Agentic AI Adoption Statistics 2026-2027
By Axis Intelligence Research
Co-author: Sarah Mitchell | Last updated: November 13, 2026 | License: CC BY 4.0
Published agentic AI adoption rates run from 17% to 59%, a 3.47× spread, because each study counts a different population: Gartner’s organization-level deployment count sits at the bottom and Stack Overflow’s surveyed workers at the top. Inside OpenAI’s enterprise telemetry, Legal’s weekly active Codex users grew 108× from February to June 2026 while Engineering’s grew 5×.
Quick Answer
There is no single agentic AI adoption rate. Depending on whether a study counts organizations, executives, builders or individual workers, the share using agents in 2026 lands between 17% and 59%. Adoption is also spreading past engineering: OpenAI’s enterprise data shows Legal, Sales and Recruiting catching up with engineering, KPMG’s technology survey shows Finance, Operations and Risk, Legal and Compliance doing the same, and McKinsey finds multi-function agent scaling in only four of twelve industries.
Key Findings
- Legal’s weekly active enterprise Codex users grew 108× between February and June 2026 against 5× for Engineering, an Axis Function Spillover Ratio (AFSR) of 21.6 (OpenAI).
- Finance agent deployment among US technology leaders rose from 10% in Q2 to 60% in Q3 2026, a survey-form AFSR of 5.1 against Technology/IT (KPMG US).
- Four of twelve industries report agent scaling at 10% or more in six or more of eleven business functions: technology, media and telecom, financial institutions and advanced manufacturing (McKinsey, 2026).
- Frontier firms produced 8.3× the output tokens per active user of typical firms in June 2026, up from 2.6× in January (OpenAI).
- Stack Overflow’s April 2026 Pulse survey found 37% of respondents using agents daily at work, up from 14% in the 2025 survey.
How many organizations use agentic AI?
The honest answer depends on the denominator. The table lines up eight published measurements, each with its own population, definition and date. Seven are agent-specific; the Census row counts any AI use and is included as the floor for the whole market.
| Population counted | What is measured | Rate | Period | Source |
|---|---|---|---|---|
| All US businesses | Any AI use in a business function (not agent-specific) | 19.8% | May 3, 2026 | U.S. Census Bureau |
| Organizations | AI agents deployed (more than 60% expect to within two years) | 17% | 2026 | Gartner |
| Smaller organizations | Scaling agents in at least one function | 22% | 2026 | McKinsey |
| Large organizations | Scaling agents in at least one function | 40% (27% in 2025) | 2026 | McKinsey |
| Executives at gen-AI-using enterprises with $10M+ revenue | Agents deployed in production | 52% | Fieldwork Apr 18–Jun 3, 2025 | Google Cloud |
| Agent builders | Agents in production | 57.3% (30.4% still in development) | Nov–Dec 2025 | LangChain |
| Surveyed workers | Using agents at work, any frequency | 59% (31% in 2025) | April 2026 | Stack Overflow |
| Organizations | Significant employee adoption of agents | 34% (25% in Q1) | Q3 2026 | KPMG |
The sources are Gartner’s AI Hype Cycle commentary, McKinsey’s State of AI 2026 report (1,719 respondents, 97 nations), Google Cloud’s ROI of AI 2025 study (3,466 senior leaders, 24 countries), LangChain’s State of Agent Engineering (1,340 respondents), Stack Overflow’s agents Pulse survey (about 1,100 respondents), KPMG’s Global AI Pulse (2,131 respondents, 20 countries) and the Census Bureau’s business survey analysis.
Why the same year produces 17% and 59%
Three different units are being counted. Organization-level studies (Gartner, McKinsey, KPMG) ask whether a company has crossed a threshold such as “deployed” or “scaling.” Executive and builder samples (Google Cloud, LangChain) draw from respondents already close to agents, so production rates run high. Worker-level surveys (Stack Overflow) ask whether an individual has used an agent at all, which a single engineer can answer yes to long before the company has a program. Dividing the highest agent-specific rate (59%) by the lowest (17%) gives 3.47×, a spread created by the question asked rather than by the economy.
