AI Automation Statistics 2026: 44 Verified Numbers on Adoption, Time Saved and Failure

Most AI automation statistics measure one thing: whether a company uses AI at all.

Almost none measure how narrowly. That second number decides whether automation changed the work or just added a writing assistant to the marketing team.

So this page leads with depth. Among US firms that have adopted AI, 57% use it in three or fewer business functions, 65% hold generative AI to three or fewer tasks, and 66% use it only to augment work people already do. One Census Bureau working paper, April 2026, nationally representative.

Then 44 figures across seven themes. Each carries its publisher, its date, a link you can open, and a label when it is a forecast or when the publisher sells what the number flatters. Compiled 15 September 2026.

Quick answer

A fifth to just over a third of businesses use AI in a business function: 19.8% in the United States (US Census Bureau, May 2026), 20.0% in the European Union (Eurostat, December 2025), 35% of UK businesses with ten or more staff (ONS, July 2026). Almost all use it narrowly. 57% of US adopters run it in three or fewer functions, and the work is writing, document analysis and information search (US Census Bureau, CES Working Paper 26-25, April 2026). Workers report about 11 hours a week saved and 6.4 hours a week managing the tools (Glean, a vendor survey, 2026). Adopting firms show 4% higher labour productivity (Bank for International Settlements, Working Paper 1325, January 2026). As of early 2025, 42% of companies had abandoned most of their AI initiatives before production (S&P Global Market Intelligence, March 2025).

What gets automated

The best data here sits in a US Census Bureau working paper almost nobody quotes. CES WP 26-25, April 2026, runs the 2026 AI supplement to the Business Trends and Outlook Survey and asks adopters what they actually do with it.

The answer is about three things.

57% of adopting US firms use AI in three or fewer business functions. 65% limit generative AI to three or fewer tasks, led by writing, document analysis and information search. And 66% deploy it exclusively to augment existing work, while 2% report an AI-linked decrease in employment. Same paper, same fieldwork window of November 2025 to January 2026.

The typical AI adopter in the United States is a company where a few people write faster.

The three functions are the same three functions

Among adopting firms the leading functions are sales and marketing at 52%, strategy and business development at 45%, and IT at 41%.

Function use and changed work are two different measurements. 23% of firms, or 41% weighted by employment, report AI used inside worker tasks. That gap is the distance between owning a tool and changing a job.

Where the tasks land

The Census Bureau asked workers directly, through the Household Trends and Outlook Pulse Survey fielded March 2026: information search 37%, writing 32%, generating ideas 32%, interpreting, translating or summarising 31%, administrative work 27%, data analysis 21%, tutoring 16%, customer support 12%.

Seven of those eight are reading and writing. The eighth, customer support at 12%, is the only one whose output reaches somebody outside the company.

The UK technology mix agrees: large language models 18%, visual content creation 16%, machine learning for data processing 12%, image processing 6%, robotics 2% (ONS, 20 July 2026). Robotics at 2% ends the factory-floor version of this story.

The counterweight is agent growth. Microsoft reports 15x growth in active agents year over year, and 18x inside large enterprises, from Microsoft 365 telemetry, March 2025 to March 2026. Microsoft sells those agents, and telemetry counts every agent created, including the ones switched off a week later. AI agent vs workflow automation unpacks why the two get counted as one.

Who is automating

Two credible answers to the same question sit fifty points apart.

19.8% of US businesses used AI in any business function in the period ending 3 May 2026, from the Census Bureau's biweekly, nationally representative Business Trends and Outlook Survey. Around 70% of firms are actively using AI across four advanced economies, from a Federal Reserve Bank of Atlanta working paper covering nearly 6,000 CFOs, CEOs and executives, published 24 March 2026.

Both are correct. The Census sample includes the two-person plumbing firm. The Atlanta Fed panel does not. When a stat gets quoted without its population, that is the move to watch for.

Adoption by firm size, in three jurisdictions

Size bands differ by agency and are not comparable across rows. UK figures are BICS wave 159 (fieldwork 5 to 28 June 2026, 38,637 responses). The EU figure covers enterprises with ten or more employees only. All read 15 September 2026.

The firm-size gradient is the most reliable pattern in the dataset, holding in every jurisdiction and every wave. Selling below ten employees? Small business AI adoption statistics is the narrower cut.

Workers answer differently again. 55% of US workers had used AI on the job for at least one of eleven listed tasks, per the Census Bureau's March 2026 worker survey, read in CBS News coverage.

A fifth of firms. More than half of workers. The tools came in from below. Agent-specific counts sit on the companion page: AI agent adoption statistics 2026.

Time saved, and the time it costs back

Two numbers come from the same survey. Only the first gets quoted.

