Two lines that used to move together came apart this year. Organizations above $1 billion in annual revenue went from 27% to 40% scaling AI agents in at least one function. Organizations below that line went from 22% to 22% (McKinsey, 2026).
Not slower. Unmoved.
That is the finding worth sitting with from McKinsey's state of AI survey, published August 25, 2026 โ 1,719 respondents across 97 nations, fielded May 4 to June 8, weighted by each nation's contribution to global GDP. In the year agents stopped being a demo and became a production pattern, one cohort converted and the other did not convert at all.
The reflex explanation is budget. The data does not support it. The second reflex is skills. The data does not support that either. What is left is less comfortable and more actionable: scaling AI agents is an execution-capacity problem, and execution capacity is the one input a lean operations team cannot buy on a purchase order.
What the Survey Actually Measured
A caveat first, because it changes how you should read the number. McKinsey's cut is by revenue, not headcount. "Smaller organizations" means under $1 billion in annual revenue โ a bucket that contains plenty of companies considerably larger than a 200-person operation. Treat 22% as directional for the mid-market, not as a mid-market census.
With that stated, the structure of the finding is unusually clean:
- Scaling agents in one or more functions: >$1B organizations moved 27% โ 40%. Smaller organizations held flat at 22%.
- Enterprise-wide AI scaling (any AI, not just agents): 54% of large organizations against roughly one-third of smaller ones.
- EBIT impact: 37% of all respondents attribute at least some earnings impact to AI โ unchanged year over year.
- AI high performers (at least 5% of EBIT attributable to AI, with self-described significant impact): flat at about 6% of respondents.
McKinsey also validated the year-over-year comparison against the 552 respondents who completed both the 2025 and 2026 waves, which matters โ a widening gap in a repeated cross-section is often just a changed sample. This one survives the panel check.
So the aggregate story is stasis: EBIT impact flat, high performers flat. Underneath the flat aggregate, one cohort pulled away by 13 percentage points on the single capability that defined the year.
Budget Is Not the Binding Constraint
The most common explanation I hear from operations leaders is cost โ agents are expensive to run, inference bills are unpredictable, the enterprise can absorb that and we cannot.
The survey tests this directly. Only about one in five respondents say AI-related operating costs, including token costs, constrained their use of AI. 60% expect their organization to increase AI investment over the next year. 28% already spend more than 10% of total enterprise ICT budget on AI (McKinsey, 2026).
A cohort that is planning to spend more, and does not identify spend as its limiter, is not stalled by money.
There is a related finding that quietly reinforces the point: 32% of respondents decided against buying at least one software product or feature because they could build it in-house with agentic coding tools. The organizations scaling agents are not just spending differently. They are re-pricing build-versus-buy. That is a capability shift, not a budget line.
And It Is Not the Skills Gap Either
The second explanation is talent โ the enterprise has AI engineers and the mid-market does not.
The cleanest counter-evidence comes from a survey whose sample is the mid-market rather than enterprise data translated downward. CLA's Heartbeat Index, released in August 2026 across 722 small and middle-market organizations, found 69.9% of leaders believe their workforce already has the skills needed for future success. Yet only 49.5% reported any positive improvement in efficiency from recent technology or AI investments, and just 20.2% reported highly positive impacts (CLA, 2026).
That is an inversion, not a gap. The population under study says the capability is present and the results are absent.
Hold one caveat: self-reported skill confidence is a soft instrument, and 69.9% may be measuring optimism. But even discounted, it tells you what mid-market leaders believe โ and therefore where they are not looking for the problem.
If budget is ruled out by the spenders and skills are ruled out by the workforce that has them, the remaining candidate is the work itself.
Scaling an Agent Means Owning a Redesigned Workflow
Here is the mechanical difference between a pilot and a scaled agent, and it is not model quality.
A pilot runs beside the existing process. Somebody's spare Thursday, a copy of the data, an output that a human reviews and then re-enters into the real system. Nothing downstream depends on it. It can be abandoned without a meeting.
A scaled agent is the process. It means the handoffs get rewritten, the exception path gets defined and owned, the audit trail satisfies whoever signs off, and a named person is accountable when the agent is wrong at 2am. None of that is procurement. All of it is organizational surgery performed by the same three people already running the quarter.