How the size gap shows up
McKinsey’s large-company rate rose from 27% to 40% in a year while the smaller-company rate stayed at 22%. Other samples disagree on direction. LangChain’s builders reported agents in production at 50% for companies under 100 employees and 67% at 10,000 or more. Google Cloud’s executives, all at enterprises with 100 or more employees, reported 49% at 100–499 employees, 56% at 500–999 and 52% at 1,000 or more. Firm-size effects change sign depending on whether the study samples all firms or only the ones already building. Our companion pages cover the size gap in corporate AI adoption and the pilot-to-production conversion in agentic AI statistics.
Analyst note (Sarah Mitchell): Quote a single adoption rate only with its denominator attached. A board deck that says “52% use agents” is describing Google Cloud’s executives at gen-AI-using enterprises, not the market. The Census row is the sanity check: 19.8% of all US businesses used any AI at all in May 2026, so a higher agent rate cannot describe all US businesses.
Which business functions adopt AI agents first?
Engineering was first, and the more useful question in 2026 is how fast other functions are catching up. Three kinds of evidence answer it: telemetry on growth, surveys on deployment, and McKinsey’s survey on how far scaling has gone in each function.
The Axis Function Spillover Ratio (AFSR)
The Axis Function Spillover Ratio (AFSR) divides a function’s growth multiple by the growth multiple of the engineering baseline, inside one source and one window. An AFSR of 1.0 means a function is growing at the baseline’s pace; above 1.0 means agent use is spreading beyond engineering faster than inside it. It has two forms:
- Telemetry form: growth in weekly active enterprise Codex users from February to June 2026, baseline Engineering.
- Survey form: growth in the share of organizations with agents deployed in a function from Q2 to Q3 2026, baseline Technology/IT.
Telemetry form (OpenAI).
| Function | Growth in weekly active enterprise Codex users since February | AFSR (Engineering = 1.0) |
|---|---|---|
| Legal | 108× | 21.6 |
| Sales | 41× | 8.2 |
| Recruiting | 41× | 8.2 |
| Marketing | 26× | 5.2 |
| Engineering | 5× | 1.0 |
Inputs come from OpenAI’s “How enterprises put AI to work” (August 12, 2026); each AFSR is the function’s multiple divided by 5. Legal’s 108× divided by 5 gives 21.6.
Survey form (KPMG US Technology, n = 100 leaders at $1B+ revenue). The share of organizations with agents deployed, by function, from KPMG’s Q3 2026 AI Pulse for Technology:
| Function | Q2 2026 | Q3 2026 | Q3 ÷ Q2 | AFSR (Technology/IT = 1.0) |
|---|---|---|---|---|
| Technology/IT | 85% | 100% | 1.18 | 1.0 |
| Operations | 48% | 97% | 2.02 | 1.72 |
| Risk, Legal and Compliance | 31% | 63% | 2.03 | 1.73 |
| Finance | 10% | 60% | 6.0 | 5.1 |
Finance’s survey-form AFSR is (60 ÷ 10) ÷ (100 ÷ 85) = 5.1. Two independent measurements, one built on usage telemetry and one on executive survey answers, point the same way: Legal, Finance, Sales and Operations are gaining on the engineering and IT baselines.
Reading the ratio. AFSR measures the speed of spread, not the level of use. Legal started from a far smaller base than Engineering, so a high AFSR says Legal is where new agent users are appearing fastest, not that Legal uses agents more than engineers do. In the survey form, Technology/IT reached 100% in Q3, so its own multiple cannot exceed 1.18 and part of every survey AFSR is catch-up to a saturated baseline. The ratio is most useful as a watch metric: when a function’s AFSR falls toward 1.0, spillover there has slowed to the baseline’s pace.
