The frontier lab hired KPMG to build an internal Supply Chain & Fulfillment Orchestration platform — essentially asking the consulting firm to design the kind of AI-native workflow system that KPMG now intends to sell broadly.

That “client-zero” deployment, as KPMG calls it, is the foundation of a new alliance between the two companies announced today, in which KPMG has been named an OpenAI Elite Partner — the highest tier in OpenAI’s partner network. After partnering to design and deploy an AI-native Supply Chain & Fulfillment Orchestration platform, now they are taking that model to market together.

Colleen Kapase, Vice President of Strategic Global Partnerships and Ecosystems at OpenAI, told Fortune that Elite Partner status is reserved for a “limited group of global partners” that have the reach, scale and delivery capabilities to support enterprise AI adoption worldwide.

“We’re beyond experimentation,” Chad Seiler, KPMG’s U.S. industry leader for technology, media and telecommunications, told me in an interview. “This is about large-scale enterprise deployment.”

The core of what KPMG is selling is a bet on how enterprise software changes from here. Right now, most employees interact with work through applications: they log into systems, navigate screens, click through modules. Seiler’s argument is that this is ending.

“When we say headless, we’re really talking about decoupling the experience of work from the underlying systems and screens and modules while keeping those systems in place as a system of record,” he said.

In the model he’s describing, employees stop navigating software and start describing what they want done. AI agents interpret the intent, coordinate across backend systems, and execute — or escalate to a human when judgment is required.

The endpoint, in his telling, is voice. “Over time, you’re going to be talking more than you’re typing,” Seiler said. “Instead of just interacting with your ERP system or your CRM system in a traditional way with clumsy UIs that are limited in what they can do, you’re kind of unleashed and you can have literally conversations with your systems and take actions with it and take actions not only within that system, but connect that to other data sets and other systems all through an intelligent agentic layer.”

A KPMG blog post published on July 20, co-authored by Swami Chandrasekaran and Matteo Colombo, frames the shift more formally: “Tried and true SaaS isn’t going away. Its user interface is evolving. More precisely, a new work surface is emerging.”

The underlying databases and enterprise applications don’t go away. They become infrastructure, invisible to the user, running underneath a layer of agents that handles translation between human intent and machine execution.

A sandwich, squashed

To make the abstraction concrete, the interview touched on a “sandwich” framework from Princeton’s Arvind Narayanan, whose “AI as Normal Technology” research breaks work into three layers: a decide layer on top, an execute layer in the middle, and a deliver layer on the bottom — what he calls the “decide, execute, deliver sandwich.”

Narayanan’s argument is that AI compresses only the execute layer, which was never more than a third of the work to begin with, while the decide and deliver layers — judgment and accountability — resist compression and may expand.

I showed Seiler the framework on a screen-share during our interview. He agreed immediately that it matched what KPMG was seeing in deployments — the image did look like a skinny hamburger patty. But he added a wrinkle: speed itself creates new verification burdens at the deliver layer, and a new kind of overhead that wasn’t in Narayanan’s original thesis — more talking about work rather than doing it.

“The worker spends less time learning the geography of the software and more time focusing on the outcomes they’re trying to achieve,” Seiler said. In his reading, the decide bun doesn’t just hold steady as the execute patty shrinks. It can expand, absorbing the coordination work that used to live in the middle. But someday, headless work will get so good that it could shrink away.

When reached for comment, Narayanan argued that software engineers have always spent a surprising fraction of their time writing specifications and product requirements documents, which fits into the “bun” section of his metaphor. But it’s not only that judgment and accountability structurally resist compression, he argued, although to an extent, they do; it’s also that “AI is rapidly increasing the ambition and complexity of projects, so the *ceiling* of judgment and accountability moves up, even as AI moves the floor up.”

When reached for comment, Narayanan referred Fortune to another blog post where he tackled the SaaS issue, saying it’s slightly distinct from the decision sandwich. On LinkedIn, he warned about “lock-in,” where the AI agent will become “the main queryable repository of all … tacit knowledge, creating dependence and stickiness,” meaning it is effectively a coworker “that you can’t fire without *every* team losing workflows and know-how.”

On timelines, Seiler and Narayanan are more aligned than their respective positions might suggest. Narayanan frames organizational adaptation to AI as a decades-long process — closer to factory electrification than overnight disruption. KPMG offers an implicit hedge in the same direction, cautioning against wholesale reinvention: “The most successful organizations will be deliberate about where they reinvent — and where they do not,” noting that “the same workflows that have been in place for years may continue to be the best fit.”

Meanwhile, Narayanan said he doesn’t think AI is urgent to the extent that frontier labs portray it to be, e.g., superintelligence by 2027, “but nonetheless it is more urgent a shock than most organizations are used to dealing with.” In other words, it takes a long time to reinvent the sandwich.

Why KPMG says humans still matter

The most interesting point Seiler makes isn’t about the technology. It’s about why a consulting firm is well-positioned in a world where the technology is commoditizing.

His answer: decades of client-specific institutional knowledge that no frontier model has. “We know their business models, their people, their culture, their systems, their data, their politics, their silos,” he said, “in an intimate way at scale that some of these frontier models don’t.”

Kapase agreed that KPMG brings “deep enterprise transformation experience,” particularly across highly regulated industries, the public sector and cybersecurity, where governance and implementation expertise are critical. She pointed to public-sector modernization and a product called Daybreak Cyber as key aspects of the partnership, inaddition to KPMG’s client-zero work within OpenAI. OpenAI is committed to broad access across its ecosystem, she added, so KPMG is not receiving exclusive access to unreleased OpenAI capabilities.

Seiler, for his part, described the OpenAI alliance as additive rather than exclusive: KPMG maintains parallel partnerships with other frontier labs, including Anthropic, and doesn’t expect large clients to standardize on a single AI provider.

“We don’t think we’re going to see a lot of cases where we’re going to have one client that’s completely just using one frontier model to run everything,” he said. Some clients, he acknowledged, are already using cheaper alternatives — including open-source models from China — for narrower tasks, as a cost and resilience hedge.

Regarding the rise of open-source models, Kapase said OpenAI’s focus is on “helping customers get greater value from OpenAI.” She noted that GPT‑5.6 delivers more intelligence from every token and stronger performance per dollar, with Sol 54% more token-efficient on agentic coding tasks. KPMG employees on their Advisory and internal teams have been using OpenAI capabilities daily since the firm integrated it into the internal AI tool aIQ Chat in 2023, Kapase noted, with KPMG identifying Codex-related use cases as it builds AI-enabled capabilities for clients.

What distinguishes the OpenAI deal, in Seiler’s telling, is the go-to-market dimension. “It’s one thing to work with the labs, and then it’s another thing to also work with them and go to market with them.” The client-zero deployment is the proof point: if KPMG could build this for OpenAI itself, the pitch to every other enterprise client becomes considerably easier to make.

Narayanan also referred Fortune to a blog post where he tackled the SaaS issue, saying it’s slightly distinct from the decision sandwich. On LinkedIn, he warned about “lock-in,” where the AI agent will become “the main queryable repository of all … tacit knowledge, creating dependence and stickiness,” meaning it is effectively a coworker “that you can’t fire without *every* team losing workflows and know-how.”

KPMG closes on a note that sounds less like a technology announcement and more like a management consulting memo — which is probably the point. Agentic AI adoption, the firm writes, is “a portfolio of business decisions to be made, not a technology migration.” For a firm whose value proposition has always been helping large organizations make hard decisions carefully, that is less a hedge than a positioning statement.