McKinsey 2026: Agent Scaling Hit 40% at Large Firms, Stayed at 22% Elsewhere

McKinsey's State of AI 2026 survey found that 40% of respondents at organisations with more than $1 billion in revenue say their company is scaling AI agents, up from 27% a year earlier. At smaller organisations the figure sat at 22% in both years. The size gap in AI agent adoption went from 5 points to 18 in twelve months, and the survey cannot tell you why.

The short answer. McKinsey's 2026 survey says large companies are scaling AI agents far faster than smaller ones: 40% versus 22% of respondents. The gap is real enough to survive sampling noise. What the survey cannot show is why it opened, how deep the scaling goes, or whether any of it pays.

What McKinsey's State of AI 2026 reports on AI agents

The report, titled The state of AI in 2026: On the road to ROI, was published on 25 August 2026. It rests on an online survey in the field from 4 May to 8 June 2026, with 1,719 participants in 97 countries.

The agent finding is narrow and specific. Among respondents at organisations with more than $1 billion in annual revenue, 40% say their company is scaling AI agents in at least one function. A year earlier the same group said 27%. Among respondents below that revenue line, the share is 22%, unchanged.

The rest of the report moved more slowly. 44% say AI is scaling across the enterprise, up from 38%. 37% attribute at least some EBIT impact to AI, roughly where it was. About 6% qualify as high performers, defined as attributing at least 5% of EBIT to AI and calling the impact significant.

Those outcome numbers, and why they look so different from vendor surveys, are taken apart in the comparison of NVIDIA's and McKinsey's AI ROI surveys. This post stays on one variable the headlines skipped past: company size.

What "scaling" means in this survey

Scaling is a self-reported stage, not an audited state. In McKinsey's framing it sits above experimenting and piloting, and the bar is at least one business function. It is a respondent's judgement about their own company, not a count of deployed agents.

For context, the 2025 edition reported 23% of respondents scaling agentic systems somewhere and 62% at least experimenting, per CX Today's summary of that edition. The 2026 coverage I could open reports the split by size but not a single all-respondent agent figure, so this post does not quote one.

"At least one function" is a low bar for a large company. A bank with 40 functions that runs one agent on its IT service desk counts as scaling. So does a bank running agents across 12 functions. The survey groups them together, and so does every headline written from it.

The AI agent adoption gap by company size went from 5 points to 18

Put the published shares side by side and the change is stark. The table below uses McKinsey's figures. The gap column is our arithmetic.

Percentages are of survey respondents, weighted by national GDP, not of companies. The enterprise-wide row is from FM Magazine's reporting of the 2026 report, which describes the smaller group as "about one-third".

The ratio tells the same story more sharply. In 2025 a respondent at a large organisation was about 1.2 times as likely as one at a smaller organisation to report scaled agents. In 2026 the multiple is about 1.8. That is the computed figure worth remembering from this survey, and none of the coverage I read stated it.

Flat at 22% is the more interesting number

A gap that widens because leaders sprint is normal for a new technology. Large firms adopted cloud, mobile and analytics first too. What is unusual here is that the smaller group did not move at all, in a year when agent features arrived inside software these firms already pay for.

Remember what "smaller" means in this survey. The line is $1 billion of revenue, so a $900 million manufacturer with a 60-person IT department sits in the 22%. This is not a story about corner shops. It is a story about the mid-market and the lower enterprise standing still.

My read is that the flat 22% is a capacity signal rather than an interest signal. These firms are not ignoring agents. They appear stuck at the step between a working pilot and a supported production system, which is where engineering, data and governance costs pile up. The specific ways pilots die at that step are catalogued in the breakdown of AI agent pilot failure modes.

Who answered the McKinsey survey, and how much the size split can bear

McKinsey surveyed 1,719 participants in 97 countries. To adjust for differences in response rates, the data are weighted by each respondent's nation's contribution to global GDP. About 36% of respondents say they work for organisations with more than $1 billion in annual revenue.

