A program of EffectiveSC
How We Use AISources checked October 24 to 29, 2026
Where things stand / What to do for power users, casual users and non-users
How the groups of people using AI at work are changing, the groups inside them, how to tell a power user, and the project's judgment on what to do for each and where the return is.
Robb Wilson, who leads this research, works at OneReach.ai, which sells software for orchestrating AI agents. EffectiveSC, which publishes How We Use AI, plans to offer AI services of its own, among them memory and routing. This page's judgments favor funding a few people to build assistants and workflows that others reuse inside the tools they already have, which sits close to those businesses. The cards report what each source measured or what people told it. The framework and the judgments are the project's own reading, not measurements, and no study has measured the return from each group. Cards that cite Microsoft, OpenAI or a training company about their own products or customers say so. Claude, Anthropic's AI, helped compile this research.
See also: Who publishes How We Use AI, What else we plan to sell or offer
That was the starting hypothesis, and the answer depends on the question asked. In Gallup's survey of US employees, the share who never use AI in their role has fallen from about four in five to under half, and daily use is the band that grew. The occasional middle has held steady, so people leaving non-use aren't piling up there; in technology and finance, where daily use is highest, it has started to shrink. In the St. Louis Fed's survey of generative AI, use on every workday hasn't risen, while use in the past week rose until its latest reading dipped, so on that question heavy users are a smaller share of users, as the hypothesis expected. On depth the hypothesis holds: the share of workers who do most of their work with AI hasn't moved. So it is wrong on Gallup's frequency bands, open on the Fed's and right on depth. Who uses AI shows the main trends, and the cards here add the rest.
30%
30% of US employees used AI at work a few times a week or more in May 2026, up from 11% in May 2023.
Who was asked: US employees aged 18 and older working full or part time for organizations
Gallup asks how often people use AI in their role, not how much of their work is done with it.
39.2%
Past-week use of generative AI for work rose from 28.2% of employed adults in Q3 2024 to 39.2% in Q2 2026, then fell to 34.8% in August 2026.
Who was asked: Employed US adults aged 18 to 64
Generative AI only, in an online panel weighted to the population. The latest reading is a single dip, and use on every workday in the same survey hasn't risen.
2%
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.
Who was asked: US workers
Pew asks how much of their work is done with AI, a different question from Gallup's. The all-or-most share was the same a year earlier, so the growth went into some of the work.
See also: How the groups are changing, Non-users, and why, Technology leads daily use by a wide margin
Sorting people only by how often they use AI mixes two things that move differently. Frequency, how often, is the axis every tracker measures, and it is spreading like any popular technology. Depth, how much of the work is done with AI that the person can still check, is where the value seems to sit, and it stays concentrated: a daily user can still be shallow. Across both sits a role, the builder, whose assistants, templates or standards other people use. Nobody counts builders, so the project says a few percent. Each group splits further, and the split decides what to do. Non-users divide by their reasons: no use for AI in their job, attachment to their current way of working, worry about privacy or compliance, objection on principle, or no access, plus invisible users who meet AI in search and apps without counting it. Casual users divide into those who treat AI as a better search box, those who polish their own writing, those who meet it only inside existing apps, and those who tried it and stalled. Many frequent users still take one step at a time. Power users divide into delegators, specialists in one domain, builders, orchestrators of several agents, gardeners who write down what works, skilled people their organization can't yet absorb, and people who hide their use. The project defines a power user by depth: someone who regularly hands AI multi-step work and checks the result. Frequency only finds candidates. Three questions sort people: do they hand AI multi-step work at least weekly, does their AI-assisted work ship without heavy rework, and do other people use something they built, documented or shared? Yes to the first two is a power user, and yes to all three is a builder, the scarcest profile. No survey yet sizes these groups on both axes at once, and the published shares called power users run from a few percent to most workers, because each source draws its own line.
5.5%
In Section's June 2026 report, 5.5% of US knowledge workers meet its proficiency bar, which now includes agent and automation skills; 73.5% are experimenters.
