July 23, 2026. OpenAI on July 22 introduced Presence, an enterprise platform for deploying and managing trusted AI agents across voice and chat. It is not a new model. It is the scaffolding companies need to put agents into production safely: policies and standard operating procedures, guardrails, approved actions, simulations, evaluations, and a review process that keeps improving the agent after launch. VentureBeat and Help Net Security reported the same framing. For anyone running customer facing automations, the news is less the product and more the standard it sets.
What OpenAI launched
- A deployment platform, not a model, that connects agents to company systems and defines how they should behave across voice and chat for customer support and internal service requests.
- Built in guardrails that intervene when a conversation moves outside the company's boundaries, plus least privilege access, so an agent can only reach the data and systems a given workflow requires.
- Simulations and evaluation tools to test agents before they go live, rather than learning on real customers.
- A post launch loop where OpenAI's Codex reviews production sessions and escalations and suggests improvements that staff test and approve before anything changes.
- Proof by dogfooding: OpenAI says Presence already runs its own English language phone support and resolves 75 percent of inbound calls without a human, a vendor claim on its own line. The platform is launching through a limited availability program, not as a self serve product.
What it means for operators
Most small and midsize businesses will not get into a gated enterprise program, and that is fine, because the valuable part is the checklist, not the login. Presence is OpenAI codifying what a production customer facing agent actually requires, and you can hold your own builds to the same bar today on the tools you already run, whether that is GoHighLevel Conversation AI and Voice AI or a custom stack. The pattern to copy: write the standard operating procedures the agent must follow, set guardrails and a human escalation path, scope its data access to the minimum, simulate and evaluate before launch, review escalations every week, and require a human to approve behavior changes. That is the difference between an agent that deflects calls and one that quietly damages the brand. It is exactly how our AI automation agency ships AI automations for clients.
The honest read
The 75 percent resolution figure is OpenAI's own, measured on its own support line, so treat it as a target to test, not a promise. And a governed agent is still only as good as the policies and evaluations behind it: the platform makes discipline easier, it does not replace it. If you are deploying a voice or chat agent that talks to customers, build the guardrails first and the features second.
Frequently Asked Questions
Presence is an enterprise platform OpenAI launched on July 22, 2026 for deploying and managing trusted AI agents across voice and chat. It is not a model. It bundles policies, guardrails, approved actions, simulations, evaluations, and a post launch review loop so companies can run customer facing agents safely.
Not directly yet. Presence is launching through a limited availability program, not a self serve product, so it is aimed at larger deployments. Smaller teams can still adopt its approach, guardrails, least privilege data access, evaluations, and human approved changes, on tools they already run such as GoHighLevel or a custom build.
A chatbot builder focuses on conversation flows. Presence focuses on governing an agent in production: it enforces policies, limits what data the agent can touch, tests it with simulations and evaluations before launch, and reviews real sessions to suggest changes that a human approves. It is a deployment and control layer, not just a bot editor.
No. The useful part is the standard it sets. You can apply the same checklist, standard operating procedures, guardrails and escalation, minimum data access, pre launch simulation and evaluation, and human approved updates, to any voice or chat agent you build today.