OpenAI Presence: Enterprise Agents You Rent, Not Own

OpenAI Presence, announced July 22, 2026, is a governed enterprise agent platform that OpenAI describes as "a battle-tested product that helps enterprises deploy trusted AI agents" — agents that answer questions, resolve issues, use company systems, take approved actions, and escalate to people when needed. What makes it worth a strategy conversation isn't the product. It's how OpenAI is selling it.

There is no self-serve tier. There is no public pricing. Deployments are led by OpenAI's Forward Deployed Engineers and select global systems integrators, scoped individually per customer. The biggest AI lab in the world just launched an enterprise agent product that you cannot buy off a pricing page — you engage it, the way you'd engage a consultancy. For every business weighing agent automation, that structure says more than any benchmark.

This post covers what Presence actually ships, which of its numbers to trust (and which are vendor-stated and unverified), why the delivery model quietly validates the senior-led custom-build thesis, and a rent-vs-own decision matrix nobody else has run for this launch — when a rented, governed platform fits, and when owning your agent system outright is the better long-term position.

Presence launched with support for real-time voice and chat agents. OpenAI's outreach materials describe a broader ambition spanning email and other channels, but VentureBeat notes OpenAI has not confirmed email support is live at launch — as of July 22, it's voice and chat.

The deployment philosophy is deliberately narrow. Each Presence deployment starts scoped to a single job — resolving billing issues, supporting insurance claims, handling employee IT requests — and the agent receives only the knowledge and system access that job requires. OpenAI's canonical example is billing-issue resolution end-to-end: understanding the request, verifying the customer, looking up account information, applying policy, and taking an approved action. Crucially, the company — not the agent — sets policy: what the agent can do autonomously, what needs approval, and when a human takes over.

If you've followed our coverage of the build-vs-buy calculus for enterprise agents, the one-job-scoping pattern will look familiar — it's the same discipline any serious custom agent build applies: narrow scope, explicit permissions, human escalation as a designed path rather than a failure mode. Presence productizes that discipline. The question this post keeps returning to is who owns the result.

Presence bundles six components into what OpenAI presents as a complete operational layer for production agents: policies and SOPs, guardrails, approved actions, simulations, evaluation tools, and a Codex-powered improvement process. OpenAI does not name a specific model version as Presence's engine — the announcement says only that the platform runs on OpenAI's models and continues to advance as those models improve.

Two pieces deserve attention. First, pre-deployment simulation: teams can test an agent against common requests, edge cases, and higher-risk scenarios before it reaches real users, with graders checking whether the agent reached the right outcome, followed policy, used tools correctly, and escalated appropriately. Guardrails then intervene mid-interaction when a live conversation moves outside the company's defined boundaries.

Second, the post-launch loop: production sessions, escalations, and quality signals feed back into the system, where Codex — using the Presence plugin — investigates those signals and suggests updates. Teams test the suggestions against the live production version before approving a controlled rollout. That's a genuinely thoughtful improvement architecture. It is also entirely OpenAI's: the tooling, the plugin, and the loop live on the vendor's side of the fence.

That framing is correct, and it's the same conclusion buyers should carry into every agent conversation — reliability engineering, not capability demos, is where enterprise agent projects live or die. It's also how to tell a governed agent from agent-washing: the presence (or absence) of simulation, evaluation, escalation rules, and approved-action boundaries is the fastest scorecard a buyer can run on any vendor's "agent" claim.

The proof point OpenAI leads with is its own support line. Presence runs OpenAI's English-language phone support at 1-888-GPT-0090, handling open-ended requests, verifying callers, using account context, and taking approved actions. OpenAI states that this line now resolves 75% of inbound issues without human assistance, reached within weeks of meeting or exceeding OpenAI's own frontline-human-quality benchmarks — and that the Codex-driven improvement loop reduced human handoffs by 15 percentage points in just 10 days.

Both numbers are vendor-stated. VentureBeat, which covered the launch, is explicit on this point: "Those figures are company-reported and have not been independently verified." The Register and VentureBeat both quote the same 75% figure directly from OpenAI — repetition across outlets is corroboration of the claim's existence, not an audit of its accuracy.

Here is the structural fact that matters most: Presence is in limited general availability, and OpenAI's own announcement states that "Deployments are led by OpenAI Forward Deployed Engineers and select global systems integrators... Presence is not yet available as a self-serve product." Forward Deployed Engineer is a title The Register notes is borrowed from Palantir — a role where engineers embed with customer operations for months at a time.

