A program of EffectiveSC
How We Use AISources checked October 24, 2026 to November 1, 2026
Where things stand / Power users as enablers
What the evidence says about power users lifting their colleagues, which ways of doing it cost them least, and why a small protected core that builds for everyone else is this project's judgment, not a finding.
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 memory and routing services for AI. Findings that favor shared assets, such as agents, templates and memory that a team shares, sit next to those interests, so this page keeps measured results, what people and organizations report, and the project's own judgment apart, and labels each. Claude, an Anthropic model, helped compile this research.
The best-supported version is narrow: put the expertise inside the tool people already use. At one firm's support desk, an AI assistant lifted the least-skilled agents most, let newer agents do as well as more experienced ones, and did nothing significant for the most skilled, whose best practices, its authors suggest, it was passing on. Some help that once came from a colleague now comes from the tool: one person working with AI matched a two-person team without it, and top open-source maintainers who became eligible for a free coding assistant worked alongside fewer peers. Directly, power users mostly reach colleagues who are already capable: engineers were likelier to try a new coding agent once their peers had, but in a randomized trial, giving people an AI assistant didn't change how their close colleagues used the same tools. Through what they build, power users reach the least experienced.
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.
Who was asked: 5,172 customer-support agents at one Fortune 500 software firm
One firm's support desk, studied before today's AI agents. The least-skilled fifth of agents, by skill measured before the assistant, not new hires.
Agents two months into the job with the AI assistant performed as well as or better than agents with more than six months' tenure without it.
A comparison by time on the job within one firm, not a test of training. In their working paper, the authors call it suggestive evidence that the assistant passes on the best practices of more able agents.
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. The agents with the most skill to pass on gained least from the tool.
In a randomized experiment with 776 P&G professionals, an individual working with AI matched a two-person team without it.
A randomized workshop study at one company; here AI stood in for a teammate.
79%
Top open-source maintainers who became eligible for free GitHub Copilot worked in repositories with 79% fewer peers; with controls the drop is 49%.
Eligibility for a free tool, not use of it; the measure is how many peers worked in the same repositories.
Disclosure: Some of the authors work at GitHub and Microsoft, which make Copilot.
+216%
Microsoft engineers were likelier to first try Copilot CLI when over a quarter of their skip-level peers had used it (+216% odds), a quarter of their code-review peers had (+54%), or their manager had (+82%).
Odds of first trying the tool among Microsoft engineers, an association; it doesn't show that peers caused it, or what anyone went on to produce.
Disclosure: The authors work at Microsoft, which makes Copilot CLI.
1.4 hours
In a Microsoft-run experiment across 66 firms and 7,137 knowledge workers, those offered Copilot spent 1.4 fewer hours a week on email (12%); those who used it saved about 2 hours (17%). Meeting time didn't change.
The change stayed with the people who had the tool: close colleagues' use of Teams and Word didn't change, the authors report.
Disclosure: Microsoft ran the experiment, and some of the authors work at Microsoft Research.
See also: What experiments measure, Judgment: What this means for a company
This order is the project's reading of its own three reports on enablers, not a measured ranking: no study has compared these ways head to head. Cheapest for the power user first. Capture what they already do: AI drafts a reusable prompt, skill or workflow from their best work, and they check it, minutes an item, provided the experts whose know-how goes in agree to it. Share how they use AI, by posting examples and showing their work, also cheap, though it mostly reaches colleagues who are already capable. Set the quality bar that shared work is held to, a recurring cost that grows with every shared asset; without one, the cost of shallow AI work lands on whoever receives it. Build shared agents, assistants and templates that run inside other people's work: one build, then upkeep, and the widest reach when people use them, though most public custom GPTs drew little use. Keep the shared memory those assets draw on current, a cost that grows with how fast the work changes, so shared memory is two jobs, cheap to capture and costly to keep right. Teach, a cost that grows with every learner, and one that managers, trainers and peers a rung up can carry instead; managers who expect AI use go with far more skill than training does. Build custom software, the costliest to build and keep up, which pays where a process is stable. Last, answer questions one at a time, the default and the costliest, since every request comes out of the helper's own work; the next section shows who carries it. The first three enable without slowing power users down. The fourth is likely the best use of their build time when it is aimed at high-volume work done by less experienced people and judged by how often others reuse it.