That makes the size split a comparison of two subgroups, not two random samples of companies. The detail matters most for the number that did not move.

Respondents, not companies

Every percentage in the report is a weighted share of people. Two respondents from one company count twice. A strategy lead and an IT lead at the same firm may answer differently about whether agents are "scaling", and both answers enter the total. Nothing in the material I could open says responses were deduplicated by firm.

GDP weighting means answers from the largest economies count for more than answers from smaller ones. That is a defensible choice for a global economic picture. It also means a size split can partly reflect where large companies are headquartered, a point picked up in the limitations section below.

How much noise sits in a 22%

36% of 1,719 is roughly 620 respondents at $1 billion-plus organisations and about 1,100 below that line. If those were simple random samples, the 95% margin of error would be about 3.9 points around 40% and 2.4 points around 22%. They are not random samples. This is an opt-in online survey, and weighting shrinks the effective sample, so the real uncertainty is wider.

The 18-point gap survives that noise comfortably. A 13-point move at large firms is too big to be sampling error alone. The claim that smaller firms are "flat" is less safe. A real move of 2 or 3 points either way would be invisible at this sample size. So "no detectable change" is the honest wording, and it is the wording I would use.

Enterprise AI scaling by company size in surveys that publish their samples

Two US data sources publish their method and break results out by firm size. Neither measures agents. Both measure something more basic, and the contrast with McKinsey turns out to be the most useful thing in this post.

Census Bureau: AI use in any business function

The Census Bureau's Business Trends and Outlook Survey is a biweekly, nationally representative survey of US businesses. Census statisticians Adam Grundy, Cory Breaux and Dhanapati Khatiwoda published a May 2026 analysis covering 14 December 2025 to 3 May 2026. It put the national AI use rate at 19.8% for the period ending 3 May 2026.

The Census analysis found 37% of firms with at least 250 employees reported using AI, and 32% of firms with 100 to 249 employees. Fewer than 20% of firms with four or fewer employees did. Most telling, AI use rose among firms with at least 20 employees but did not change significantly among firms with fewer than 20.

That is the Census version of McKinsey's flat line, at a much smaller scale of firm. One caveat travels with it. Census changed its question in November 2025 from AI use "in producing goods or services" to use "in any business function", so older trend lines do not join cleanly.

Ramp AI Index: paying for any AI product

Ramp measures the share of businesses with an observed payment for an AI product or service in a given month. The panel is more than 70,000 US businesses on its card and bill pay platform. The series are published on FRED. In September 2026, 50.0% of small businesses, 62.5% of medium businesses and 65.7% of large businesses were paying for AI, per the FRED series.

Ramp's Summer 2026 Business Spending Report defines small as 1 to 24 employees, medium as 25 to 99 and large as 100 or more. The FRED notes do not restate those bands, so treat them as Ramp's general convention rather than a confirmed AI Index definition.

The year-on-year change runs the other way from McKinsey. Small businesses rose 11.5 points, about 30% in relative terms. Large businesses rose 14.0 points, about 27%. The point gap widened from 13.2 to 15.7, but the ratio of large to small narrowed from 1.34 to 1.31.

Here is the pattern I take from the three. The deeper the commitment a survey measures, the wider the size gap it finds. Paying for a tool, small firms are catching up. Using AI inside a business function, small firms stalled. Scaling agents into production, the below-$1B group stalled and large firms pulled away. That is my interpretation of three instruments that were never designed to be compared, not a finding any of them makes.

Why AI agent adoption is pulling ahead at large firms

None of the coverage of the 2026 report I could open quotes McKinsey explaining the size gap. What follows is reasoning from the report's other findings, labelled as such.

Agents need plumbing that large firms already paid for. An agent that acts, rather than one that drafts, needs an identity, scoped permissions, API access to the systems it touches, logging and someone on call. Large firms built most of that for earlier automation waves. A $400 million company often has to build it first, and the cost of doing so is set out in the enterprise AI agent cost model.

