A program of EffectiveSC
How We Use AISources checked October 24 to 29, 2026
Where things stand / What people use AI for
What people actually use AI for at work and at home, how heavy and light users differ and how far each gets, and which uses pay off.
In surveys of workers and in the logs of the biggest chat products, writing and finding information come first. Most of the time people use AI to help with part of a task, not to hand over the whole of it.
51%
In May 2026, among US employees who use AI at work, the most common uses were writing and editing (51%), search or research (49%) and general assistance or problem-solving (39%).
Who was asked: employees who use AI at work
People could name more than one use.
about 40%
Writing is about 40% of work-related messages on consumer ChatGPT, and about two thirds of all writing requests ask it to rework text the user supplied (edit, critique, translate or summarize).
almost 60%
In one week of February 2026, looking up information and refining content made up almost 60% of 175,274 Microsoft 365 Copilot Chat prompts.
31%
In March 2026, 31% of US workers said they had used AI at work to interpret, translate or summarize information and 27% for administrative tasks, the next most common uses after searching, writing and generating ideas; fewer had used it for coding or modeling, customer support or managing logistics or supply chains.
Who was asked: US workers
The question counts AI broadly and names no time period.
under 10%
In April 2026 logs of Google's Gemini products, attempts to automate a task end to end were under 10% of conversations about non-routine cognitive work, such as hypothesis testing or creative design; most of that use was partial drafting, review, ideation and finding information.
Who was asked: interactions across the Gemini App, Google AI Mode and the Gemini API
Google's own classification, which it calls an initial attempt.
Disclosure: Google makes Gemini.
See also: How many people use AI at work
Most use of consumer chatbots has nothing to do with work. People look things up, ask about health and have fun, and few turn to AI for companionship.
73%
Non-work messages rose from 53% to 73% of consumer ChatGPT use between June 2024 and June 2025.
25.1%
In 2025, 25.1% of people aged 16 to 74 in the EU used generative AI tools for private purposes in the previous three months, against 15.1% for work and 9.4% for formal education.
Who was asked: People aged 16 to 74 in the EU
86%
Among US adults who use AI chatbots, 86% use them to search for information, 51% for fun or entertainment, 48% to create or edit images or videos and 27% to get news (February 2026).
Who was asked: chatbot users n=2,605
Shares of chatbot users, who are about half of US adults.
42%
Among US adults who use AI chatbots, 42% use them to get medical advice and 40% for diet and fitness information (February 2026).
Who was asked: chatbot users n=2,605
Shares of chatbot users, who are about half of US adults.
one in ten
In February 2026, one in ten US adults said they use AI chatbots for emotional support or advice, and a smaller share for companionship.
Who was asked: 5,119 US adults, Pew's American Trends Panel (probability-based)
A share of all US adults, not only chatbot users.
A usage log counts what one product's users typed, and a survey counts what people say they did. The share of use that is work depends on the product measured, and the two methods disagree about which tasks AI does.
82% to 89%
In a May 2026 paper, Microsoft estimated that 82% to 89% of Microsoft 365 Copilot Chat use in enterprises is for work and about 10% personal, against 27% work use on consumer ChatGPT in OpenAI's study, so the share of use that is work depends on the product.
Who was asked: Microsoft 365 Copilot Chat usage in enterprise tenants
Each company classified its own product's logs, in its own way.
Disclosure: Microsoft makes Copilot and studied its own product.
over 86%
In April 2026 logs, over 86% of interactions with Google's conversational AI, the Gemini app and AI Mode, happened outside work.
Who was asked: de-identified interactions across the Gemini App, Google AI Mode
Consumer use only: enterprise and Workspace accounts are left out.
Disclosure: Google makes Gemini.
74%
In November 2025, 74% of sampled requests to Anthropic's API, mostly from businesses, were work-related, against 46% of conversations on its Claude.ai chat app.
Claude itself classified the transcripts.
Disclosure: Anthropic makes Claude; this site is compiled with Claude.
46%–61%
Chat logs from OpenAI, Anthropic and Microsoft put 46%–61% of generative AI use in the top ten tasks, against 22% in a survey of US adults run from August 2025 to May 2026. The authors say chat classifiers, which can't see a user's job, pile chats into a few generic tasks.
Who was asked: Real-Time Population Survey: US adults aged 18-64
The survey asks each worker only about the main tasks of their own job.
People who use AI often do more with it: more coding and automation at work, and more personal uses outside it. Light users stick to a few. Age changes the mix too.
