Why Most AI Adoption Projects Start at the Wrong Layer

Direct Answer

Most AI adoption projects fail because they start at the tool layer instead of the coordination layer — a team buys a model, a copilot, or an agent and points it at a workflow nobody has actually redesigned. McKinsey found workflow redesign is the strongest predictor of EBIT impact from AI use, and MIT’s NANDA lab found 95% of enterprise generative AI pilots produce no measurable return. The model was rarely the reason.

Key Takeaways

• Workflow redesign, not model selection, is the strongest predictor of AI’s business impact. McKinsey found only 21% of adopters had fundamentally redesigned a workflow, and that group captures most of the measured gains.
• MIT NANDA’s 2025 study of enterprise generative AI found 95% of pilots deliver no measurable P&L return, and traced the gap to organizational learning and workflow integration, not model quality.
• The “wrong layer” is the tool layer: the model, chatbot, or single-step agent. The right layer is the coordination layer: who hands off to whom, in what order, and where a human has to stop the process.
• Adding an agent to a broken handoff usually makes the handoff faster, not correct. Speed does not fix a step that was never supposed to exist.
• Deciding what stays human comes before deciding what gets automated, not after.

What “The Wrong Layer” Actually Means

Every workflow has at least three layers stacked on top of each other. There’s a strategy layer — what the business is trying to get done. There’s a coordination layer — the actual sequence of handoffs, approvals, and decisions that make the work move from one person to the next. And there’s a tool layer — the software, spreadsheet, inbox, or model that executes each step.

Most AI adoption projects start at the bottom of that stack. Someone picks a tool — a chatbot for support tickets, a copilot for drafting emails, an agent that reads invoices — and installs it at one point in an existing process. The process itself, the coordination layer, never gets touched. Nobody asks why the invoice moves through four inboxes before it gets approved, or why a support ticket needs three separate sign-offs before a refund goes out. The AI just does one of those steps faster.

That’s the wrong layer. It’s wrong not because the tool is bad, but because the tool was asked to speed up a route that was never designed on purpose. Most of these routes are what I’d call 1995 workflows: data moving between email, PDFs, and spreadsheets because that’s what was available thirty years ago, not because anyone decided that was the right way to move a claim, a purchase order, or a booking through an organization. Layer AI on top of that, and you get a faster version of a route nobody would design today.

What the McKinsey and MIT Numbers Actually Show

Two separate research efforts converged on the same finding from different directions in 2025, and the pattern is worth laying out plainly rather than summarized.

Read those four rows together and the story isn’t “AI doesn’t work.” Adoption is nearly universal — 88% of organizations already use it somewhere. The gap is between adoption and transformation: only 6% see enterprise-wide financial impact large enough to matter, and the group that does overwhelmingly overlaps with the 21% that redesigned a workflow rather than bolting a tool onto an existing one.

MIT NANDA’s researchers, after reviewing more than 300 public AI deployments and interviewing dozens of enterprise leaders, put it directly: the core barrier to scaling wasn’t infrastructure, regulation, or talent. It was learning — systems and teams that never adapted the surrounding process once the tool was live. That’s a coordination-layer failure wearing a technology costume.

What This Is Not

It’s worth naming what the data does not support, because the easy conclusions are the wrong ones.

It is not primarily a data-readiness problem, even though bad data is real and makes everything harder. Clean data flowing into the wrong layer still produces a faster version of the wrong process.

It is not a model-quality problem. The frontier models used in 2025 and 2026 pilots were, by any reasonable measure, capable enough for the tasks companies assigned them. The failure rate didn’t track model releases.

It is not primarily a training or change-management problem in the usual corporate sense — the kind fixed with an onboarding deck and a Slack channel. Training people to use a new tool inside an unchanged process just makes them faster at the same broken handoff.

It is a layering problem. Teams treated AI as a faster version of an existing step instead of a reason to ask whether that step should exist in its current form at all.

Where Should AI Adoption Actually Start?

Start by mapping the coordination layer before selecting a tool. That means writing down, honestly, what the current route actually is: who touches the document or ticket first, what triggers the next handoff, who has to approve it, where it sits when nobody’s looking at it, and what happens when something goes wrong.

Then, for every handoff in that map, ask one question: does this step exist because of a real constraint — liability, a licensed act, a regulatory requirement, a trust relationship with a customer — or because that’s simply how the work has always moved between two people? I’ve written through this exercise for invoice processing and for insurance claims, and the pattern repeats: most steps in a 1995 workflow exist because of missing infrastructure at the time, not because the step is load-bearing today.

Only once that map exists does it make sense to decide where AI executes a step directly and where a human has to remain the checkpoint. Picking the tool first inverts the order, and inversion is exactly what the McKinsey and MIT numbers are describing when they say workflow redesign is the variable that predicts impact and its absence is the variable that predicts failure. The handoff itself is usually the thing worth interrogating before anyone opens a vendor comparison spreadsheet.

What Stays Human When You Redesign the Coordination Layer

Redesigning the coordination layer doesn’t mean automating every step in it. Some steps stay human on purpose, and naming them explicitly is part of the redesign, not an afterthought bolted on after launch.

Liability decisions stay human — the moment where someone is accountable if the call is wrong. Licensed acts stay human, because a credential or a signature carries legal weight a model can’t hold. Irreversible customer-facing actions, like a refund that can’t be clawed back or a contract that’s been sent, stay human until the system has enough track record to earn trust at that step specifically, not generally. And judgment calls that depend on context outside the data the system can see — a long-standing customer relationship, a one-off exception — stay human because that’s precisely the kind of situation a coordination layer redesign should route to a person, not around one.

I’ve written more specifically about where that human checkpoint actually belongs in a workflow, and the short version is: the checkpoint belongs at the coordination layer, placed deliberately, not wherever the tool happened to stop.

Summary

The reason most AI adoption projects don’t show up in the numbers isn’t the model. It’s that the model got installed at the tool layer while the coordination layer — the actual sequence of handoffs a business runs on — stayed exactly as it was in 1995. McKinsey’s research ties workflow redesign directly to EBIT impact; MIT NANDA’s research ties its absence directly to pilots that never scale. Adoption starts in the right place when the coordination layer gets mapped and redesigned first, and the tool gets chosen second.