51.1%
On the WebArena web-navigation benchmark, GPT-4 agents that induced reusable workflows from their own past runs (Agent Workflow Memory) raised overall success 51.1% relative to the BrowserGym baseline, 12.0 points absolute, and outperformed SteP, which uses 14 workflows written by human experts. On Mind2Web the relative gain in step-wise success was 24.6%.
A benchmark of web agents on an older model, not a workplace; it shows AI can draft reusable routines from past work.
76.7%
In a randomized experiment with 61 full-time professionals, 23 of 30 (76.7%) authorized AI systems to reuse their materials, against 30 of 31 (96.8%) who were asked about human professionals; what they chose to share didn't differ significantly. Their reasons moved toward loss of control, replacement and downstream governance, and they favored consent, control and credit.
A small preprint with simulated tasks; the comparison is with human professionals, not colleagues.
61%
61% of Microsoft's Frontier Professionals say their teams share AI tips, new agents, learnings and mistakes, against 36% of other AI users.
What AI users say their teams do; it doesn't show how much is shared or whom it reaches.
Disclosure: Microsoft sells the AI tools its survey asks about.
54%
54% of Microsoft's Frontier Professionals say their teams discuss quality standards for AI-assisted work, against 29% of other AI users.
Teams that discuss standards, with no timing; it doesn't say anyone sets them before the work starts.
Disclosure: Microsoft sells the AI tools its survey asks about.
40%
40% of US desk workers received AI “workslop” in the past month, costing about two hours per instance.
Polished-looking AI work that lacks substance, as the people who received it reported it.
Disclosure: BetterUp, which ran the survey with a Stanford lab, sells coaching to companies.
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.
OpenAI's top firms by use, set against its median firm; the report doesn't say the Custom GPTs and Projects are shared, and the gap shows where effort goes, not what it causes.
Disclosure: OpenAI's own usage data about its own product.
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.
Public GPTs only; it says nothing about reuse inside a company.
roughly one in four
OpenAI's December 2025 enterprise report says leading firms turn on connectors that give AI secure access to company data inside core tools, and that roughly one in four enterprises had not.
OpenAI doesn't state the base.
Disclosure: OpenAI's own report about its own customers.
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).
A training company's own tests; associations, not the effect of any one course, and the June comparison is between organizations, not people.
Disclosure: Section sells AI training to companies.
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.
Each edition compares different groups, so the two multiples aren't a trend; associations, not effects.
Disclosure: Section sells AI training to companies.
$6M
The State Department says its AIRMOD pilot used generative AI to digitize, translate and extract key data from over 115K documents, saving 76.5K hours and avoiding $6M in costs (reported in its July 2026 playbook).
One agency's account of one automation pilot; it doesn't say how the hours or costs were counted. It was the playbook's only dollar saving, and it came from automating a workflow, not from the agency's chatbot.
Disclosure: The State Department reporting on its own project.
See also: What memory products keep, and who controls it, When memory goes wrong, Old tools or new
Before AI, a small share of employees already supplied a large share of the help in most organizations. The research on helping finds a price for it: in one firm, people who spent more time on citizenship work such as helping got smaller raises and advanced more slowly, and in a daily study, helping lifted helpers' mood but cut into their sense of progress on their own work. Few AI users say they're rewarded for reinventing their work with AI when the results fall short, and a sizable share of the most capable say they are likely to change employer within a year. These studies come from before AI or from surveys; none measures what enabling costs a power user's own output.
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.
Network studies from before AI, published in a business magazine; value-added collaborations, not all help.
In archival data on 3,680 employees of one professional services firm, those who spent more time on organizational citizenship behavior, such as helping others, had lower salary increases and advanced more slowly, controlling for time spent on task performance.
Who was asked: 3,680 employees of one professional services firm (archival data)
An association in one firm's records, from before AI, holding time on core tasks constant; it doesn't show that helping caused the smaller raises.
In a study that surveyed 82 employees for 10 workdays, daily citizenship behavior such as helping others was associated with positive affect but interfered with employees' perceptions of their work goal progress.