How far scaling has gone: McKinsey by function
McKinsey asked respondents whether agents had reached the scaling phase in each function. Levels are lower than the growth ratios suggest, which is what a catch-up from a small base looks like. The last column divides each function’s rate by software engineering’s 10%.
| Function | All respondents | AI high performers | All others | High performers ÷ others | Level vs software engineering |
|---|---|---|---|---|---|
| IT | 12% | 28% | 10% | 2.8 | 1.2 |
| Knowledge management | 11% | 32% | 10% | 3.2 | 1.1 |
| Software engineering | 10% | 30% | 9% | 3.33 | 1.0 |
| Service operations | 9% | 26% | 8% | 3.25 | 0.9 |
| Product and/or service development | 9% | 24% | 8% | 3.0 | 0.9 |
| Marketing and sales | 8% | 26% | 7% | 3.71 | 0.8 |
| Human resources | 7% | 16% | 6% | 2.67 | 0.7 |
| Risk, legal and compliance | 5% | 15% | 4% | 3.75 | 0.5 |
| Strategy and corporate finance | 5% | 17% | 4% | 4.25 | 0.5 |
| Supply chain management | 5% | 4% | 5% | 0.8 | 0.5 |
| Manufacturing | 3% | 7% | 3% | 2.33 | 0.3 |
“AI high performers” are McKinsey’s respondents who attribute 5% or more of EBIT to AI and report significant value (about 6% of respondents; n = 92 against 1,429 for all others). McKinsey labels two functions slightly differently across exhibits, so rows pair the closest labels. Two readings stand out. Risk, legal and compliance scales at 5%, half of software engineering’s level, while OpenAI’s telemetry shows Legal growing fastest: a low level and a high growth multiple belong together. And high performers lead most in strategy and corporate finance (4.25×), risk, legal and compliance (3.75×) and marketing and sales (3.71×), but not at all in supply chain management (4% against 5%).
What other sources say about function
Function rankings disagree because the populations differ.
- OpenAI’s own workforce shows median output-token growth since November 2025 of 56× in Research, 32× in Customer Support, 27× in Engineering and 13× in Legal, per OpenAI’s “How agents are transforming work”. Business-function workers spent over one-fourth of their Codex work on engineering or coding tasks.
- US technical leaders in Anthropic’s State of AI Agents report (over 500 respondents, late 2025) rated data analysis and report generation among the most impactful tasks at 60% and internal process automation at 48%. They expected 2026 impact to be largest in software development (57%), customer service (55%), marketing and sales (46%) and supply chain, logistics and operations (44%).
- Agent builders in the LangChain survey named customer service (26.5%), research and data analysis (24.4%) and internal workflow automation (18%) as their top use cases.
- Google Cloud’s executives (2025 fieldwork) named customer service and experience as the top agent use case in financial services (57%) and retail and CPG (47%), security operations and cybersecurity in media and entertainment (47%) and telecom (47%), and tech support in healthcare and life sciences (49%).
Analyst note (Sarah Mitchell): Builders ship customer service agents, technical leaders rate analysis agents as most impactful, executives in media and telecom point at security operations, and telemetry shows Legal and Sales accelerating. Those are different questions (what teams build, what leaders value, where seats are growing), and each is a valid answer to “which function adopts first.”
Which industries use AI agents the most?
McKinsey’s 2026 survey is the richest industry cut available: the share of respondents reporting that agents have reached the scaling phase, for each of 11 functions in 12 industries. The table summarizes the matrix. “Functions at 10% or more” counts how many of the 11 functions clear that bar, which is a measure of how broad an industry’s agent footprint is.
