The 6% Problem: Why AI Isn't Delivering ROI Despite Accelerating Investment
- David Golub
- 2 days ago
- 6 min read
McKinsey's latest state of AI report is subtitled "On the road to ROI." Its headline finding is that the organizations getting there are “fundamentally redesigning” their workflows: nearly three-quarters of high performers, against about a quarter of everyone else.
In a sidebar 18 pages later, associate partner Tara Balakrishnan explains why adoption keeps moving more slowly than the technology does. "The limiting factor," she writes, "is increasingly the organization's ability to absorb change."
Both of those claims are in the same report. Only one is actionable.
AI high performers, the respondents attributing at least 5% of EBIT to AI and calling its impact significant, totaled about 6% of the sample in 2025. In 2026 they are about 6% of the sample. Flat. Almost everything else in the report went up.
Everything, Except What Matters Most
The survey ran from May 4 to June 8, 2026, with 1,719 respondents across 97 nations, and by every measure of deployment it is a story of acceleration.
Organizations scaling AI across the enterprise rose from 38% to 44%. Use of AI in three or more business functions rose from 51% to 56%. Regular use in at least one function now stands at 89%. More than a quarter of organizations now put over 10% of their entire information and communication technology budget into AI, and 60% of respondents expect to increase that investment again next year.
The individual-level results are better still: 80% of respondents say AI has improved their own productivity and about half say it helps them make better decisions, and those figures hold steady from the C-suite down to individual contributors.
Then the enterprise line. Some 37% report that AI contributed positively to their organization's EBIT, essentially unchanged from a year ago.
More deployment, more functions, more money, more productivity, same financial result. That's not a story about AI being oversold. Gains that consistent are likely real.
This is a story about where those gains stop.
These Practices All Go to 11
The report compares high performers against everyone else on 11 organizational practices: transformative ambition, human in the loop, active cost management, strategic workforce planning, performance management, workflow redesign, a transformation office, a semantic layer for company data, senior-leader role modeling, budget commitment and impact tracking.
High performers lead on all 11.
That result isn't a checklist, it's a warning. When winners are ahead across every practice, you probably measured the same thing 11 times. These are not 11 independent levers a company can pull in any order. They look more like 11 symptoms of a single underlying capability: an organization's capacity to change itself.
Which is precisely what Balakrishnan named in the sidebar.
One honest caveat on the exhibit: the high-performer base is 92 respondents against 1,429 others, so treat the individual practice gaps as directional rather than precise. The pattern across all 11 is the durable part.
The Sidebar Was the Killer Insight
Read McKinsey’s two claims together and the prescription inverts.
Fundamental workflow redesign is the most absorption-expensive thing an organization can attempt. Prescribing it to companies whose binding constraint is absorption capacity is a diet that requires you to already be thin.
We also have a full year of evidence that the advice does not transmit on its own. Last year's report was built around how organizations are rewiring to capture value, and workflow redesign among high performers climbed from 55% to nearly 75% over the same period. The recipe was public. Deployment deepened in every direction measured. The share of companies converting any of it into EBIT did not move.
To be fair, the report cannot tell us which way the causality runs, and it does not claim to. High performers may redesign workflows because they can absorb change, or they may have learned to absorb change by redesigning workflows. The survey is correlational and McKinsey presents it that way. The problem starts downstream, when a correlation gets read as a recipe and a year of budget goes into following it.
Why this matters for anyone writing next year's AI investment plan: if capacity to change is the real constraint, the plan's currency is not dollars or seats or use cases. It is how much change your organization can finish.
Absorption Capacity Is Built, Not Spent
"Your constraint is absorption capacity" sounds like a counsel of despair. It is the opposite, once you notice how the constraint behaves.
Capacity to absorb change works like a muscle, not a budget. Big-bang redesign consumes it and frequently returns nothing, because the redesign never completes, and an unfinished transformation leaves an organization more resistant than it found it. Anyone who has watched a two-year process redesign get quietly rescoped knows the real cost. The next proposal starts in a deeper hole than the last one did.
A change small enough to finish does the reverse. It ships, the organization discovers it survived, and the following one meets less friction. Capacity is not spent by planning. It is built by completing.
That reframes the sequencing question. The goal for the coming year is not the most ambitious redesign the budget will support. It is the largest number of changes the organization can actually finish.
Where AI ROI Actually Shows Up
All of the above raises the obvious question. Finish what?
We have argued for some time that the value sits at the boundaries between systems rather than inside them, and that most of what companies call an automation problem is a coordination problem. The handoff between two teams, two tools or two datasets is where context evaporates and where nobody owns the loss.
The new data is consistent with that. High performers are furthest ahead scaling agents in knowledge management (32%) and software engineering (30%), both of which span boundaries by nature rather than sitting inside one desk.
There is a second reason to pick the boundary, and the absorption argument is what surfaces it. A redesign scoped to a single boundary is absorption-cheap. It touches two teams, not 12. It has an owner on each side. It can be finished inside a quarter.
So the cheapest unit of change and the highest-value unit of change turn out to be the same unit. That is less a happy coincidence than a consequence of one fact: boundaries are small, and boundaries are where the loss already is.
It is also why adoption is not the same thing as ROI. Deploying a tool inside a role produces the 80% productivity number. Closing a boundary produces the EBIT number. This report measured both, and only one of them reached the P&L.
Testing Is About the Absorption Budget
When we published Minimum Viable Autonomy, the scoping rule answered a technical question. Pick a domain "narrow enough to be manageable but complex enough to demonstrate functional competence," because an agent has to prove it can do the job before anyone asks whether the market wants it. Before agents can walk to product-market fit, they need to crawl to task competence.
The absorption constraint gives that same rule a second job. Narrow is not only how you prove the agent works. It is how you spend scarce organizational capacity in increments that pay themselves back before you draw on the next one.
Which means the scoping conversation now carries two questions instead of one. Can we prove this agent is competent in this domain? And can this organization finish this change? A scope that passes the first test and fails the second is how a working pilot becomes a shelved pilot.
For what it is worth, eight weeks is a realistic target to get a boundary-scoped change into production. Not because the work is trivial, but because a scope chosen to finish is a different scope than one chosen to impress.
The Question for Next Year's Plan
The report's framing is that organizations need to transform, not merely adopt. That is true, and it is not actionable, which is a reasonable explanation for why the high-performer share has not moved in a year.
The actionable version is smaller and considerably less satisfying. Find the boundary losing the most context. Scope the change so it finishes. Ship it, and let the organization learn that it can. Then do it again with slightly more range than you had before.
The companies in that 6% are not there because they read the list of 11 practices and worked through it. They are there because they built the capacity to change, one completed change at a time, and the practices followed.
One question McKinsey doesn't suggest, but should have: does your AI plan describe the AI you intend to deploy, or the change your organization needs to finish?
McKinsey called it the road to ROI. Most companies are on it. Almost nobody is arriving.
Agentic Foundry: AI For Real-World Results
Learn how agentic AI boosts productivity, speeds decisions and drives growth
— while always keeping you in the loop.