The delivery arm behind this is the OpenAI Deployment Company, a consulting entity OpenAI launched in May 2026 — about two months before Presence — reinforced by its May 2026 acquisition of the consultancy Tomoro and backed by Bain & Company. On pricing, an OpenAI spokesperson told The Register: "During this limited GA phase, deployments are scoped individually based on each customer's use case and implementation needs. Broader pricing details to come as availability expands." VentureBeat reported it asked OpenAI about pricing twice and had not received an answer at publication.

Read that as a business model, not a product footnote. OpenAI could have shipped Presence as an API with a pricing page. It chose embedded engineers, hand-scoped engagements, and per-customer commercial terms. That is a public admission — from the lab with the most capable models and the most self-serve distribution on the planet — that enterprise agents are an implementation business. The model is necessary; the senior humans wiring it into your systems, policies, and edge cases are what make it work. It's the same thesis we've argued in our build-vs-buy decision framework — Presence is just the largest company yet to price it that way.

Timing adds one more caveat for buyers doing diligence: Presence launched one day after a disclosed OpenAI/Hugging Face security incident involving frontier models under internal evaluation — context worth reading via our coverage of the containment incident when you assess sandboxing and tool-permission risk.

Presence didn't arrive in a vacuum. About a week earlier, on July 15, Anthropic launched Ode — a consulting organization built around forward-deployed engineers helping companies integrate Claude into complex workflows. VentureBeat draws the distinction: Ode centers on deployment services, while Presence combines implementation services with a defined operational layer — policies, simulations, evaluations, approvals, and production updates — as a named, branded product. Two labs, one week, the same bet: the money in enterprise AI is in implementation, not model access.

The rented-platform field is crowded too. A well-funded rival platform, Sierra — co-founded by Bret Taylor, who also chairs OpenAI's board, a piece of industry trivia rather than anything sourced as relevant to Presence's design — reports strong enterprise traction and launched with a wider channel set than Presence's voice-and-chat scope. Salesforce's Agentforce ships inside Service Cloud using CRM data, and Decagon targets high-volume structured support automation for Zendesk, Salesforce, and Kustomer shops. Against that field, Presence's launch scope is notably narrow: two channels, no self-serve tier, no published compliance-certification list, and no pricing.

The named customers reinforce how early this is. BBVA is exploring AI-powered voice support for everyday banking needs in Mexico. SoftBank is testing natural Japanese-language customer conversations. IAG, the Australian insurer, is exploring timely support during high-demand events such as severe weather. Exploring and testing — those are the verbs in every single case. None of the three is described as being in production.

Even the most enthusiastic launch quotes stay carefully inside "explore." That's not a criticism — it's exactly the hedging you'd expect from regulated enterprises three days into a platform's existence. But it means the only production Presence deployment anyone can point to is OpenAI's own phone line, graded by OpenAI, reported by OpenAI.

Mainstream coverage frames Presence against Sierra, Decagon, and Agentforce — rented platform versus rented platform. The comparison enterprise buyers actually face has a third column: building an agent system you own. We've made that case before in branded SaaS vs a custom-built system; here's how the three options compare on the dimensions that decide total cost and control over a multi-year horizon.

The pattern in the matrix: Presence wins on day-one governance maturity and loses on everything that compounds — portability, improvement-loop ownership, cost transparency. That trade is rational for a Fortune-500 support organization that wants a vendor to hold the risk. It's a poor trade for a mid-market company whose support workflows, CRM logic, and customer data are the actual moat — which is why our AI transformation engagements start from the owned column and rent only what genuinely isn't differentiating.

The analyst counter-signal is worth holding next to the launch enthusiasm. Gartner predicted that by 2027, half of organizations that were planning to shift customer service to AI will abandon those plans — a standing forecast from a June 2025 press release that The Register resurfaced in its Presence coverage, not fresh July 2026 research. That is a separate finding from Gartner's agentic-AI project-cancellation forecast, which we've covered on its own — two distinct predictions about two different things; don't let anyone blend them into one scary number.

Our forward read: Presence's launch shape — FDE-led, per-customer pricing, one job at a time — is likely to define how enterprise agents get sold through 2027, because it matches how they actually succeed. Expect a self-serve tier eventually, but expect the governed, implementation-heavy tier to remain the revenue center; once Bain-backed consulting economics are in the model, they rarely leave it. For most businesses, the practical hedge is to own the layer closest to your customers — the CRM logic, the escalation rules, the data pipelines — regardless of which agent engine sits behind it. That's the architecture our CRM automation practice builds: agent-ready workflows you own, with the model layer kept swappable on purpose.