OpenAI Presence Proves AI Agents Still Need Hands

OpenAI Presence launched on July 22, 2026, and the most important thing about it isn’t the demo. It’s the fine print: every deployment is led by human engineers. OpenAI — the company with the most capable models on the planet — is not selling you an AI agent you switch on. It’s selling you an AI agent with a team of Forward Deployed Engineers attached, delivered through a six-stage rollout and priced per customer. That detail is the whole story, and it says something uncomfortable about where enterprise AI actually is in 2026: the model was never the bottleneck. The hands are.

What OpenAI Presence actually is

Presence is a managed service for putting AI voice and chat agents into real customer and internal workflows — billing disputes, insurance claims, IT requests — with guardrails, escalation rules, and a Codex-powered process that studies production sessions and proposes improvements before they roll out. On OpenAI’s own announcement, the company says its English-language phone support line hit a 75% resolution rate without human help within weeks, and cut human handoffs by 15 percentage points in ten days. Impressive numbers, and I believe them.

But look at how it’s sold. As AI News reported, Presence is in limited general availability — not self-serve — with OpenAI’s Forward Deployed Engineers and selected integrators leading every deployment. Pricing is unpublished and set per customer. The three named customers — BBVA, SoftBank, and IAG — are design partners, not proof of scale. This is a consulting engagement wrapped around a model, not a product you buy and turn on.

Read the delivery model, not the demo

That delivery model is the tell. Handing customers a capable agent isn’t the hard part anymore; getting it to run safely inside a real business is. When the constraint shifts from what the model can do to how many engineers you can put in the room, you’ve stopped selling intelligence and started selling deployment.

The industry’s own analysts have been saying this out loud. Gartner predicts that over 40% of agentic AI projects will be canceled by the end of 2027 — not because the models got worse, but because of unclear value, escalating costs, and inadequate controls. Presence is OpenAI’s answer to that failure rate: throw expert humans at every rollout so it doesn’t become another canceled project. It works. It just doesn’t scale to everyone, and it isn’t cheap.

The bottleneck is the hands

Here’s what I mean by “the hands.” A model can reason about your business all day. What’s scarce is the safe, accountable path from the AI decided something to the change is live in your systems, on the record, and reversible. That path is the hands — and it’s what Presence spends its six stages and its engineers building for each customer, by hand, every time.

Scoping, security review, staged rollout, escalation to a person when the agent is unsure — none of that is model work. It’s accountability work. It’s the scaffolding that lets a business trust an autonomous system to touch real workflows. The frontier lab just confirmed, with its product roadmap, that this scaffolding is the expensive, unsolved part.

What this means if you don’t have OpenAI’s engineers

Most businesses will never get a Forward Deployed Engineer. But the underlying need is identical whether you’re BBVA or a five-person agency: give an AI safe hands into your real systems, and keep a record of everything it does.

For the system most businesses actually want to change — the CRM — that gap is exactly what we built Sentinel to close. The pitch has never been a smarter model. It’s giving the AI you already use the hands to develop against your CRM, safely, without a consulting engagement attached to it.

How Sentinel gives your AI hands — without a forward-deployed engineer

Presence supplies accountability with people: engineers who scope, review, and watch. Sentinel supplies it with a safety layer instead. Every change your AI makes is logged, and a snapshot is taken before every deploy; Salesforce changes go sandbox-first with tests required; only one write key is live at a time so two sessions never collide.

The mechanics are deliberately un-magical. Your AI connects over MCP to a dedicated per-client server, not a laptop, so the audit log, the snapshots, and the key custody all live in one place that’s always on. And that server is auto-provisioned the moment you buy — VM, static IP, DNS, and SSL, no engineer required. Where Presence is limited-availability and engineer-led, this is self-serve.

To be clear about the difference: Presence deploys customer-support agents; Sentinel is for building and changing your CRM. Different jobs. But the lesson about hands is the same one, and it’s the whole reason the safety layer exists.

Freedom plus visibility, not a human on every step

There are really only two ways to make an autonomous system safe to run in production. Put a person in the loop for every action — accurate, and exactly why Presence is engineer-led and priced per customer. Or make every action visible and every deploy recoverable, so you can move fast and still prove what happened and undo it. Sentinel bets on the second. It logs everything and blocks nothing, which is what lets several team members’ AI sessions develop against the same org without a coordinator standing over them.

That bet also shows up in the price, which — unlike Presence — is published: $2,500 one-time onboarding on your first Sentinel, then $500/month per Sentinel. You can own more than one. No custom quote, no engagement fee, no engineers billed by the hour.

The takeaway

The Presence launch isn’t a threat to read nervously — it’s a signal to read carefully. The most advanced AI company in the world just told you, through its own product design, that deploying agents into real work still takes hands, and that those hands are the expensive part. The only real question left for your business is whose hands, and how accountable.

If the thing you want changed is your CRM, you don’t need OpenAI’s engineers on site. You need your own AI to have safe, logged, reversible hands into your org. Start your Sentinel and give it those hands.