Glean's Work AI Institute surveyed 6,000 full-time digital workers in the US, UK and Australia between December 2025 and January 2026. They said AI automation saves them about 11 hours a week. They also reported 6.4 hours a week feeding context, checking outputs, debugging and switching tools, 37% of their AI-interaction time against 36% spent producing work.

The second number eats most of the first. Glean sells an AI work platform, so both figures are its own survey of its own market. Publishing the unflattering one is why I kept them.

Building agents for small teams, the 6.4-hour figure matches what I see: the work changes shape, from doing the task to specifying it, checking it and running it again. So Gravity pauses customer-facing steps for a person by default.

Government measurement is more sober. Among US workers who used AI at work in the prior week, 31% saved one to two hours and about a quarter saved less than an hour (Census Bureau, August 2026, fielded March 2026, read in CBS News coverage).

An experiment where AI made people slower

METR ran a randomised controlled trial: 16 experienced open-source developers, 246 real issues in their own repositories, mainly Cursor Pro with Claude 3.5 and 3.7 Sonnet. They forecast a 24% speedup.

They took 19% longer. Afterwards they still estimated they had been 20% faster.

Two caveats, both METR's own. These were early-2025 tools, and METR published a redesign note on 24 February 2026 saying it is rerunning the study. Treat the 19% as one measurement of one small cohort.

What survives is the perception gap. People cannot feel their own throughput, which is why self-reported hours saved should never be the only evidence in the room.

Most firms measured nothing at all

The Atlanta Fed asked about 6,000 executives in four countries what AI did to their firm over three years. More than 80% reported no impact on productivity, output or employment. Asked about the next three years, the same firms predicted a 1.4% productivity gain, 0.8% more output and 0.7% lower employment.

The same split shows inside the vendor survey. 13% of workers say their organisation performs significantly better because of AI, against 75% who say AI makes them personally more productive.

Personal speed is easy. Organisational throughput is a different project.

Does automation pay back

One causal estimate anchors this section, and it is smaller than anything a vendor will show you.

The Bank for International Settlements matched survey responses to company accounts for more than 12,000 non-financial firms across the EU and US. Adopters show 4% higher labour productivity on average, through capital deepening. Working Paper 1325, 23 January 2026. The same paper finds no adverse employment effects, the largest gains at medium and large firms, and higher wages at adopters.

Four percent is real and defensible. It is also roughly what a well-executed software rollout has always delivered.

Where the larger gains come from

The biggest credible effects sit at the bottom of the skill distribution. Across 5,179 customer support agents in a staged rollout, published in the Quarterly Journal of Economics, issues resolved per hour rose 14% on average, 34% for novice and low-skilled workers, and by very little for experienced ones.

By function, the Stanford AI Index puts the range at 14 to 15% in customer support, 26% in software development and 50% for marketing output. That is a modelled synthesis by the AI Index, not one study. The Index also estimates US consumer surplus from generative AI at $172 billion a year by early 2026, up from $112 billion. Also a synthesis.

What executives report about their own returns

Self-reported ROI runs far ahead. Wharton Human-AI Research with GBK Collective, 28 October 2025, found 72% of enterprise leaders formally measure ROI, three in four report positive returns, and four in five expect ROI within two to three years.

Google Cloud with National Research Group surveyed 3,466 senior leaders in 24 countries: 52% have deployed AI agents, 39% have launched more than ten, and 74% report ROI within the first year, published 4 September 2025. Google Cloud sells the agents.

Put 74% next to 4%. Both claim to be about returns. One asks people how it feels; the other reads the accounts.

What organisations spend

There is no credible figure for the size of the workflow automation market. Here is the working.

What exists is national accounting. The Bureau of Labor Statistics uses software investment as its proxy for AI and automation adoption, and it is accelerating: 11.1% compound annual growth from 2019 to 2024, against 7.9% from 2007 to 2019 (Monthly Labor Review, 4 May 2026).

Above the software sits the concrete. The Federal Reserve Board puts major technology firm capital expenditure at $412 billion a year, 1.31% of US GDP as of the fourth quarter of 2025 (FEDS Notes, 3 April 2026).

Gartner forecasts worldwide AI spending of $2.59 trillion in 2026, up 47% from $1.76 trillion in 2025, published 19 May 2026. A forecast. Inside it, spending on AI models is forecast to grow 110% and add $6 billion, which Gartner attributes to agents running inside multistep workflows. Also a forecast, and the only line in the spending picture specifically about agent automation.

The market-size numbers are not usable

Published estimates size the 2026 workflow automation market between $27.09 billion and $29.95 billion, which reads like consensus until you look for a published method and find none. That range comes from four commercial research firms, none of which publishes a method, which is why none is linked here. Run the same exercise on robotic process automation and the answers land between $6.08 billion and $35.27 billion.

A range that wide is not an estimate.

So none of it appears in the 44. I dropped those numbers because there is nothing behind them to check. For prices that can be checked line by line there is the AI agent pricing index, and Gravity's own pricing sits on one page.

