APIs Out, Consultants In: The Enterprise Agent Platform War Has Started

The most telling detail in OpenAI's launch of Presence this week is not a benchmark or a price. It is a distribution decision: you cannot sign up. Deployments run through OpenAI's Forward Deployed Engineers and hand-picked systems integrators, in a limited general-availability program. The company that built the world's most famous self-serve API has concluded that the enterprise agent market cannot be won with an API at all.

The 95 percent problem

The industry's dirty secret — repeated in every earnings call sotto voce — is that the overwhelming majority of enterprise AI pilots fail, by some estimates 95 percent. They fail not because models are too weak but because integration is brutal: permissions, legacy software, escalation paths, compliance sign-off, the thousand small vetoes of enterprise IT. Model capability became commodity; deployment competence did not.

Every major player has now drawn the same conclusion within months of each other. Microsoft embedded some 6,000 engineers in customers under its Frontier Company program. Google is selling Gemini Enterprise as a governed suite, not an endpoint. Meta pitches Business Agent through its ads sales machine. Nvidia and ServiceNow are wiring Project Arc directly into IT workflows. And OpenAI now fields consultants who charge, as one wag put it, boots-on-the-ground prices — while eating its own dog food on a support line that resolves 75 percent of calls autonomously.

Why agents change the lock-in math

An API customer can switch providers in an afternoon; the flash-tier price war exists precisely because model access is fungible. An agent deployment is different in kind. It encodes a company's policies, permission structures, escalation protocols and tribal knowledge into a specific platform's governance layer. It is touched by consultants who trained your staff. Ripping it out is not a model swap; it is a systems migration with compliance implications.

That is why the labs are racing to be first through the door of every Fortune 500 IT department, even at consulting-margin economics they once disdained. The prize is not this year's deployment revenue. It is becoming the default operational substrate — the place where the agents live — before rivals do. Switching costs are the moat that model quality never was.

Watch the integrators, and watch Asia

Two second-order effects deserve attention. The global systems integrators — Accenture, Deloitte, Infosys, TCS — are being deputized as distribution for the labs, even as agentic coding hollows out their traditional billable-hours model from below. They are simultaneously channel and prey, which is not a stable arrangement; Indian IT's $7 billion acquisition spree reads as preparation for exactly this squeeze.

And the enterprise agent war will not stay Western. China's platforms consolidated their consumer agents months ago, and Alibaba, Tencent and ByteDance are now packaging enterprise agents for a domestic market where US labs cannot follow. SoftBank's presence on OpenAI's launch-customer list is the tell: in Asia's boardrooms as everywhere else, the question has moved from "which model?" to "whose platform runs our operations?" — and that answer, once given, will not be revisited for a decade.