22% vs. 8%
In May 2026, US employees who use AI at work a few times a week or more were nearly three times as likely as less frequent users to use it for coding assistance (22% vs. 8%) and automation (21% vs. 8%), and more than twice as likely to use it for task, scheduling or project management (21% vs. 9%).
Who was asked: employees who use AI at work
19%
In 2025, 19% of all monthly active ChatGPT Enterprise users had never used its data analysis tool, 14% had never used reasoning and 12% had never used search; among daily active users, the shares were 3%, 1% and 1%.
Who was asked: De-identified, aggregated ChatGPT Enterprise usage
Monthly users include the daily users.
Disclosure: OpenAI's own usage data about its own product.
66%
In May 2026, 66% of internet-using US adults who use AI chatbots daily had used AI in the past six months for at least one social, emotional or personal purpose, from entertainment and personal advice to companionship or romantic chat, against 49% of those using chatbots a few times a week and 35% of those using them a few times a month.
Who was asked: 4,268 US internet-using adults interviewed online by YouGov
Any one of several personal uses counts, entertainment included.
39.3%
In 2025, 63.8% of people aged 16 to 24 in the EU had used generative AI tools in the previous three months, nearly twice the 32.7% of all 16 to 74 year olds. A far larger share of the young used them for formal education (39.3% against 9.4%) and private purposes (44.2% against 25.1%), but about the same share for work (15.8% against 15.1%).
Who was asked: People aged 16 to 24 in the EU
Shares of each age group, not of AI users.
See also: Who is more likely to use it
Few people hand most of their work to AI, and the typical job uses it for a small share of its tasks. People who use it in more ways report bigger gains, and beginners fall short most often.
21%
In September 2025, 21% of US workers said at least some of their work was done with AI (all, most or some of it), including 2% who said all or most; 19% said some, 65% not much or none, and 12% hadn't heard of AI use in the workplace. In October 2024, 16% said at least some.
21%
In April 2026 logs of Google's Gemini products, AI was used for only 21% of the tasks in the median occupation with any AI use, and only 3% of occupations showed AI use for over 75% of their tasks.
Who was asked: interactions across the Gemini App, Google AI Mode and the Gemini API
Enterprise and Workspace use are left out, so work use may be understated.
Disclosure: Google makes Gemini.
90%
In May 2026, the more kinds of tasks US employees used AI for at work, the more often they said it had a positive effect on their productivity: 45% of those using it for one or two purposes, rising to 66% for three or four, 78% for five or six and 90% for seven or more.
Who was asked: employees who use AI at work
Self-reported; Gallup says the link doesn't show that more uses cause bigger gains.
15%
From October 2025 to April 2026, Claude Code sessions in which Anthropic's classifier rated the person a novice reached verified success, such as passing tests, 15% of the time and at least partial success 77%, against 28-33% and 91-92% for sessions rated intermediate or above.
Who was asked: About 400,000 interactive Claude Code sessions from about 235,000 people
Expertise and success are judged by Anthropic's classifiers from transcripts.
Disclosure: Anthropic makes Claude; this site is compiled with Claude.
3.8%
Only 3.8% of Section's June 2026 respondents wrote workable instructions for a custom AI assistant (scored 3 of 5 or higher in a hands-on test).
See also: The heaviest users
Controlled trials find large gains on well-defined tasks such as writing, and none on others, such as medical advice. Across whole jobs the measured effect is small, and many companies abandon most of their AI projects.
37%
In a preregistered online experiment in early 2023 with 444 college-educated professionals, those randomly given ChatGPT finished an occupation-specific writing task 10 minutes (37%) faster than the control group, and evaluators graded their work 0.45 standard deviations higher.
Who was asked: 444 college-educated professionals recruited online
Short online writing tasks with an early version of ChatGPT.
0.37 standard deviation
In a randomized field experiment with 776 Procter & Gamble professionals in 2024, individuals working with AI showed a 0.37 standard deviation gain in the rated quality of their solutions to real product-innovation challenges over individuals working without it, and two-person teams with AI a 0.39 standard deviation gain; the authors say individuals with AI performed at a level comparable to teams without AI.
Who was asked: 776 Procter & Gamble professionals (R&D and commercial)
A one-day virtual workshop at one company.
Disclosure: P&G gave gifts to a Harvard Business School institute, and one author was a paid consultant to P&G.
24.3%
In a 2025 study of 48 freelance workers, those who chose to use AI for single steps finished tasks 24.3% faster than workers without AI, while the few who handed AI whole tasks were 17.7% slower, partly from checking and fixing its work and partly because less expert workers chose to hand tasks over.
