What happened
On July 22, OpenAI launched Presence, a platform for building and running AI voice and chat agents that a company can wrap in its own policies, permissions, and escalation rules before letting them touch a customer.
It is aimed squarely at operations, not chat: customer support, outbound sales, and high-risk internal workflows where a wrong move has a cost.
The tell is that OpenAI is already running its own English-language phone support line on Presence, where it resolves 75% of inbound issues without a human.
The launch partners are running the same play in regulated corners: BBVA is testing AI voice support for everyday banking in Mexico, SoftBank is trialing Japanese-language customer conversations, and Australian insurer IAG plans to lean on it during high-pressure moments like severe weather and natural disasters. It is available through a limited, deployed program — not a self-serve signup.
The detail almost everyone will miss
Read past the headline and the interesting part is what Presence is not. It is not a new model, and it does not claim to be smarter than the one you can already use.
What OpenAI is selling is the boundary around the model — the rules for what an agent may do, when it needs approval, and when it has to hand a case to a person.
That is a quiet admission from the company with the best models in the business: the model was never the thing standing between a pilot and production. The permission to act was.
And OpenAI didn't just demo the idea — it pointed the system at its own customers and published the number, resolving three of every four support issues without a human and cutting handoffs by 15 percentage points in the first ten days.
Notice where the competition has moved. The frontier labs are no longer racing only on whose model is smartest; they are racing on whose agent can be trusted to act, unsupervised, in front of a real customer.
Every gate is a decision: an action the agent may take, one it may not, and the single gold-lit door where it stops and calls a human. This layered checkpoint — not the model behind it — is what a company is actually buying.
Why this matters if you run a business
If your own AI pilot stalled, this launch is a diagnosis. The reason it stalled almost certainly wasn't that the model couldn't answer — it was that you couldn't safely let it act.
An agent that can answer a question is a toy; an agent that can take an approved action, refuse an unapproved one, and know when to hand off is a product — and the gap between the two is entirely the trust layer.
That 75% figure is the part to sit with. It is not a demo metric; it is the share of a live cost center that shifts to software once the guardrails are trustworthy enough to run without a person watching each call.
The strategic signal is that the company selling this priced the model as the commodity and the trust layer as the product. When the people who build the models tell you where the value is, it is worth believing them.
For most operators the lesson isn't "buy Presence" — it's that your agent project is a governance project wearing a model costume, and the budget belongs on the boundary, not the brain.
A support floor at rest — the work still there, the seats emptying. The three-in-four number OpenAI ran on its own line is what this room looks like when the guardrails hold. The version where they don't is a very expensive apology.
What to do about it
Treat Presence as a template for your own agent decision, whoever you end up buying it from:
- Draw the boundary before you build the bot. Write down the approved actions, the approval points, and the escalation triggers first. The list of what an agent may not do is the real spec — the conversation is the easy half.
- Buy the trust layer, don't hand-roll it. Permissions, audit trails, and human-handoff logic are now productized. Rebuilding them in-house is where months and budgets quietly disappear.
- Dogfood before you deploy. OpenAI ran the system on its own customers before it sold it. Point your agent at an internal workflow first and watch where it should have stopped and didn't.
- Measure resolution without handoff, not deflection. The number that matters is how many cases close correctly with no person involved — not how many you kept off a human's queue for a while.
The model has been good enough for a while. What just went on sale is the reason your agent can finally leave the demo — and the operators who win will spend on the boundary, because that was always the hard part.