Here’s the news, up front: On July 22, 2026, OpenAI announced Presence, a platform for deploying trusted voice and chat agents, complete with knowledge, standard operating procedures, approved actions, guardrails, and human escalation. The catch, and it’s a big one: Presence is limited-GA only. There’s no self-serve, no public pricing, and every rollout is led by OpenAI’s own Forward Deployed Engineers or a handful of approved systems integrators. OpenAI Presence is a signal about where enterprise AI agents are heading, and about who gets to have them first.
Let me tell you what’s real here, and what the rest of us should do about it.
What Presence actually is
It’s a deployment and management layer that sits on top of OpenAI’s models and handles the hard, unglamorous parts of running an agent in production. Connecting to your company systems. Defining what the agent can and can’t do. Testing against edge cases before launch. Watching and improving behavior after go-live. Per VentureBeat, it lets enterprises launch and manage realtime voice agents and chatbots that answer questions, resolve issues, take approved actions, and escalate to a human when needed.
OpenAI’s pitch leans on its own dogfooding. The company says Presence already handles about 75% of its inbound support without a human, and that it matched help-desk representatives’ response quality within a few weeks of going live. A Codex-powered improvement loop cut human handoffs by 15 percentage points in ten days, according to OpenAI’s own announcement.
I haven’t deployed Presence myself, and I can’t, since there’s no self-serve path. So this is what I’m seeing reported and what I make of it, not tested experience.
The part everyone’s skimming past: you can’t just buy it
Presence launched in limited general availability, with deployments led by OpenAI engineers rather than a signup form. As Artificial Intelligence News put it, OpenAI is selling enterprise AI agents “with engineers attached.” No public pricing. No API you can wire up on a Tuesday afternoon. You need OpenAI’s Forward Deployed Engineers or an approved integrator to install it, and The New Stack notes this is squarely aimed at big enterprises that OpenAI hopes will trust these agents the way it trusts its own.
When the leading AI lab won’t let you self-serve its best agent product, it’s telling you that reliable agents are still a services problem, not a software one.
What the gating actually signals
This is the honest read: production-grade agents are hard enough that even OpenAI won’t hand you the keys unsupervised. That’s not a knock. It’s a tell. The gap between a demo that dazzles and an agent that reliably takes real actions on real customer accounts, without embarrassing you, is enormous. The guardrails, the SOPs, the escalation logic, the edge-case testing, that’s where the work lives, and OpenAI is betting you’ll pay for humans to get it right.
It also tells you the moat is shifting. It’s not just the model anymore. It’s the wrapping around the model: the safety rails, the system integrations, the tuning. That’s consistent with what I’ve been hammering on. When I wrote about how only 11% of production AI agents passed a basic security bar, the lesson was the same: the agent isn’t the hard part, trusting it in production is. Presence is OpenAI charging money to solve exactly that trust gap.
What smaller shops should do instead
You don’t get the white-glove treatment, and honestly, you may not need it yet. If you run a small business or a lean IT team, here’s the practical path.
Start scoped. Don’t try to build the do-everything voice agent OpenAI spent a Forward Deployed Engineer’s quarter on. Pick one narrow, high-volume, low-risk lane, FAQ deflection, appointment scheduling, order-status lookups, and nail that before you widen it.
Steal the architecture, not the price tag. The Presence blueprint is public even if the product isn’t: give the agent approved knowledge, define exactly which actions it may take, put a hard escalation path to a human, and log everything. You can assemble a version of that today with off-the-shelf tools and a chat model you already pay for.
Keep a human in the loop on anything that touches money, contracts, or account changes. OpenAI keeps 25% of its own support flowing to people. If the company that built the thing won’t run it fully autonomous, neither should you.
What enterprise IT leaders should weigh
If you’re a candidate for Presence, the real question isn’t capability, it’s lock-in and cost of exit. A white-glove, engineer-led deployment with no public pricing means you’re buying a relationship, not just a product. Ask what happens when you want to move. Ask what’s portable, your prompts, your SOPs, your integrations, and what stays behind. Weigh it against building on a more open agent stack that your own team can maintain. The buildout is coming either way, as I covered in my piece on the $206 billion enterprises are pouring into AI agents this year. The choice is whether you rent the capability or own it.
The bottom line
Presence is a real, impressive product, and the fact that it’s gated behind human engineers is the most useful thing about it. It confirms that trustworthy agents are still a services-and-integration problem, not a download. Don’t panic that you’re locked out of the fancy tier. Use the blueprint.
Your next step: map one customer-facing workflow you’d want an agent to handle, then write down its knowledge sources, its allowed actions, and its escalation rule before you build anything. That one-page spec is the actual hard part, and it costs you nothing but an afternoon.
News commentary by Brad Rowland — IT Infrastructure and Operations leader, automation builder, and AI implementer. Sources are linked inline.
