenario where almost every decision requires expert judgment.
What limitations are known
As of July 26, 2026, OpenAI had defined several boundaries.
- Limited availability. Presence is available to qualified enterprise customers through limited GA.
- No self-service signup. Access is discussed with OpenAI’s account team.
- No public pricing. Implementation cost and scope are calculated for each customer.
- No single integration specification. Channels, models, volume, data retention, and SLA depend on the project.
- OpenAI data is not the same as an independent benchmark. The 75% self-resolution figure and the 15 percentage point reduction in handoffs refer to OpenAI’s own English-language support.
- Documentation does not make the agent production-ready. OpenAI’s documentation separately emphasizes the need for review, integration, testing, security review, and approval.
These limitations should be included in the procurement checklist. Without them, a polished case study can easily turn into a promise the vendor never made.
How implementation works
OpenAI describes six stages of managed deployment.
- Define the business outcome, success criteria, and first process.
- Connect systems, transfer policies, permissions, and escalation paths.
- Complete security, privacy, and legal review.
- Run simulations, evaluations, and acceptance testing.
- Launch a limited production rollout and measure results.
- Improve the agent based on evidence, not isolated impressions.
AI Dawn recommends starting with one job and one channel. For example, not “automate the contact center,” but “handle order-status requests after successful verification, read three fields, and hand off to an agent in two predefined exception cases.”
What metrics to define before the pilot
Metrics should separate speed, outcome, risk, and cost.
| Category | Example metric | Why it matters |
|---|---|---|
| Outcome | share of requests resolved without repeat contact | shows true task completion |
| Quality | share of policies applied correctly | protects against fast but incorrect answers |
| Escalation | share and reasons for handoffs to a human | shows the limits of automation |
| Risk | number of unresolved actions | checks access control behavior |
| Speed | time to result | compares the agent with the current process |
| Cost | total cost of one resolved request | includes models, integrations, and operators |
| Stability | change in quality between versions | catches regressions after improvements |
You cannot optimize only for containment rate. If the agent stops handing off complex cases to a human, the metric may rise while risk rises too. You need a pair: independent resolution plus verified outcome quality.
Company readiness checklist
Before speaking with OpenAI or an integrator, answer "yes" or "no" to 10 questions.
- Is there one process with an owner?
- Is the monthly volume and current cost known?
- Are the rules documented and free of conflicts?
- Do the systems have documented APIs?
- Can permissions be limited to the minimum necessary?
- Are the actions that require a human defined?
- Is there a test set of common and rare cases?
- Can the company store and review session logs?
- Are the data and retention requirements agreed on?
- Is a budget set aside for post-launch support?
If there are more than three negative answers, the process should be cleaned up first. AI does not fix unclear accountability; it scales the consequences faster.
FAQ
Is OpenAI Presence already available?
Yes, but only to suitable enterprise customers through a limited general availability program. It is not a self-service product.
How much does OpenAI Presence cost?
There is no public price list. Pricing and scope are determined for each specific implementation.
Does Presence only work with voice?
No. During limited GA, voice and chat use cases are supported.
How is Presence different from a regular chatbot?
Presence includes managed implementation, policies, permissions, simulations, evaluations, logging, escalations, rollout, and rollback. A chatbot may not have these operational controls.
Can a similar system be built through the API?
Yes. But the customer will need to implement integrations, permissions, testing, observability, versioning, and support on their own.
Will Presence replace support agents?
OpenAI does not promise full replacement. The platform is specifically designed with handoff to a human where judgment, approval, or risk handling is needed.
Bottom line
OpenAI Presence shows a mature form of an enterprise AI agent: not a standalone model, but a managed system with permissions, tests, escalations, and version control. The platform is interesting for large organizations with high-volume processes, but at the outset it requires formalized workflows, technical readiness, and a budget for managed deployment.
The first step is to choose one narrow process and calculate its current cost, errors, handoffs, and resolution time. If the baseline metrics are unknown, you cannot compare a pilot to reality.