Agentic AI Statistics 2026: 150 Sourced Data Points

How to Read These Numbers

"AI agent" has no fixed definition. McKinsey's survey uses one of the stricter versions: systems that act in the real world and can plan and carry out several steps of a workflow on their own. Other surveys leave the term to the respondent, which lets a chatbot count. Before comparing two figures, check three things.

A Federal Reserve Board staff note made the same point in April 2026. It cited research that found estimates of work-related AI adoption ranging from about 5% to 40% across 16 US surveys, and concluded that the right estimate depends on the question being asked.

Official Statistics on AI Use

The US Census Bureau and Eurostat do not yet ask about agents specifically, but they give the most reliable baseline for AI use of any kind. In the US, between 18% and 23% of firms use AI, depending on the survey and period, with far higher rates at large firms and in information and finance. In the EU, 20% of enterprises with ten or more staff used AI in 2025, from 42% in Denmark to 5.2% in Romania.

Agent Adoption Surveys

Surveys that ask about agents give answers from 19% to 79%. Set side by side with their samples, the spread mostly reflects the question. "Being adopted" covers far more than "scaling", and large organisations are ahead: McKinsey's scaling rate rose from 27% to 40% at companies with more than $1 billion in revenue, while smaller ones stayed at 22%.

For a consumer view of what people will let agents do, see our readout of a 5,067-person survey on agent permissions.

What Agents Are Used For

Task-level data is scarce, but two sources go further than a general adoption question. The US Census Bureau's 2026 AI supplement asked firms which business functions and worker tasks use AI. Anthropic's Economic Index classifies samples of real Claude conversations, with privacy protections, by task and by how much is delegated. The Census data show use concentrated in a few functions and mostly augmenting human work. Anthropic's data show automation rising.

The Anthropic figures describe one company's traffic, not the economy. They are still among the few published measurements of what delegated AI work looks like task by task.

Multi-Agent Orchestration and Protocols

Coordinating several agents on one workflow is still early, but it is growing. In KPMG's quarterly survey of large US firms, the share orchestrating multiple agents across workflows doubled from 9% to 18% in one quarter. The research record is sobering: a UC Berkeley study of seven open-source multi-agent frameworks found failure rates of 41% to 86.7%, and sorted the causes into 14 failure modes.

The plumbing is spreading faster than the practice. The download counts below are the only figures on this page that we collected ourselves, from the public npm and PyPI statistics on September 18, 2026. The npm window is August 18 to September 16, 2026, compared with the same dates in 2025; PyPI figures are 30-day totals.

Capability and Reliability

METR measures how long a task, in skilled-human time, an AI agent can complete. The 50% figure, where the agent succeeds half the time, is the one usually quoted, and it has been doubling roughly every four months for models released since 2023. The 80% figure matters more for unattended work, and for each of the four models in the table it is roughly four to ten times shorter. METR has not yet published horizons for newer models such as Claude Opus 4.7 and GPT-5.5.

Controlled studies point the same way. The best agent in Carnegie Mellon's TheAgentCompany, a simulated software firm, completed 30% of its tasks on its own. METR's 2025 trial found experienced developers took 19% longer when allowed AI tools, although they believed they had been sped up. METR's 2026 follow-up suggests that has changed, but METR itself calls the new data weak evidence. For how often coding agents game their own tests, see our census of published reward-hacking rates.

Value, ROI and Failure Rates

Returns are real but concentrated. In McKinsey's 2026 survey, 80% of respondents say AI has made them personally more productive, but only 37% report a positive contribution to their organisation's earnings (EBIT), and about 6% qualify as high performers. BCG's study of more than 1,250 firms finds 5% achieving value at scale and 60% achieving no material value.

Two figures need care. PagerDuty's 171% is the average ROI executives expect from agentic AI, not a return anyone measured. And Gartner's forecast that over 40% of agentic AI projects will be cancelled by the end of 2027 is a prediction, not a count of failures.

Cost and Governance

Cost is now a constraint in its own right. About one in five of McKinsey's respondents say AI operating costs, including tokens, have limited their use of AI. KPMG finds only 26% of large US firms can see in real time what running AI costs them. On governance, only 21% of companies planning agents told Deloitte they have a mature governance model.

Unit costs vary more than most budgets assume. Our index of frontier model API prices and our look at how the agent harness changes the cost of the same model show where that spread comes from.

Security Incidents and Breach Costs

IBM's 2026 breach study, covering 602 organisations, puts the average breach at $4.99 million and AI-enabled breaches at $6 million. Figures on agents specifically come from security vendors' own surveys, so read them as indicative. HiddenLayer found one in eight reported AI breaches linked to agentic systems. Gravitee found 88% of organisations had confirmed or suspected agent incidents, while only 14.4% put every agent live with full security approval.

Spending and Market Size Forecasts

Headline market figures look far apart but mostly measure different things. Gartner's $2.59 trillion for 2026 and IDC's $1.3 trillion for 2029 cover all AI spending, including infrastructure. The "AI agents market" estimates are much smaller and closer than they appear: two research firms both put 2025 at about $7.8 billion to $7.9 billion. Their headlines differ ($52.6 billion against $294.7 billion) mainly because one forecasts to 2030 and the other to 2035. Compare end years before comparing numbers.

Work and Jobs

Half of US employees now use AI at work at least occasionally, and 15% use it daily, according to Gallup's May 2026 data. The job effects measured so far are small at the level of firms: 2% of AI-using US firms report AI-related employment cuts. They are sharper for young workers in exposed jobs, with employment of US software developers aged 22 to 25 down nearly 20% since 2024. Expectations run ahead of outcomes: 39% of McKinsey's respondents expect AI-related headcount declines next year.

Vendor-Reported Figures

Company disclosures are useful for scale, but they follow each company's own definitions, and those change. Salesforce's Agentforce revenue figure, for example, has included Slackbot and its Headless 360 platform since the second quarter of fiscal 2027, so it is not directly comparable with earlier quarters. For Microsoft's Copilot numbers in context, see our analysis of Microsoft's FY26 Q4 results.

Widely Quoted Figures, Checked

A handful of agentic AI statistics circulate widely with the wrong source, the wrong wording or no source at all. These are the ones we see most often, checked against the primary publication.

Methodology

All 150 rows are available as a CSV download, with each figure's source, date and type. The publications behind them are listed in the Sources section at the end of this article.

Sources

Every figure on this page comes from one of the 44 sources below, grouped by who published them.