How Many Companies Have Actually Deployed AI Agents?

Every figure in this post comes from a primary document: the original survey writeup, the government data series, the academic paper, or the analyst release. We opened every URL on August 5, 2026 and confirmed the number is on the page. Aggregator roundups, the ones with titles like 50 AI Agent Statistics for 2026, were excluded as sources entirely, because most of their numbers cannot be traced past a chain of blogs citing each other.

Three rules governed what got in. Every figure carries its population and its field dates, because 42 percent of enterprise IT leaders and 42 percent of all US businesses are different facts wearing the same number. Vendor surveys of their own customers are cited where they are honest and labeled every time, and they are never counted as independent evidence. And nothing is averaged: when two credible sources disagree, both numbers are here with the difference named.

What would change our conclusion: a large-sample independent survey that asks specifically about agents in production and lands above 50 percent. As of August 2026, no such result exists. Strict wording lands at 10 to 25 percent in every independent instrument we could find, and we looked.

Line the published surveys up side by side and the pattern is hard to miss. The strict questions cluster low. The loose questions cluster high. The vendor-commissioned questions cluster highest.

Nobody in this chart is fabricating data. Gartner's 17 percent counted organizations that have deployed agents to date, asked of 2,501 CIOs between May and June 2025. PwC's 79 percent counted executives who say agents are already being adopted at their company, asked of 308 US executives in April 2025, more than half of them director level. Adopted includes every pilot, proof of concept, and single-team experiment in the building. Deployed to date, asked of the person who owns the infrastructure, does not. That single wording difference is worth about 60 percentage points.

Four things move the number, and once you know them you can predict any survey's result from its methodology page.

Question wording is worth the most. KPMG demonstrated this inside its own instrument: its Q1 2026 pulse of 237 US leaders at billion-dollar firms found 54 percent actively deploying agents, and the same survey's technology-sector cut came in at 80 percent. One quarter later the Q2 wording shifted to currently using AI agents and produced 53 percent. Three numbers, one survey family, five months.

Seniority is second. PwC's respondent pool was 54 percent director level, and directors say yes to adoption questions more readily than the CIOs who own the uptime. Gartner asked the people who carry the pager, and got 17.

Firm size is third. KPMG surveys companies above 1 billion dollars in revenue. Salesforce's Connectivity Benchmark starts at 1,000 employees. The Census Bureau's BTOS series samples actual US businesses, most of which are small, and it found about 10 percent using AI in production as of October 2025. Every big adoption number you have seen excludes the majority of companies that exist.

And fourth, who is asking. Salesforce found 83 percent of enterprise IT leaders saying most or all teams have adopted agents in a survey it commissioned of 1,050 IT leaders in late 2025. Boomi's commissioned Forrester study of 409 IT decision-makers found 86 percent beyond pilot in 2026. A vendor's number about its own market sitting 60 points above the independent measurement is not a scandal. It is a finding, and it should change how much weight the number gets in your planning.

Ask a different question: not whether an organization has agents anywhere, but whether agents run wide and deep enough to matter. Every source that measures depth converges on a much smaller number.

Among companies already adopting agents, only 17 percent have them embedded in almost all workflows, per PwC's April 2025 survey of 308 US executives. IBM's 2025 study with Oxford Economics found 16 percent of executives reporting agentic AI scaled enterprise-wide, against 70 percent calling it important to their future. McKinsey's 2025 survey of 1,993 organizations found no single business function where more than 10 percent of respondents are scaling agents. Accenture's 2025 survey of 2,000 companies classified 8 percent as front-runners scaling AI with strategy attached. Deloitte's August to September 2025 survey of 3,235 leaders found 25 percent had moved even 40 percent of their AI experiments into production.

So the honest deployment picture in one sentence: most organizations are somewhere, almost none are everywhere, and the everywhere number is 8 to 17 percent depending on the bar. Part of the confusion is that agent means different things in different shops, which is the same definitional fight we walked through in RPA versus AI agents: plenty of things reported as agents are scripted bots wearing the season's name.

One count runs the other way and is worth naming rather than hiding. An arXiv study titled Measuring Agents in Production, fielded July to October 2025, found 82 percent of the 111 practitioners who answered its deployment question had systems in production or pilot. That is a survey of people who build agents for a living answering about their own systems. It measures the practitioner frontier, not the economy, and the gap between 82 and 10 is a decent estimate of how far apart those two worlds currently sit.