Cost of the AI itself does not look like the main brake. About one in five respondents say AI operating costs, including tokens, have constrained their AI use. For agents specifically, Civic's summary of the report puts that at about one in ten. If inference prices were what held smaller firms back, the agent number would be the higher one.

Workflow redesign is a fixed cost, and fixed costs favour scale

The report describes high performers as much more likely to redesign workflows around AI. The others insert it into existing ones. Redesign means process analysts, change management, retraining and months of parallel running. Most of that cost does not shrink with the size of the business.

A large firm spreads that fixed cost across far more transactions. A firm one tenth the size pays a similar design bill for one tenth of the volume, so the payback on the same agent is slower. I think this, more than model access or price, is why the gap will keep widening through 2027. What the minority that does scale actually did differently is covered in the analysis of what the organisations scaling AI agents did differently.

There is a quieter mechanism too. A firm with more functions has more chances to have one agent in production. "At least one function" therefore favours size mechanically, even if adoption per function were identical. That point cuts against the rest of this section, and it returns below.

What the State of AI survey does not measure

The survey is the most cited annual read on enterprise AI, and it is careful about its own scope. The problem is what headlines infer from it. Five things it does not tell you, each of which changes how the 40% should be read.

Depth. One agent in one function and dozens of agents across a business both count as scaling. The survey does not report how many functions, agents or tasks sit behind the answer. Counting deployments is its own trap, as the case against AI use case counts as a metric sets out.

Outcomes per agent. 37% report some EBIT impact from AI overall. No published figure connects agent scaling to profit, cost per task or headcount, by size or otherwise.

Reversals. The survey captures a stage on the day a respondent answers. A firm that scaled an agent in 2025 and pulled it back in 2026 shows up as a lower stage, with no record of the failure. Flat can mean nobody moved, or that some advanced while others retreated.

Firm-level truth. Answers are perceptions from individuals, weighted by country. There is no verification against deployment logs, contracts or spend.

Variation inside "below $1B". A $5 million agency and a $950 million distributor sit in one bucket. Whatever is happening inside the 22% is invisible, including whether the mid-market is moving while the smallest firms fall back.

Where this argument is weakest

Three objections deserve a straight hearing, and one of them could undo the headline.

First, the "at least one function" mechanism. Large firms simply have more functions. If each function had the same chance of scaling an agent, firms with more functions would still report scaling more often. That alone can produce a gap without any difference in capability. The survey data I could see cannot separate the two explanations, and this post's capacity argument rests on judgement, not on a test.

The size split may partly be a geography or vocabulary split

GDP weighting gives heavy influence to the largest economies, where many $1 billion-plus respondents work. If agent adoption in those economies rose faster for reasons unrelated to firm size, part of the gap is geography. The report coverage does not publish the size split by region, so this cannot be checked from outside.

Vocabulary matters too. A smaller firm may run a vendor's support bot that resolves tickets on its own, and its respondent may not call that "scaling AI agents". A large firm's respondent describing a bespoke build will. Ramp's data shows small firms buying AI products quickly, and some of those products are agentic. The 22% may undercount agent use that smaller firms consume rather than build.

Finally, the comparison across three surveys in this post is interpretive. McKinsey, Census and Ramp count different units, in different countries, against different definitions. The "deeper commitment, wider gap" pattern fits all three, and three non-comparable points are not a trend.

Frequently asked questions

Where to start if your firm sits in the 22%

Do not try to close an 18-point gap with a strategy deck. Close it one function at a time.

This week, pick the single function where an agent already runs in a pilot and write down three things. Who owns it in production? Where are its actions logged? What would it cost to switch off? If any of the three is blank, that blank is your scaling plan, and it is cheaper to fill than a new pilot.

Then read the next McKinsey edition, or any adoption survey, with two questions ready: what counted as "scaling", and how many respondents sit behind the subgroup you care about. If the report cannot answer both, quote its direction and not its decimal places.

Related reading