What OpenAI actually shipped
On 22 July 2026 OpenAI released Presence, a platform for building, deploying and managing real-time voice and chat agents inside a company. It is not a new model. It is the scaffolding that sits around a model once you try to put it in front of customers or staff: written company policies and standard procedures, permission controls that limit which data and systems an agent can reach, and escalation rules that hand a conversation to a person when it moves outside the agent's remit.
Before anything goes live, Presence runs simulations. Graders score whether the agent reaches the right outcome, follows policy, uses its tools correctly and escalates when it should, tested against common requests, edge cases and higher-risk scenarios. After launch, an OpenAI Codex loop reviews real interactions and proposes changes to the agent's behaviour - but a staff member has to test and approve each one before it ships. Presence is offered as limited general availability, through OpenAI account teams and Forward Deployed Engineers, not as a self-serve product.
The model was never the thing stopping you
For two years the public argument about agents has been about capability - is the model good enough to trust with a customer. Presence is a quiet admission that this was the wrong question. The models were good enough some time ago; what stopped owners deploying them was the absence of a boundary. Who decides what the agent is allowed to touch, what happens when it hits a case it was not built for, and how you prove after the fact that it behaved. Those are governance questions, not intelligence questions.
Presence turns exactly that boundary into a product. The value it sells is not a smarter voice - it is the permission model, the escalation path and the simulation harness that let you sign off on a deployment. Read plainly, the message to owners is that the hard, expensive part of an agent was always the scaffolding, and a vendor has now decided that scaffolding is worth charging for.
The 75 percent hides a 25 percent
OpenAI's proof point is its own support line, 1-888-GPT-0090, which it says now resolves 75 percent of inbound issues without a human and met or exceeded its internal benchmarks for frontline human quality within weeks. That number will travel fast, and it will become the bar your own board points at. It is worth reading the other way round: one in four calls still escalates to a person, and those are the hard, angry, unusual or legally loaded ones.
The economics of a support operation do not live in the easy 75 percent. They live in the 25 percent that escalates, where the cost per contact, the liability and the reputational risk concentrate. An agent that clears the routine work can make that residue harder, not easier, because the person who picks up now inherits only the calls the machine could not close. Plan the 25 percent deliberately, or the headline number will flatter a queue that quietly got worse.
This is a dependency, not a tool
Because Presence ships as limited availability through account teams and Forward Deployed Engineers, adopting it is closer to a consulting engagement than installing software. Your policies, your escalation logic and your permission maps get encoded inside one vendor's platform. That is convenient at launch and expensive to leave, because the switching cost is no longer the model - it is the governance layer you spent months tuning.
The defensible move is to treat the policy, the escalation rules and the permission model as your own portable assets, documented outside the platform that runs them. The agent can be swapped. The rulebook that says what it may do, and who it calls when unsure, is the part you cannot afford to rent. Own it in a form you could hand to a different vendor tomorrow.
Decide who owns the agent before it answers
The unglamorous question to settle first is liability. When an agent takes an approved action - moves money, changes an order, promises a refund - the accountability does not sit with the model. It sits with whoever signed off the policy that let it act. In the EU that lands on real regimes: data-access limits under the GDPR, the high-risk duties the AI Act begins applying from 2 August 2026, and, in several member states, a works council that has a say before an agent handles staff-facing tasks.
Presence gives owners better tools to answer these questions than they had last week. It does not answer them for you. Before the agent picks up a single call, name the person who owns its rulebook, the escalation path for the hard quarter of contacts, and the record you would show a regulator. Those three answers, not the model behind the voice, decide whether the deployment is safe.
Read next: The Door Left Open Was the Package Registry | The Attacker Was a Vendor Running an Evaluation
