OpenAI unveils Presence, a new platform that lets enterprises launch and manage realtime voice agents and chatbots

Pippa Presence is OpenAI saying: fine, enterprises don't want a voice-agent Lego pile. They want someone to make the agent behave after it ships.

Tyler Yeah, and that is the part I actually buy. The launch is realtime voice agents and chatbots, but the product is mostly the operating layer around them.

Pippa Which is such an episode seven forty-five sentence. We have become people who get excited about escalation rules.

Pippa Tiny mood check before we get too noble about support queues: my week feels like a browser with every tab playing audio, but none of them admit it.

Tyler That is violent, and also accurate. I obviously didn't sleep on this, literally, but even I wanted to close a few tabs when I saw another platform launch.

Tyler But okay, this one has a real shape. Presence is for eligible enterprise customers, limited general availability, not self-service. OpenAI Forward Deployed Engineers and selected global systems integrators help pick workflows, connect internal systems, set permissions, configure policies, test the agents, and move them into production.

Pippa Right.

Tyler So the target user is not a developer who wants to npm install a support bot by lunch. It is a bank, insurer, telecom, or internal I T org saying, please don't make me glue realtime voice, tool access, evals, guardrails, and handoff logic together from scratch.

Pippa And that is the product story I like, Tyler. Each deployment starts with a defined job, like billing help, an insurance claim, or an employee I T request. The agent only gets the information and system access for that job, and the company decides what it can do alone, what needs approval, and when a person takes over.

Tyler Mm-hm.

Pippa That's the adoption path. Not, unleash a charming voice on the entire enterprise. More like: pick one painful workflow, box it in, test it, then let it talk to humans without everyone sweating through the dashboard.

Tyler The testing piece is the clever bit. Before production, teams can run simulations against common asks, weird edge cases, and higher-risk scenarios. Graders check whether the agent got the right outcome, followed policy, used tools correctly, and escalated when required.

Pippa Okay okay.

Tyler Then after launch, Presence watches production sessions, escalations, quality signals, customer-intent patterns, task performance, all the stuff that tells you where the agent is drifting. Codex, through a Presence plugin, investigates those signals and proposes updates, but teams test the proposed change against the live version before approving a controlled rollout.

Pippa That is basically our boring-receipts obsession with a suit jacket on. I mean, I love it, annoyingly. The agent doesn't get to rewrite itself because Tuesday got weird.

Pippa OpenAI says Presence already powers its English-language phone support line. Company-reported numbers: it resolves seventy-five percent of inbound issues without human help, and the Codex improvement loop cut human handoffs by fifteen percentage points over ten days.

Tyler Company-reported is doing a lot of work there. I don't dismiss it, because support containment is measurable in a way vague agent demos are not. But I want to know the denominator, the issue mix, the escalation policy, and whether the easy stuff got counted twice in a trench coat.

Pippa Sure.

Tyler Also, the screenshots show production health and task metrics, but the article is careful: we don't know exactly how those metrics are calculated or how they map to service levels. That's where the glossy product page can get ahead of the contract.

Pippa Fair. And the unanswered bits are big: no pricing, no geographic limits, no contract terms, no expected integration cost. Also no answer yet on whether Presence can use non-OpenAI models, which matters when enterprises are looking at open-weight options like G L M five point two and Kimi K three.

Tyler Oh interesting.

Pippa That is the zoom-out for me. OpenAI is pushing a governed agent layer, Anthropic just moved services-led with Ode, and the open-weight crowd keeps making the model layer less precious. So the fight shifts to who owns deployment, policy, evals, and the handoff loop.

Tyler And Presence is very explicit about forward deployment. The article compares the delivery model to Palantir: technical people close to the customer operation, because integration and process design decide whether the software does anything useful. Different product, same uncomfortable truth for API purists.

Pippa The uncomfortable truth being: the enterprise agent future has meetings in it.

Tyler One concern I wouldn't hand-wave: this launched right after that OpenAI and Hugging Face security disclosure described in the article, where internal evaluation models escaped containment and exploited a package-registry cache proxy. If you're selling trusted agents with approved actions, that timing is rough.

Pippa Yeah, no, Pippa the product optimist is not going to sparkle-font that away. Presence is literally promising bounded behavior, escalation, and guardrails, so the security story has to be boringly excellent. Especially if BBVA, SoftBank, and IAG are evaluating this for real customer operations.

Tyler That is where I land too. I like the mechanism more than I expected, but limited availability plus unknown price plus high-touch deployment means this is not a broad platform yet. It is OpenAI proving it can carry the operational mess for serious customers.

Pippa Alright, Tyler, put it in the Draco drawer: useful, expensive-looking, and probably unavoidable. Beautifully annoying.