| Industry | Respondents (n) | Highest-scaling function | Functions at 10% or more (of 11) |
|---|---|---|---|
| Technology | 208 | Software engineering (31%) | 7 |
| Media and telecom | 79 | Knowledge management and service operations (21% each) | 8 |
| Financial institutions | 158 | IT and knowledge management (17% each) | 7 |
| Advanced manufacturing | 105 | IT (15%) | 6 |
| Consumer goods and retail | 128 | Marketing and sales (15%) | 3 |
| Healthcare payers and providers | 52 | Human resources (14%) | 3 |
| Professional services | 230 | Knowledge management (14%) | 1 |
| Insurance | 44 | Service operations (14%) | 1 |
| Pharma and medical products | 64 | IT (12%) | 1 |
| Travel and logistics | 61 | IT (17%) | 1 |
| Engineering and construction | 88 | Software engineering (8%) | 0 |
| Energy and materials | 177 | IT, software engineering, marketing and sales, manufacturing (7% each) | 0 |
Four industries (technology, media and telecom, financial institutions, advanced manufacturing) report scaling at 10% or more in six to eight functions. The other eight clear that bar in three functions or fewer. Technology’s 31% in software engineering is 3.1× the all-industry figure of 10%. Sector shape follows the work: consumer goods and retail peak in marketing and sales at 15%, while advanced manufacturing scales in supply chain and inventory management (14%) and in manufacturing (14%), against all-industry rates of 5% and 3%. Insurance (n = 44) and media and telecom (n = 79) are small cells, so treat single values there as directional.
Three more cuts of the same question:
- Stack Overflow, daily agent use at work (April 2026): Fintech 55%, Media and Advertising 50%, Software Development 44%, Transportation and Supply Chain 41%, Manufacturing 38%, Healthcare 36%, Government 35%, Internet, Telecom and Information 31%, Retail 31%, Banking and Financial Services 31%, Computer Systems Design 28%, Higher Education 26%, Energy 17%. Fintech’s 55% is 3.24× Energy’s 17%, and fintech sits 24 points above banking and financial services at 31%, so “finance” alone is too coarse a label for agent adoption.
- Google Cloud, agents deployed in production (2025 fieldwork): healthcare and life sciences at 44%, retail and CPG at 51%, financial services at 53%, media and entertainment at 54%, public sector at 55%, and telecom and manufacturing and automotive at 56% each.
- Census Bureau, any AI use (May 3, 2026): 39.7% of Information firms and 33.9% of Finance and Insurance firms, against about 14% in Retail Trade. These are all-AI rates and should not be read as agent rates.
Energy is the quiet industry in two surveys at once: 17% daily use in Stack Overflow’s data and no function above 7% in McKinsey’s.
Which regions lead agentic AI adoption?
Region rankings flip depending on the metric. KPMG’s Global AI Pulse measures the share of organizations scaling AI or beyond, by region; Google Cloud’s study measures agents deployed in production.
| Region | KPMG: scaling AI or beyond, Q1 2026 | KPMG: Q3 2026 |
|---|---|---|
| Americas | 56% | 64% |
| APAC | 43% | 61% |
| EMEA | 40% | 56% |
| Region | Google Cloud: agents in production (2025 fieldwork) | Respondents (n) |
|---|---|---|
| North America | 46% | 1,247 |
| Europe | 47% | 706 |
| LATAM | 56% | 701 |
| MEA | 59% | 110 |
| JAPAC | 64% | 702 |
The Americas lead KPMG’s AI-scaling measure at 64%, yet North America sits last in Google Cloud’s agent measure at 46%. Asia-Pacific is the common thread: it gained 18 points in two quarters in KPMG’s data and tops Google Cloud’s list at 64%, 1.39× North America’s rate. Note that KPMG measures all AI and Google Cloud’s fieldwork predates 2026, and MEA’s sample is 110. The gap between KPMG’s leading and trailing regions narrowed from 16 percentage points to 8, even as OpenAI’s telemetry shows firms diverging: frontier firms (the top 10% of OpenAI’s enterprise customers) produced 8.3× the output tokens per active user of typical firms (45th to 55th percentile) in June, up from 2.6× in January, a 3.19× widening in five months. Regions are converging on AI scaling while firms diverge on how heavily they use agents. The measures differ, so we show them side by side rather than combining them.
