01OpenAI’s July 22 announcement of Presence is important for a reason that has little to do with another leap in model intelligence. It packages the operating controls around an agent—the policies, permissions, evaluations, approved actions, monitoring, and human handoffs required to make AI dependable inside a real business process.
The headline is not another model
OpenAI describes Presence as a managed enterprise product for high-volume, high-stakes voice and chat workflows. Each deployment begins with a specific job, such as resolving billing issues, supporting an insurance claim, or handling an employee service request. The agent receives only the knowledge and system access needed for that job, while the organization defines what it may do, when approval is required, and when a person must take over.
That bounded design is the meaningful shift. A demonstration asks whether an agent can complete a task once. A production system has to complete thousands of variations while respecting identity, policy, data access, and exception rules. Presence is currently offered through limited general availability as a managed deployment, which also acknowledges that this operating work is not yet a simple self-service configuration exercise.
Governance is becoming part of the product
Presence brings together standard operating procedures, guardrails, approved actions, simulations, evaluation tools, and an improvement process. Those components are not paperwork around the product; they are the product that makes model reasoning usable. An agent that can call tools and change systems needs a stronger control plane than a chatbot that only drafts a response.
The practical architecture separates permission from capability. A model may understand how to issue a credit, update a record, or schedule a service, but the workflow should still determine whether that action is approved for this customer, this amount, and this moment. High-impact steps need explicit policies outside the model, scoped credentials, and a clear escalation path.
The data and analytics work underneath
A governed agent is only as reliable as the operational data behind it. Analysts and data teams have to identify authoritative sources, document metric definitions, expose effective dates, reconcile conflicting records, and preserve lineage. When an agent can take action, stale or ambiguous data stops being a reporting inconvenience and becomes an operational risk.
Business analysts also translate policy into testable decisions. They define what evidence is required before the agent proceeds, what uncertainty should trigger an abstention, which exceptions deserve human judgment, and what a successful handoff contains. That work connects business intent to technical guardrails.
Use a balanced production scorecard
Production measurement should combine outcome, quality, risk, and economics. Useful measures include task completion, first-contact resolution, grounded accuracy, policy compliance, correct tool use, escalation precision, rework, latency, and cost per successful outcome. OpenAI’s agent-evaluation guidance similarly emphasizes traces, graders, datasets, and repeatable evaluation runs rather than relying on a few polished examples.
Containment by itself is dangerous because an agent can avoid escalation while doing the wrong thing. A better scorecard pairs containment with verified resolution quality and customer impact. The deeper signal from Presence is that the enterprise AI race is moving from who has the most impressive demo to who can operate, measure, and improve agents without giving up control.