Who was asked: 5,026 US knowledge workers across industries
A training company's survey with its own definition of proficiency, which now includes agent and automation skills.
Disclosure: Section sells AI training to companies.
54%
54% of Microsoft's Frontier Professionals say their teams discuss quality standards for AI-assisted work, against 29% of other AI users.
Microsoft's survey of people who already use AI. It defines the group partly by shared, repeatable practices, so part of this gap is built in.
Disclosure: Microsoft sells the AI tools its survey studies.
10%
In Microsoft's readiness index, 19% of AI users are in the Frontier zone, 10% have strong skills but lack the systems to apply them (blocked agency), 16% are stalled and 50% emergent.
Who was asked: AI-using knowledge workers in 10 markets
Zones of Microsoft's own readiness index, built from what people report. They are shares of people who use AI, not of all workers.
Disclosure: Microsoft sells the AI tools its survey studies.
28.5%
Of about 334,000 public custom GPTs in early 2024, 28.5% had more than 10 conversations; the store's top-listed GPTs drew about a third of conversations.
Who was asked: Public custom GPTs collected from OpenAI's GPT Store by January 20, 2024
Public assistants in one store, early in its life; it says nothing about reuse inside companies.
See also: The heaviest users, Heavy and light users do different things, How far each group gets, Estimate: Who uses the tools that building takes
The goal is to turn private gains into shared capability without burning out the people who produce them. Find power users by what they hand to AI, from usage data rather than surveys, since many hide their use. Give them sanctioned tools with real data access, protected time and a builder's mandate: one assistant, template or workflow that a team reuses, judged by how many people use it each week and how often its output needs fixing, not by their own usage. Credit that reuse in reviews and pay, since few people who use AI say reinvention is rewarded. Govern what they build the way spreadsheets should have been governed: a named owner, a review before wide use, a log and an off switch. Set norms for checking, because experienced users can feel faster than they are, and expertise built into a tool does least for the most skilled. Expect large personal gains that stay personal unless they're captured, more output for others to review, and people other employers want. Measure reuse, rework and retention.
61%
Among working respondents in 47 countries, 58% intentionally use AI at work at least every few months (31% weekly or daily); 72% in emerging and 49% in advanced economies. 61% have at least rarely hidden their AI use and 55% have presented AI output as their own.
Who was asked: 48,340 people in 47 countries (32,352 working), Nov 2024 to Jan 2025
An online survey across many countries, in which hiding use even rarely counts, so it is a loose bar.
13%
Only 13% of AI users say they're rewarded for reinventing their work with AI even when the results fall short.
Who was asked: 20,000 knowledge workers in 10 markets who use generative AI at work at least occasionally
Microsoft's survey of people who already use AI.
Disclosure: Microsoft sells the AI tools its survey studies.
19%
In METR's early-2025 randomized trial, 16 experienced open-source developers took 19% longer with AI tools, yet afterward believed AI had saved them 20%.
METR has since said newer data suggest a speedup, but it calls that evidence weak because of who took part.
40%
40% of US desk workers received AI “workslop” in the past month, costing about two hours per instance.
Who was asked: 1,150 full-time US desk workers (BetterUp Labs and Stanford Social Media Lab), September 2025
Desk workers describing output they received from colleagues; it isn't a measure of power users' own work.
Disclosure: BetterUp, which ran the survey with a Stanford lab, sells coaching to employers.
29%
PwC's AI front-runners, with scarce skills and strong AI capabilities, are 14% of workers; 29% of them say they are very or extremely likely to change employer within a year.
Who was asked: 49,364 workers in 48 countries and regions, 29 sectors, May to June 2026 (weighted by gender and age)
A consultancy's survey, with no comparison for other workers.
Disclosure: PwC sells consulting on AI and the workforce.