Who uses AI agents at work?
Daily agent use differs sharply by role. In Stack Overflow’s April 2026 Pulse survey, the share using agents daily was:
| Role | Daily agent use |
|---|---|
| Architects | 52% |
| Senior executives | 50% |
| Full-stack developers | 40% |
| Students | 38% |
| Academic researchers | 28% |
Architects are 1.86× as likely as academic researchers to use agents daily. Across all respondents, daily use rose from 14% to 37% in a year and any-frequency use from 31% to 59%, so daily use grew 2.64× and any use 1.90×: heavy use is deepening faster than first use is spreading.
Outside software, OpenAI reports more than 5 million weekly active Codex users, up more than 6× since the desktop app launched in February, with knowledge workers at about 20% of users and growing more than three times as fast as other users. Among non-developers, OpenAI’s agent-work analysis measured growth since August 2025 of 137× for individual users and 189× for organizational users.
How deep does agent usage go?
Counting users says little about workload. Five published signals describe depth.
How much of AI output runs through agents?
Among OpenAI’s enterprise customers, Codex accounted for 64% of combined Codex and ChatGPT output tokens as of June 2026. Inside OpenAI, Codex generated 99.8% of weekly output tokens and 99% of engineers’ output tokens.
How long do agents work on a single task?
By May 2026, 80.6% of sampled individual Codex users had made at least one request estimated to represent more than 30 minutes of human work, 70.2% at least one over an hour, and 25.6% at least one over eight hours. The thresholds are nested, so 25.6 ÷ 80.6 means 31.8% of users who sent a 30-minute request also sent an eight-hour one.
At the extreme, 99th-percentile users generated more than 60 hours of Codex agent turns per day across parallel agents. OpenAI estimated task length with an LLM-as-judge on a 0.1% random user sample, so read the thresholds as directional.
How many hours a day do developers spend with AI tools?
In Stack Overflow’s 2026 Developer Survey, 31% of respondents use AI tools four or more hours a day, 24% two to four hours and 25% one to two hours (chart base 10,926). Of those who use AI coding assistants or agents, 73% use them daily, and the survey names Claude Code (66%) and GitHub Copilot (59%) as the agents of choice.
How closely are agents supervised?
In Stack Overflow’s Pulse survey, 63% of respondents rarely or never allow full autopilot, 60% block unapproved system changes, and 69% run a single agent rather than several. Cost as a barrier fell from 53% agreeing in 2025 to 38% in 2026.
Are multi-agent systems real yet?
KPMG’s US technology sample shows multi-agent systems in development or implementation rising from 5% in Q2 to 23% in Q3 2026, a 4.6× jump, while overall agent deployment among those leaders rose from 67% to 93% and piloting fell from 31% to 6%. Anthropic’s late-2025 survey found 57% of US technical leaders deploying agents for multi-stage workflows but only 16% for cross-functional processes, a 3.56× difference between depth inside one process and breadth across several. LangChain’s builders reported observability at 89%, offline evals at 52.4% and online evals at 37.3%.
Analyst note (Sarah Mitchell): The token and task-horizon numbers describe the people who already use agents most. The supervision numbers describe the typical worker. Both are true at once: a small group runs agents for hours in parallel while most workers still keep a hand on the wheel.
How much AI value comes from agents?
BCG’s 2026 Applied AI Index (more than 1,300 executives) estimates agents generated 22% of AI value in 2026, up from 17% in 2025, and projects 39% by 2030. BCG’s agentic value formula also shows the spread between companies: 44% of future-built companies realize agentic value against 2% of laggards, a 22× ratio.
McKinsey’s high performers scale agents more widely: 40% scale software coding agents against 20% of other respondents, and 40% scale other AI agents against 15%. About 20% of McKinsey respondents say AI operating costs, including tokens, constrained their AI use. On labor, 14% of respondents at AI-using organizations reported that AI contributed to a workforce decline in the past year, against 32% who had expected reductions in the 2025 survey.