Who was asked: 82 employees surveyed for 10 workdays (experience sampling)
A small daily study from before AI; progress is what the employees reported each day, not measured output.
13%
Only 13% of AI users say they're rewarded for reinventing their work with AI even when the results fall short.
What AI users say about their own employers.
Disclosure: Microsoft sells the AI tools its survey asks about.
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.
What workers say they intend, with no comparison figure for other groups.
Disclosure: PwC sells consulting, including on AI.
This is the project's judgment, not a finding: power-user results for everyone, and power-user operation for a few. The few are scarce by every measure of depth, and nobody counts how many people build things others use. Protect them: give them sanctioned tools connected to company data, time set aside to build, credit when others reuse their work, a second owner and a review date for every shared asset, and leave to turn down one-off requests. Don't push everyone toward heavy use. The largest gains from expertise built into a tool went to the least skilled, who didn't have to become power users to get them; judgment decides who gains from AI advice; just past the edge of what AI can do, it made skilled consultants worse; and even the most advanced users keep some work away from it. The middle should inherit what the core builds, inside the tools it already uses, with managers who expect it to use them. Whether people keep the gain once the tool is gone is mixed: in one experiment a sizable gap came back, though the people who gained most kept part of it. No study sets the right size for the core, so start with the people already building, and let reuse show how many more are needed.
2%
About 2% of US workers say they do all or most of their work with AI, in both October 2024 and September 2025.
Who was asked: US workers: adults employed full or part time with one job or a primary job
How much of their work is done with AI, as workers report it.
2.7%
In Section's January 2026 tests, 2.7% of knowledge workers were AI practitioners and 0.08% experts (under 3% combined); 69% were experimenters and 28% novices.
A training company's hands-on tests of knowledge workers at large firms in three countries; one edition, not a trend.
Disclosure: Section sells AI training to companies.
5%
Only 5% of employees use AI in ways that transform their work, and the most-trained employees are 55% more likely to leave.
EY's own definition of transformative use, in a single reading, so no trend.
Disclosure: EY sells consulting, including on AI.
43%
43% of Microsoft's Frontier Professionals say they intentionally do some work without AI to keep their skills sharp, against 30% of other AI users.
What AI users report in a vendor's survey; the comparison is other AI users.
Disclosure: Microsoft sells the AI tools its survey asks about.
nearly 10%
A GPT-4 business assistant on WhatsApp had no average effect on 640 Kenyan small businesses: owners below the median did nearly 10% worse, while those above it may have gained over 15% (not statistically significant).
Small-business owners in Kenya, not employees, getting advice from an AI mentor, not expertise built into a tool.
60% to 70%
On tasks inside AI's frontier, below-average consultants improved 43% with GPT-4 against 17% for above-average ones; on a task just outside it, consultants were right 60% to 70% of the time with AI against about 84% without.
Consultants on one task built to sit just outside the model's ability.
In a randomized online experiment, a generative-AI assistant raised performance on an incentivized business problem-solving task for all participants, most for lower-education ones: the gap between higher- and lower-education participants fell from 0.548 to 0.139 standard deviations, closing three quarters of it. In a follow-up without AI, treated participants did no worse than controls and lower-education participants kept part of their gain, but a sizable education gap re-emerged.
Who was asked: 1,174 adults aged 25 to 45 in a randomized online experiment outside firms
One online experiment on one task, in a working paper; it is about education, not experience on the job.
See also: The heaviest users, Estimate: Who uses the tools that building takes, What goes with wider use
No study compares shared memory, shared agents, peer helpers and training head to head, so the claim that shared memory is the top lever goes beyond the evidence, and so does the order on this page. A memory store people have to go and search isn't what the support study tested; it tested expertise delivered inside the work. No study shows power users lifting their teammates directly as a measured effect, measures what enabling costs a power user's own output, follows the same people from casual to deep use, or sets how much of a power user's time enabling should take or how many builders an organization needs. Nobody counts how many people build things others use. The field study the project has proposed would randomize teams to shared memory, shared agents, peer helpers or the organization's usual training, and weigh what colleagues gain against what the enabler gives up; until something like it runs, the ranking on this page is a judgment.
The AI at Work Benchmark
Score your organization's practices, measure your people with an anonymous pulse, and see both against the published evidence.