Whatever the true adoption number is, the spending number is enormous and accelerating. US organizations with revenue above 1 billion dollars plan an average of 202 million dollars in AI investment over the next 12 months, per KPMG's Q2 2026 pulse of 204 senior leaders, barely down from 207 million in Q1. BCG's AI Radar, fielded across roughly 2,400 executives in 16 markets and published January 2026, has companies planning to spend 1.7 percent of revenue on AI in 2026, up from roughly 0.8 percent a year earlier, with the 640 CEOs in the sample steering about 60 percent of their AI budgets to agents specifically. Gartner's May 2026 forecast puts worldwide AI spending at 2.59 trillion dollars for the year, up 47 percent.

Now set one number against that wall of money: 26 percent. That is the share of the same KPMG population that has full real-time visibility into what their AI agents cost to operate, from the same Q2 2026 survey. Companies are budgeting nine figures against a cost base three quarters of them cannot see. PwC found 88 percent of US executives planning to increase AI budgets in the next 12 months because of agentic AI back in April 2025, so this is not a blip. It is the operating posture of the market: spend first, instrument later. We wrote about where that ends for smaller teams in the point where a no-code stack stops working, and the enterprise version of the same mistake just has more zeros on it.

Here the data splits cleanly in two, and the split is the most useful thing in this post.

Ask adopters how it is going and the news is good. Two thirds of companies adopting AI agents, 66 percent, report increased productivity, and 57 percent report cost savings, both from PwC's April 2025 survey of 308 US executives. That is a real signal from real deployments.

Ask the finance function and the temperature drops about thirty degrees. McKinsey's 2025 survey of 1,993 organizations found 39 percent could attribute any EBIT impact to AI, and only around 6 percent qualified as high performers with 5 percent or more of EBIT moved. IBM's 2025 study of 2,000 CEOs across 33 countries found just 25 percent of AI initiatives had delivered their expected ROI. BCG's 2025 Build for the Future research classed 60 percent of companies as getting minimal material value from AI, and a separate McKinsey growth-practice piece from 2025 put organizations seeing no significant bottom-line change at nearly eight in ten.

Self-reported productivity is a feeling. EBIT is a measurement. Both are honest answers to different questions, and the 40-point gap between them is where most of the disappointment in this market lives.

The distribution matters more than the average, though. BCG's same 2025 research found the top 5 percent of companies, the ones it calls future-built, pulling 5 times the revenue impact and 3 times the cost reduction of everyone else. And the macro signal agrees that the prize is real: PwC's Global AI Jobs Barometer, built on close to a billion job ads and thousands of company filings, measured productivity growth from 2018 to 2024 at 27 percent in the industries most exposed to AI against 9 percent in the least exposed. The value exists. It is just concentrated in the companies that deployed properly, which is exactly what you would expect if deployment quality, not deployment count, is the scarce thing.

A trend needs the same measurement, on the same population, at three or more points. Almost nothing in this market clears that bar, which is why almost everything called a trend in the roundups is actually two numbers and a line drawn between them. Three series clear it.

McKinsey's State of AI series has organizations using AI in at least one business function at 55 percent in April 2023, 72 percent in early 2024, and 88 percent by mid 2025, on samples of 1,363 to 1,993 global respondents. Stack Overflow's annual developer survey has developers using or planning to use AI tools at 70 percent in 2023, 76 percent in 2024, and 80 percent in 2025, on samples between 49,000 and 90,000. And the Census Bureau's Business Trends and Outlook Survey, analyzed by the St. Louis Fed, has actual US businesses using AI in production at roughly 4.4 percent across 2023 and 2024, about 7 percent in the first quarter of 2025, and about 10 percent by October 2025.

All three move the same direction. The levels differ by a factor of nine, and the St. Louis Fed's own June 2026 write-up explains why with a finding that should be stapled to every adoption chart: when the Census broadened its question wording, measured adoption roughly doubled overnight. Same firms, same technology, different sentence. How you ask is worth more than a year of actual adoption.

One more honest caveat, because two points is not a trend but some two-point deltas are still worth knowing: Stack Overflow's 2026 pulse of 1,100 technologists found AI agent use at work jumped from 31 percent in 2025 to 59 percent in 2026, and KPMG's pulse series has active deployment at billion-dollar firms going 12 percent in early 2024, 33 percent in Q2 2024, 54 percent by Q1 2026. Direction is not in dispute anywhere in this dataset. Only the altitude is.

Four figures circulate hard in this market, and each one deserves its asterisk printed at full size.

The 97 percent. Writer's claim that 97 percent of companies deployed AI agents in the past year appears in hundreds of roundups. We could not find a published population, sample size, or field methodology anywhere, and the company's own blog post blocks automated access, so the number cannot be checked against anything. A statistic with no methodology is a slogan. It may even be true of some population. Nobody can say which.