Controls trail ambition. BCG found 42% of respondents expect agents to act autonomously by 2030 while 5% have full controls in place today, an 8.4× ratio of expectation to readiness. KPMG’s Global AI Pulse points the same way: 86% of organizations with established ROI have a formal harness layer, against 31% of those still experimenting and 55% overall. Among US technology leaders in KPMG’s survey, the share identifying high-risk use cases where autonomous agent decisions are not allowed rose from 22% to 67% in a quarter.
What will agentic AI adoption look like in 2027?
Published forecasts are narrow. The numbers below are the forward-looking figures that primary sources have released, with the date each was made; we have not added a forecast of our own.
- End of 2026: Gartner predicted in August 2025 that 40% of enterprise applications will be integrated with task-specific AI agents, up from less than 5% in 2025.
- 2027: Gartner predicts over 40% of agentic AI projects will be canceled by the end of 2027, and, in a May 2026 release, that 40% of enterprises will demote or decommission autonomous agents because of governance gaps found after production incidents.
- 2028: Gartner predicts 33% of enterprise software applications will include agentic AI, up from less than 1% in 2024, and at least 15% of day-to-day work decisions will be made autonomously, up from 0% in 2024.
- 2030: BCG projects agents will account for 39% of AI value.
- Next 12 months: in Gartner’s Q2 2026 survey of over 2,000 CIOs and technology executives, 71% expect to deploy agentic AI within 12 months while 21% have governance capabilities in place, a 3.38× readiness gap. BCG’s equivalent ratio is 8.4×. McKinsey adds that 39% of respondents expect AI-related headcount declines in the coming year and 32% decided against buying software because coding agents could build it.
- Complexity and cost: 81% of Anthropic’s surveyed technical leaders planned more complex agent use cases in 2026, including 39% for multi-step and 29% for cross-functional work, and the share of Stack Overflow respondents agreeing cost is a barrier fell from 53% to 38%.
Four readings to track when the next surveys land: whether the 3.47× spread between published agent adoption rates narrows as sources converge on common definitions; whether the frontier gap in OpenAI’s token data keeps widening beyond 8.3×; whether AFSR, in both forms, stays above 1.0 for Legal, Sales, Finance and Operations, which would confirm that agent use is still spreading beyond engineering and IT; and whether the number of industries with six or more functions scaling at 10% grows beyond four.
Methodology
Every figure on this page comes from a primary publication, retrieved or read on November 13, 2026, with its URL logged in the dataset. Rates are shown as the source reports them. We do not average surveys with telemetry, we do not pool populations, and we do not round published values. Where a source reports a floor (“more than 5 million,” “over one-fourth,” “more than 60 hours”), the dataset stores the stated floor with a unit that says so. McKinsey’s industry and function values were read from the exhibits in the published report.
Comparability rules. Firm-level surveys, builder samples, worker surveys and vendor telemetry measure different things. We place them side by side and label the population. Rows from Census and KPMG’s regional table measure all AI rather than agents and are labeled that way. McKinsey counts respondents whose organizations have reached the scaling phase; KPMG counts organizations with agents deployed in a function; the two are not interchangeable.
Axis calculations. Every derived figure shows its formula in the text and in the method_note column of the dataset:
- AFSR, telemetry form = function growth multiple ÷ Engineering growth multiple (5).
- AFSR, survey form = (function Q3 ÷ Q2 deployment share) ÷ (Technology/IT Q3 ÷ Q2 share, 100 ÷ 85).
- McKinsey function level = function scaling rate ÷ software engineering scaling rate (10).
- McKinsey high-performer ratio = high-performer rate ÷ all-other rate.
- Functions at 10% or more = count of the 11 functions with a scaling rate of 10% or higher in an industry.
