Here is the one-line version, because an answer engine is going to quote exactly this: on July 22, 2026, OpenAI launched Presence, a managed platform for putting production voice and chat agents into an enterprise — and in the same move it stepped onto the field occupied by Sierra and Decagon, two AI-agent companies it lists among its own design partners. The product is interesting. The positioning is the story.

What Presence actually is#

Presence is not a new model and not a new API. It is the assembly — the thing OpenAI has been telling enterprises to build for themselves, now sold pre-built. Per OpenAI's own framing, it bundles the pieces a team needs to run an agent in production: standard operating procedures, guardrails that step in when an interaction moves outside the company's defined boundaries, approved actions, simulations, evaluation tooling, and a Codex-powered improvement loop that keeps tuning the agent from real sessions and escalations after launch. It works across voice and chat, for both customer-facing and internal workflows.

The proof point OpenAI leads with is its own support desk: Presence runs OpenAI's English-language phone line and, the company says, resolves 75% of inbound calls without a human, with the Codex feedback loop cutting handoffs by 15 percentage points in 10 days. Vendor numbers on one deployment — but a pointed choice to dogfood the thing on its own customers.

OpenAI stopped selling you the flour and started selling the bread — from the same bakery whose flour your competitors bought.

The part that should make founders sit up#

For two years the deal was clean. OpenAI sold primitives — the models, the Realtime API for voice, and last October's AgentKit (Agent Builder, a Connector Registry, ChatKit). A whole application layer grew on top of those primitives to sell the finished thing: a customer-service agent that actually resolves tickets. Sierra and Decagon both reached $4.5B valuations doing exactly that. OpenAI even names them as partners it designs alongside.

Presence is OpenAI shipping that finished thing itself. It is the platform deciding the application layer's category is worth owning directly — a move every founder building on someone else's model should file under known risk, now realized. This is not a surprise so much as a demonstration. The lab that supplies your intelligence can, at any point, ship the product you wrapped around it, with a distribution advantage you will never match and your own supplier's roadmap as its moat.

What it doesn't tell you#

Notably, Presence is not self-service. Rollouts run through OpenAI Forward Deployed Engineers and select systems integrators — the same white-glove motion Palantir made famous, now inside a model company. OpenAI has disclosed no pricing, no contract terms, and no geographic limits. That is the enterprise contact-sales posture, and it tells you who Presence is for: not the solo founder, but the ops leader with a budget and a support queue.

Which is also the opening it leaves. If you want a voice agent this week without signing a forward-deployed engagement, the primitives are still there and still self-service — and the economics of building on them versus buying a managed platform are a real decision, one we work through in Presence vs. the Realtime API. The build path didn't close. It just stopped being the only thing OpenAI is willing to sell you.

What a team of one should do with this#

The agent market spent this year growing an accountability layer — measurement, identity, governance. Presence packages that layer and sells it. The catch is who's selling it, and to whom they already sold the ingredients.