See also: What experiments measure
The goal is one rung up, not a builder. Move casual users, who use AI a few times a month or a year, to weekly use on two or three recurring tasks in their own job, and move frequent users who still take one step at a time to handing over whole steps they can check. Work through managers who expect and model use: where managers back AI, far more employees use it often. Ship assistants and templates that builders made inside the tools people already use, so the deeper path is the default. Clear the worry about privacy and compliance with approved tools and a plain policy, because that, more than access, is what infrequent users say holds them back. Widen uses on purpose: people who use AI for more kinds of tasks report gains more often, though that is self-reported and could run either way. Agree what finished means for work done with AI, so speed doesn't become someone else's rework. Skip generic prompt courses and usage targets, since trained workers still score low on hands-on tests. Expect modest gains per person: a few percent of work hours in official records and representative surveys, and several times that in vendors' and consultancies' surveys, so budget on the lower figure. The total matters because the group is large. Plan on about one rung a year for a group, a planning rule, since no study has followed the same people up. Measure weekly use of specific workflows, tasks handed over end to end, rework, and where the saved time goes.
38%
In February 2026, among US employees who use AI at work less than a few times a week, in organizations that make AI tools available, the most common reasons for not using it more often were concerns about data privacy, security or compliance (38%) and preferring to keep doing their work the way they do now (36%); 25% said they are ethically opposed to using AI, 22% that they don't believe AI can assist with their work, 2% that they don't have access and 2% that they aren't permitted. More than one reason was allowed.
Who was asked: US employees who use AI at work less than a few times a week (infrequent users), in organizations that make AI tools available
More than one reason was allowed, and the base is employers that make AI available, not all employees.
2.6x
Section's two editions measure the manager effect differently: in January 2026, workers whose managers expect AI use were 2.6x as proficient as those whose managers discourage it; in June 2026, workers whose managers expect and require AI scored 1.5x higher than those with no expectations.
A training company's tests in two editions with different comparisons, showing associations, not effects.
Disclosure: Section sells AI training to companies.
1.5x
In Section's January 2026 tests, employees trained on AI were 1.5x as proficient as untrained ones, yet still averaged 40 out of 100. In June 2026, workers at organizations that provide agent-specific training scored 44 vs 25 (1.8x).
The first edition compares trained with untrained workers; the second compares employers that offer agent training with those that don't.
Disclosure: Section sells AI training to companies.
about 3%
Two years after ChatGPT's launch, Danish register data rule out effects on earnings or hours above 2%; adopters report saving about 3% of work hours.
The time saved is what adopters report; the result on earnings and hours comes from official registers, and covers chatbots before today's agents.
40 to 60 minutes
ChatGPT Enterprise users report saving 40 to 60 minutes per active day on average.
Who was asked: OpenAI survey of 9,000 workers at almost 100 enterprises
Self-reported, per active day, by customers of the company that sells the product.
Disclosure: OpenAI's own survey of its own customers.
See also: What goes with wider use, Hours assisted and time saved
The goal is to deliver what AI does without asking non-users to adopt it. Sort them by reason, because each needs a different move: access and clear rules for the few who lack them; one credible win in their own job, shown by a peer, for those who don't believe AI can help them; AI that arrives inside the workflow for those who prefer their current way; approved tools and a plain policy for the wary; and respect, with a human path kept open, for those who object. Build AI into the systems and processes they already use, or let a teammate or an agent do the AI-shaped step for them. Expect little from persuasion: most adults who don't use chatbots don't plan to start within a year, and telling workers how much time AI saves changed their beliefs but not their use. Expect real gains where AI is built into high-volume, standardized work, and the largest for the least experienced. Measure process outcomes such as cycle time, errors and cost per case, not adoption.
39%
In February 2026, among US employees who don't use AI at work, in organizations that make AI tools available, 39% said they don't believe AI can assist with their work and 5% that they aren't permitted to use it. More than one reason was allowed.
Who was asked: US employees who don't use AI at work, in organizations that make AI tools available
More than one reason was allowed. Lack of access is rarely the reason, as Who uses AI shows.
60%
In February 2026, 60% of US adults said they read AI summaries at the top of search results, and 49% said they use AI chatbots.