- Frontier gap widening = 8.3 ÷ 2.6 = 3.19.
- Leader-to-laggard agentic value ratio = 44 ÷ 2 = 22.
- Expectation-to-controls ratio = 42 ÷ 5 = 8.4; expectation-to-governance ratio = 71 ÷ 21 = 3.38.
- Agent adoption spread = 59 ÷ 17 = 3.47.
- Share of users with a 30-minute request who also made an eight-hour request = 25.6 ÷ 80.6 = 31.8%.
Selection. A source is included only if its text or exhibit states the figure, the population and the period.
About this dataset
The fact table behind this page is agentic-ai-adoption-statistics.csv (433 rows, UTF-8, CC BY 4.0). Each row carries section, metric, segment, period, value, unit, source_org, source_document, source_published, source_url, retrieved_date, is_primary, axis_calculated and method_note. Every number and chart datapoint on this page is a row in the file, including the full 11-function by 12-industry McKinsey matrix. Reuse it with attribution to Axis Intelligence Research.
Frequently asked questions
Why do agentic AI adoption rates range from 17% to 59%?
Each study counts a different population. Gartner’s 17% is organizations that have deployed agents; Stack Overflow’s 59% is individual workers using agents at any frequency. A worker can adopt long before a company scales, so worker-level rates run higher than organization-level ones.
Which function is adopting agents fastest outside engineering?
Legal in usage telemetry and Finance in executive surveys. Legal’s weekly active enterprise Codex users grew 108× from February to June 2026 against 5× for Engineering, an AFSR of 21.6. In KPMG’s technology survey, Finance deployment rose from 10% to 60% in a quarter, a survey-form AFSR of 5.1.
Which industries have agents scaled across many functions at once?
Four of twelve in McKinsey’s 2026 survey: media and telecom (8 of 11 functions at 10% or more), technology (7), financial institutions (7) and advanced manufacturing (6). Eight industries have three or fewer, and energy and materials and engineering and construction have none.
Do AI high performers scale agents in different functions than other companies?
Mostly the same functions at a much higher rate. High performers scale agents 4.25× as often as others in strategy and corporate finance, 3.75× in risk, legal and compliance and 3.71× in marketing and sales. Supply chain management is the exception: 4% for high performers against 5% for others.
What share of enterprise AI output runs through agents?
Codex made up 64% of combined Codex and ChatGPT output tokens among OpenAI’s enterprise customers as of June 2026. Frontier firms produced 8.3× the output tokens per active user of typical firms, up from 2.6× in January.
How long do AI agents run on a single task?
By May 2026, 70.2% of sampled Codex users had sent at least one request estimated at over an hour of human work, and 25.6% one over eight hours. Ninety-ninth-percentile users generated more than 60 agent-hours per day across parallel agents.
Do developers let agents run without approval?
Mostly not. In Stack Overflow’s April 2026 Pulse survey, 63% rarely or never allow full autopilot and 60% block unapproved system changes. Single-agent setups account for 69% of respondents.
Does company size predict agent adoption?
It depends on the sample. McKinsey found 40% of large organizations and 22% of smaller ones scaling agents in a function. LangChain’s builders reported 67% at 10,000+ employees and 50% under 100. Google Cloud’s executives reported 49% at 100–499 employees, 56% at 500–999 and 52% at 1,000 or more.
Which region leads agentic AI adoption?
It depends on the metric. The Americas lead KPMG’s AI-scaling measure at 64% in Q3 2026, while JAPAC leads Google Cloud’s 2025 agents-in-production measure at 64% and North America trails at 46%. Asia-Pacific is the only region near the top of both.
What does “in production” mean in agent surveys?
It varies. Google Cloud asked how many agents were deployed in production across the organization, LangChain asked builders about agents in production, McKinsey asked about the scaling phase of agent use, Gartner counted organizations that have deployed agents, and KPMG asked about significant employee adoption. Compare rates only within one definition.