The 40 percent cancellation number. Gartner's June 2025 release predicting that over 40 percent of agentic AI projects will be canceled by the end of 2027 is cited constantly as evidence that agent projects are failing. Read the sentence again: will be. It is a forecast, not a measurement, and no published study has yet measured actual agent project cancellation rates. Gartner followed it in May 2026 with a second forecast that 40 percent of enterprises will demote or decommission agents by 2027 over governance gaps found only after a production incident. Both may prove right. Neither is data about what has happened.

The vendor adoption numbers. Salesforce's 83 percent of enterprises with most or all teams on agents, from its commissioned survey of 1,050 IT leaders in late 2025, and the Forrester-for-Boomi 86 percent beyond pilot from 409 IT decision-makers in 2026, sit 60-plus points above Gartner's independent 17. The same Boomi study found only 34 percent of those leaders trust what their agents are doing, which tells you the 86 is measuring something closer to has tried than has trusted with production.

And the c-suite double-count. IBM's 2026 study of 2,000 CEOs found 76 percent of organizations now have a Chief AI Officer, up from 26 percent in 2025. BCG's AI Radar, fielded the same season across 640 CEOs, found 72 percent of CEOs saying they personally are the main AI decision maker. Both cannot describe the same companies unless most CAIOs do not actually decide anything, which, come to think of it, might be the finding.

The most useful part of a benchmark is usually the blank space, so here is what we could not find after opening 48 URLs across 11 publishers, stated plainly so you do not waste an afternoon looking for it.

Nobody measures small business agent adoption. Every executive survey in this post samples firms above 1 billion dollars in revenue or above 1,000 employees. The Census BTOS series covers small firms but asks about AI generally, not agents. If you run a 30-person company, no number in this post describes your peer group, and anyone selling you an adoption stat about your peer group made it up.

Nobody has measured agent project failure. Only forecast it. The actual cancellation, rollback, and decommission rates of deployed agent systems are unpublished as of August 2026.

Nobody defines deployed. Not one survey in our set prints its definition of deployment next to its number. Pilot, production, and proof of concept get sorted by the respondent's own judgment, which is exactly how a director and a CIO end up 60 points apart about the same company.

Nobody resurveys the same panel except the Census and Stack Overflow. McKinsey's 55 to 72 to 88 climb is three different samples answering three years running, so some unknown share of the climb is sample drift rather than adoption.

Nobody splits built from bought. A company running one purchased support bot and a company running eleven custom agents on its own infrastructure both count once. For anyone making a build-versus-buy decision, the adoption number that would actually help does not exist.

And nobody audits ROI. Every return figure in this market is self-reported, resting on cost visibility that only 26 percent of large US firms claim to have per KPMG's Q2 2026 pulse. The honest version of most published agent ROI is: we think it is positive, and we cannot fully see the denominator.

Strip the definitional noise and three planning facts survive.

First, you are earlier than the pitch decks say. If agents at production scale sit at 10 to 25 percent of organizations, per every independent 2025 and 2026 measurement in this post, then being without them puts you with the majority, not the stragglers. The everyone-already-has-this pressure comes from surveys built to produce it. What the data actually punishes is not being late. It is deploying badly: the 60 percent of companies BCG found getting minimal value in 2025 mostly did adopt something.

Second, the winners are winning on depth, not presence. BCG's future-built 5 percent pull 5 times the revenue impact of everyone else in 2025, and the productivity divergence in PwC's jobs data, 27 percent against 9 from 2018 to 2024, accrued to industries that actually restructured work. One accounting firm in the St. Louis Fed's May 2026 Eighth District survey reported 30 to 40 percent more output capacity after adopting AI, and it got there by rebuilding how work moved, not by turning on a copilot.

Third, the labor market has already voted. Workers using agents at work went from 31 to 59 percent in a year per Stack Overflow's 2026 pulse of 1,100 technologists. Organizations with a Chief AI Officer went 26 to 76 percent in a year per IBM's 2026 study of 2,000 CEOs. AI-skilled workers carry a 62 percent wage premium across 27 countries per PwC's 2026 Jobs Barometer. Whether or not the agents are in production yet, the hiring, the titles, and the wages are being repriced now.

So the decision in front of an operator is not whether to believe 17 or 83. It is whether your specific workflow clears its own ROI bar, which no industry average can tell you. That is a question you answer by shipping one production agent against one measured workflow and reading the number, which is the entire argument for a small scoped build over a strategy engagement. It is also the honest version of what an ai automation agency is for: not handing you an adoption statistic, but getting you to your own number in weeks. At agentclaw we run that as a two-week production sprint with evals attached, and when the workflow will not clear its bar, we say so on the first call, because a benchmark post full of other people's inflated numbers is a bad place to add our own.

If you are weighing the routes, agency versus in-house hire versus off-the-shelf walks the same decision with the costs attached.