Who was asked: 5,119 US adults, Pew Research Center American Trends Panel, February 17-23, 2026
All US adults, not only workers, and Pew doesn't say how many readers of AI summaries never use a chatbot.
15%
Across 5,172 customer-support agents, an AI assistant raised issues resolved per hour by 15% on average (QJE, 2025); the 2023 working paper gives 5,179 agents and 14%.
One firm's support desk, studied before today's agents; the assistant sat inside the window where agents already worked.
See also: Non-users, and why
No study measures the return from each group, so this is the project's ranking, to pilot and measure. First come builders aimed at shared workflows, the most leverage per scarce person, because what one person builds serves many at little added cost, while coaching grows with their hours. The best evidence that expertise travels through a tool is a support assistant built from one firm's own conversation records, which lifted the least-skilled agents most and gave the most skilled no significant gain. Second comes redesigning processes so AI is the default, including for non-users: the largest total value, and the practice that most separates the companies reporting a return. Third, managers enabling casual and frequent users, which is steady and broad. Last, generic training for everyone and targets for usage, the weakest bets. No single group carries the return. Power users' gains stay private unless someone captures them. Casual users lack use cases, which builders produce. Non-users gain from AI built into their work, which builders make. And expert help is scarce, because a small share of employees already carries much of the helping. So fund the flow: builders encode the work, the middle adopts it, and non-users receive it inside their tools. Per person, power users working as builders return the most; in total, the middle does, reached through what builders make; and non-users pay off through redesigned processes, not persuasion.
36%
The least-skilled support agents gained about a third: 36% more issues resolved per hour for the lowest skill quintile (QJE; 34% in the working paper). The journal's summary gives about 30% for less-skilled and less-experienced workers.
The least-skilled fifth of agents, by skill measured before the assistant, not new hires.
The most skilled agents saw no significant productivity gain from the assistant, and small but significant declines in resolution rates and customer satisfaction.
Small declines, measured at one firm.
3% to 5%
Before AI, across more than 300 organizations, 20% to 35% of value-added collaborations typically came from only 3% to 5% of employees.
Who was asked: The authors' network studies across more than 300 organizations
Research from before AI, published in a business magazine.
about 6%
In McKinsey's survey of May and June 2026, nearly 9 in 10 respondents (89%) reported regular AI use in at least one business function and 80% said AI had improved their own productivity, but only 37% said AI had contributed to their organization's EBIT, about the same as the 39% of 2025; about 6% are high performers, who attribute 5% or more of EBIT to AI and report significant value, and nearly three quarters of them redesigned workflows against about a quarter of other respondents.
Who was asked: 1,719 participants at all levels in 97 nations
Respondents at every level, from executives to individual contributors, reporting on their own organizations; not company accounts.
Disclosure: McKinsey sells consulting on AI.
7x
Frontier firms send about 2x the median firm's messages per seat but 7x as many to GPTs; weekly users of Custom GPTs and Projects rose about 19x in 2025, and about 20% of Enterprise messages run through a Custom GPT or Project.
Frontier firms are OpenAI's top firms by use, set against its median firm; the report doesn't say the Custom GPTs and Projects are shared with anyone.
Disclosure: OpenAI's own usage data about its own product.
See also: What experiments measure, Are companies getting a return on AI?
No survey yet asks the same people how often they use AI, how much of their work they hand to it and whether others use what they build. So the groups inside the groups have no sizes that add up, and no single share can be called the power users. No study follows the same workers from casual to power use, so how many climb, and how fast, is unknown, and the reading that the casual band is a waystation fits the totals without being measured. No study sorts returns by group, compares ways of spreading power users' know-how head to head, or counts what helping others costs them. So this page doesn't claim that power users lift their teammates directly, that casual users become power users at a known rate, that power users abandon their old tools, that companies can skip enablement because power users enable themselves, or that shared memory is the best way to lift the middle. The benchmark survey this site is preparing, which isn't open yet, is meant to measure the first of these gaps.
The AI at Work Benchmark
Score your organization's practices, measure your people with an anonymous pulse, and see both against the published evidence.