OpenAI Presence: when the model vendor sells the deployment too

OpenAI introduced Presence on July 22, and the interesting part is not the technology. It is the sales channel. Presence is a managed platform for deploying what OpenAI calls trusted voice and chat agents into production workflows: customer support, outbound sales, high-risk internal processes. It bundles model reasoning with a company's policies, guardrails, escalation rules, simulations, evaluations, and Codex-powered improvement loops. OpenAI says it already handles about 75% of inbound calls on its own English-language support. And you cannot sign up for it. It is limited GA, available only through an account team.

What shipped, and what it is not

Presence is not a new API. It is the integration and governance layer, the part that system integrators and independent consultants have been delivering on top of the API for two years, now packaged and priced by the lab. The Register called it a consulting path outright, and reporting the next day described it as a deployed product with boots-on-the-ground pricing rather than self-serve software.

That puts OpenAI on the same field as Microsoft Frontier, which reportedly has 6,000 engineers embedded at customers, as Palantir, and as OpenAI's own Deployment Company, built from the Northslope acquisition. The pattern is not one company's strategy. Anthropic and Blackstone have made the same bet, reported by TechCrunch on July 15: the value is moving from training models to implementing them.

Why the labs are doing this

Because the pilots are failing, and not because of the models. The MIT study cited in the Presence coverage found that 95% of enterprise generative AI pilots produced no measurable impact, and that the failures were almost never about model quality. They were about integration, permissions, and change management.

The adoption data from the same week says the same thing from the other side. Anthropic's 2026 State of AI Agents report has 80% of companies seeing measurable return and 81% scaling multi-step agent workflows, while Gartner expects more than 40% of agentic AI projects to be cancelled by 2027 over cost, unclear ROI, and weak governance. Both can be true. The companies with attribution and governance see returns; the rest generate code (AI reportedly writes about 61% of enterprise codebases now) without visibility or control.

The model was never the bottleneck. The deployment was, and now the vendor wants to own that too.

What Presence does well

Give it credit for the shape. A policy layer, escalation rules, simulations before go-live, an evaluation loop, and an improvement cycle fed by the coding agent: that is the checklist I would hand any team building a support agent, and most teams skip half of it. If you are a large OpenAI customer with an account team and a support-heavy workload, this is a reasonable way to skip building it yourself.

What independent teams still own

• Vendor neutrality. Presence lives inside the OpenAI stack. A router that sends routine traffic to open-weight models and hard cases to a frontier model, the pattern in frontier-model-router-2026, is something a single-vendor platform will not build for you.
• Hybrid and on-prem hosting. Regulated customers and data-residency constraints are not solved by a managed cloud agent, however good its guardrails are.
• Measurable ROI per task. The lab measures its own deflection rate. You need per-task quality and cost numbers for your workload, tied to the process you changed.
• Portable guardrails. Output validation and policy checks that live in your code, not in a vendor console. See guardrails-output-validation.
• Speed and no lock-in. A small team can ship an agent on a client's infrastructure in weeks and hand over the keys. 6,000 embedded engineers cannot be un-embedded cheaply.

The honest gap

Everything about Presence is vendor-described. The 75% deflection figure is OpenAI's own support line, a workload it controls end to end, and there is no public pricing, no independent evaluation, and no customer reference yet. Whether it beats a well-built in-house agent on a real enterprise workload is unknown. What is not unknown is the direction: every major lab now sells implementation, which is the clearest signal I have seen that implementation is where the money is.