Who was asked: 48 human workers
Not randomized: workers chose how to use AI, and only a handful handed it whole tasks.
34.5%
In a randomized study in 2024 with 1,298 UK adults working through doctor-written medical scenarios, those given GPT-4o, Llama 3 or Command R+ identified the relevant conditions in fewer than 34.5% of cases and the right course of action in fewer than 44.2%, no better than a control group free to use any source; tested alone, the same models identified the conditions in 94.9% of cases.
Who was asked: 1,298 UK adults recruited on Prolific (stratified to UK demographics)
Written medical scenarios, not real illness.
1%
Linking two surveys of Danish workers in 11 AI-exposed occupations to payroll records through June 2024, economists found AI chatbots had no significant effect on earnings or recorded hours in any occupation; confidence intervals rule out average effects larger than 1%.
Who was asked: About 25,000 workers per survey round
Not randomized; covers the first year and a half after ChatGPT's launch.
42%
In S&P Global Market Intelligence's 2025 survey of more than 1,000 respondents in North America and Europe, as CIO Dive reports it, 42% of companies were abandoning most of their AI initiatives, up from 17% a year earlier, and the average organization scrapped 46% of its AI proofs of concept before production.
Who was asked: More than 1,000 respondents in North America and Europe
As a trade outlet reported it; S&P's own wording isn't quoted.
RAND's 2024 report on why AI projects fail repeats an outside estimate that more than 80 percent of AI projects fail, twice the rate for IT projects that don't involve AI; RAND cites a Fortune article for it and did not measure it.
Widely quoted as RAND's finding, but RAND took it from other sources and measured none of it.
See also: What experiments measure, Are companies getting a return on AI?
The estimate that follows is this site's arithmetic from published surveys of which tools people use. It is not a count of people who have built an agent. Robb Wilson, who leads this research, works at OneReach.ai, which sells software for orchestrating agents. EffectiveSC, which publishes How We Use AI, plans to offer AI services of its own. Claude, Anthropic's AI, helped compile this research, and the card that names Claude Code says so.
No survey counts people who build an agent, an automation or a custom assistant that someone else uses. For all employees, the closest published uses are coding assistance and automation. This site multiplies each use by the share of employees in that frequency group. The two uses overlap, so they do not add, and using the tool is a ceiling, not a headcount of builders. Among professional developers the same tools are ordinary: most use a coding agent in a typical week, and most say they build scripts or internal tools. The tighter bound, a hands-on test of writing a workable custom assistant, is in How far each group gets.
about 8%
Project estimate: about 8% of US employees used AI for coding assistance in May 2026, and about 8% used it for automation. Among the 30% of employees who use AI a few times a week or more, 22% used it for coding assistance and 21% for automation; among the 22% who use it less often, 8% did each. Less often is use at least a few times a year but not a few times a week, 52% minus 30%. Coding assistance is 22% times 30% plus 8% times 22%, which is 8.36%; automation is 21% times 30% plus 8% times 22%, which is 8.06%. Each rounds to about 8%. The two groups overlap, so they are not a combined 16%. Using either tool is not the same as building an agent that someone else uses.
Who was asked: US employees aged 18 and older working full or part time
This site's arithmetic from Gallup's published shares. The two uses overlap, and using a tool is not building an agent someone else uses.
90%
In JetBrains' weighted developer surveys, Claude Code use at work rose from about 3% (April to June 2025) to 18% (January 2026) and about 39% (May to July 2026), while GitHub Copilot fell from 29% to 21% and Cursor from 18% (January 2026) to 12%. In May to July 2026, 90% of professional developers used AI coding agents at work at least weekly, and 68% daily.
Who was asked: Professional developers worldwide
Weighted surveys of professional developers by JetBrains. Each share is of developers using the tool at work. Using a coding agent is not building an agent for someone else.
Disclosure: JetBrains sells developer tools and AI agents, and Claude Code is Anthropic's product; Claude, Anthropic's AI, helped compile this research.
77.3%
18.6% of respondents build in-house software to replace paid vendors, against 77.3% who build small scripts and automations and 63.0% team-level internal tools; among those replacing paid vendors, 54.4% build with AI. Stack Overflow calls predictions of an AI-driven "SaaSpocalypse" "greatly exaggerated".
Who was asked: Respondents to the build question
An opt-in survey of developers. Most say they build small scripts and automations or team tools. Replacing a paid product is a much smaller share, and the share who build that with AI is of that smaller group only.
See also: How far each group gets, What AI agents are used for
The AI at Work Benchmark
Score your organization's practices, measure your people with an anonymous pulse, and see both against the published evidence.