What breaks

The most quoted failure statistic in this category is one survey, and it is now eighteen months old.

S&P Global Market Intelligence's Voice of the Enterprise study, more than 1,000 respondents in North America and Europe, found 42% of companies had abandoned the majority of their AI initiatives before production, up from 17% a year earlier. The same survey put the average share of AI proof-of-concepts scrapped before production at 46%, naming cost, data privacy and security as the top obstacles. Reported 14 March 2025, so both are as of early 2025.

One survey. Two findings. Not two independent studies agreeing, which is how the pair usually travels.

Everything about 2027 is a forecast

Gartner predicts 40% of enterprises will demote or decommission autonomous agents by 2027 over governance gaps found after production incidents, published 26 May 2026. A forecast, from an analyst firm, about a year that has not happened.

I include it because you will meet it in every deck this quarter, and you should know what kind of number it is before repeating it.

Counted harms, and the ones only predicted

Documented AI incidents rose to 362 in 2025, from 233 in 2024, per the AI Incident Database compilation in the 2026 AI Index. A count of recorded harms undercounts by construction.

Hallucination rates across 26 leading models range from 22% to 94% on the Index's benchmark. The top of that range belongs to a model somebody is running in production this morning.

And the automation nobody approved is bigger than the automation anyone tracks. 66% of office professionals said they had used AI tools at work they believed were not permitted, in a PagerDuty survey of 1,250 professionals at organisations above $500 million in revenue, 11 June 2026. PagerDuty sells incident response. The figure stayed because it matches the shape of the Census worker data.

What it is doing to jobs

The clearest labour signal in the dataset is narrow, recent and about young people.

Stanford's Digital Economy Lab, on ADP payroll records from November 2022 to June 2026, measures an AI employment gap for workers aged 22 to 25 in highly AI-exposed occupations of 19% as of June 2026, up from 15% in July 2025. In levels: employment for that age group fell about 11% in highly exposed occupations while growing about 10% in less-exposed ones. Experienced workers show no comparable gap.

The lab's own caveat travels with the finding: this is not widespread, economy-wide job displacement. Keep that sentence attached to the 19% every time it gets quoted.

Employers describe less movement. Among UK businesses using AI, about half report no change in headcount and under 7% reported reductions (ONS, 20 July 2026). In the US, 2% of firms reported an AI-linked employment decrease.

Hiring demand shows the other side. Indeed Hiring Lab recorded AI mentions in 4.2% of US job postings in December 2025, the highest on its index, with AI-mentioning postings 134% above February 2020 while total postings were 6% above; marketing rose from 8.4% to 14.9% and HR from 4.4% to 8.8%.

The people are already fluent. Work-related generative AI use among US adults went from 33% to 41% between August 2024 and November 2025, with non-work use rising from 36% to 50% (Federal Reserve Board, Real-Time Population Survey, 3 April 2026).

What is thin, and why

Spending is the weakest theme here. The sources are thin, so the section stays thin.

Three sources I wanted are missing because their sites refused an automated fetch and I have not read the originals:

• McKinsey, "The state of AI in 2026: On the road to ROI", August 2026. Trade coverage attributes an EBIT-impact share and a high-performer share to it. Unread here, so uncited here.
• OECD announcement on AI use by individuals and firms, 28 January 2026, carrying a firm-level adoption series. Unread.
• ServiceNow Enterprise AI Maturity Index 2026. A vendor index with a large sample. Unread.

One more absence. The claim that 95% of AI pilots fail, usually attributed to an MIT project, is not here. I could not trace it to an MIT-hosted original, and the mirrors carrying it disagree about the method.

Fourteen figures were dropped or held in total, including four shadow-AI vendor surveys with four incompatible answers, of which one survived. 44 of the figures that survived verification appear here. No-code automation statistics covers the adjacent tooling numbers that did.

Every figure on one table

44 figures, one row per highlighted number on this page; the # links to the sentence it sits in and the publisher links to the original document where the link is recorded. Confidence is the CSV label as written: primary means the publisher collected the data; secondary means a credible re-report or model; vendor-interested means the publisher sells something the figure flatters; forecast means a prediction, not a measurement; unverified means the original blocked our read and the figure comes from a named re-report. The same rows are in data.csv (CC BY 4.0).

Use this data

Take it. The point of building the table was to have something worth citing.

All 44 rows are published under a Creative Commons Attribution 4.0 licence. Republish them, chart them, use them commercially. Credit this page and you are covered.

The machine-readable version carries every field: download the CSV, with columns for id, figure, context, publisher, date, method, source URL and confidence. Row ids match the anchors here, so #stat-B1 and CSV row B1 are the same number. A GitHub mirror sits next to the pricing dataset.

FAQ

Sources