The research points the same direction from two independent angles.
Deloitte's State of AI in the Enterprise 2026 โ 3,235 leaders across 24 countries โ found the response to AI is lopsided toward the cheap intervention. 53% are educating the workforce to raise AI fluency. Only 30% are reimagining the organization around new AI patterns and 33% are redesigning career paths. Just 34% are using AI to deeply transform products, processes, or business models; the other two-thirds remain incremental (Deloitte, 2026).
Microsoft's 2026 Work Trend Index โ 20,000 workers across 10 countries, fielded February 18 to April 20 โ quantifies what that lopsidedness costs. Organizational factors such as culture, manager support, and talent practices drive more than twice the realized AI impact of individual factors: 67% versus 32%. And only 26% of AI users report that leadership is aligned on AI strategy (Microsoft, 2026).
Training is a budget decision. Redesigning a workflow is a decision-rights decision. Large enterprises have people whose entire job is the second kind. Mid-market operations teams have the same three people who also close the month.
That is the execution-capacity story โ and to be explicit about epistemic status, the causal link is my reading, not McKinsey's finding. What McKinsey establishes is that the divergence exists and that cost does not explain it.
The Honest Counter-Argument: Maybe Flat Is Correct
Take the opposing case seriously, because parts of it hold.
Aggregate returns have not moved. EBIT impact is flat at 37% and high performers flat at about 6%. A smaller organization that sat out a year in which the measurable return did not improve has not obviously lost anything. Waiting for tooling to mature and unit costs to fall is a defensible allocation of scarce attention.
Agent-scaling rate is not a business outcome. It is an input metric. Chasing it because a survey says the enterprise is chasing it is the purest form of pilotitis โ funding motion to close a benchmark rather than a constraint.
Enterprise scaling is partly a denominator artifact. A 40,000-person company has more functions, so the odds that at least one has scaled an agent rise mechanically with size. Some of the 13-point move is surface area, not superiority.
Where the counter-argument fails is in what it costs to start late. The capability being built in that 13-point move is not agent expertise. It is the muscle of re-architecting a workflow end-to-end and living with the consequences โ which compounds, transfers across functions, and cannot be purchased later at speed. A competitor that redesigned order-to-cash this year will redesign quote-to-contract faster next year. You will still be pricing your first redesign.
The correct conclusion is not "scale agents." It is: run one real redesign this year so the capability exists when the returns do show up.
The One-Workflow Test
The practical move is to stop funding a portfolio of pilots and re-architect exactly one workflow end-to-end this quarter. Choose it with four questions. If a candidate fails any of them, it is the wrong workflow โ not because agents cannot do it, but because you will not finish.
- Does one person own the entire path? If the workflow crosses a boundary where two leaders must agree, you are buying a negotiation, not a deployment. Pick inside one span of control for the first one.
- Do you have a baseline you measured before touching it? Cycle time, error rate, cost per transaction, volume. Without a pre-deployment number you cannot distinguish a real gain from a plausible story โ and you will default to "it feels faster."
- Is there a defined exception path with a named owner? Agents do not fail loudly; they fail plausibly. Decide now who reviews what share of runs, and what happens when the agent is confidently wrong.
- Is it boring, high-volume, and internally visible? High volume gives you statistical signal within a quarter. Internal visibility means the first failure is a lesson, not a customer incident.
Then gate it. At the end of the quarter the workflow is scaled, killed, or explicitly rebooked with a new hypothesis. Silent drift โ the pilot that neither dies nor ships โ is what produces a flat line at 22% for a year.
The Gap Is Compounding, and It Is Not Made of Money
The number that should bother you is not 40%. It is 22% โ the same 22% as last year, in a cohort that says it is not constrained by cost and believes its people are ready.
Flat, with money available and skills self-reported as present, is not a resourcing problem. It is a sequencing problem. Organizations that succeeded at scaling AI agents did not buy more; they committed to owning one redesigned workflow all the way to the exception path.
Before this quarter closes, name the single workflow you will re-architect end-to-end, name the person accountable for it, and write down the baseline number today โ while it is still measurable.
