In March 2026, researchers from the University of Texas at Austin, working with KPMG, published in Harvard Business Review the most detailed picture yet of how people actually ask AI for help at work. They tracked 2,597 employees at a large company for eight months and analyzed 1.4 million requests made to AI assistants.
The result unsettled anyone who expected the tool to level the team on its own. Only about 5% of employees asked AI for help in a way the researchers called sophisticated: they gave specific instructions, asked the AI to explain its reasoning before accepting an answer, and refined the request several times until it actually worked. Everyone else asked once, took the first answer, and moved on.
The most counterintuitive finding is this: the difference had nothing to do with who used AI more. Some of the most frequent users were, at the same time, among the least sophisticated. What separated the top group from the rest was not effort. It was method — and method can be taught, copied, and published.
A minority carries the average
The researchers identified four signals that mark sophisticated use: the person returns to AI repeatedly to refine what they already asked, is ambitious with the first request instead of asking for something too small, demands that the AI explain its reasoning before accepting the answer, and picks the right tool for each task instead of defaulting to the same one out of habit. None of these signals require technical knowledge — they are work habits, not programming skills.
The same pattern shows up, with even clearer numbers, in a study published in the Quarterly Journal of Economics by researchers from MIT and Stanford. They tracked 5,179 customer support agents before and after an AI assistant was built into the support system. Average productivity rose 14%. But the average hid two very different groups: beginner agents gained up to 34% in productivity, while experienced agents barely moved — in some cases, their customers' satisfaction even dropped slightly.
The reason behind that result is what matters here. The AI assistant taught experienced agents nothing new; it already reflected what they did. For beginners, though, the system replicated the responses and approaches of the team's best agents — in practice, it turned the know-how of people who already did the job well into something any newcomer could use on day one. An agent with two months on the job, using the tool, produced as much as a six-month veteran without it.
Why generic training does not fix it
The common temptation is to close that gap with a "how to use AI at work" course for everyone. A controlled experiment published by the National Bureau of Economic Research in 2026, with 1,174 participants aged 25 to 45, helps explain why that alone falls short.
In the experiment, each participant solved a business problem, with or without access to an AI assistant. Without AI, more educated participants scored 0.548 standard deviations above less educated ones — a large gap. With AI access, that gap fell to 0.139 standard deviations: a reduction of roughly 75%. The tool alone already acted as a partial leveler.
But the study also showed its limit. Looking at the chat logs, the researchers found that more educated participants kept extracting more value from the same tool — giving more precise instructions, exploring more options, refining further. AI closes part of the distance by itself, but not all of it. What closes the rest is someone taking the request that already works — discovered by whoever has the most practice — and publishing it for those who have not gotten there yet.
What has to be in place
Standardizing the request that works is not writing a manual and hoping someone reads it. Five verifiable mechanisms make the difference.
Requests published by routine, not by person. Overdue-payment follow-ups, meeting minutes, ticket triage, a reply to a recurring complaint: every routine that repeats every week becomes one tested, published request — not a fresh invention from every person, every time.
Reach set by whoever publishes it. Whoever creates the request chooses who sees it: the whole team, one department, or only people with a given role. That avoids two opposite mistakes — locking everything behind excessive permissions, or spreading a sensitive request to people who should not see it.
Agents that run the routine in steps, not just answer a question. The next step beyond a reusable request is an agent that already knows how to carry out the whole routine — pull the information, apply the right format, leave it ready for review — without the person rebuilding the reasoning every time.
Approved knowledge behind the request. The template points to the current version of the company's process or policy, so the outcome does not depend on whoever made the request having copied the right version of the right document.
An owner and a current version. When the process changes, the published request changes with it — automatically, not because someone remembered to update a copy scattered across emails and personal notes.
This is how Skyller was designed: ready-made process templates and reusable agents, with reach set by person, group, and role, so that the way one person learned to ask for something becomes a standard available to the rest of the team.
From isolated talent to collective gain
The most expensive effect of individual improvisation never shows up on a cost spreadsheet — it shows up in turnover. When the person who figured out how to ask for a well-structured sales proposal leaves the company, that knowledge leaves with her. No one documented the request, only the result it produced while she was there.
Publishing the request flips that logic. The customer support study showed the size of the effect when this happens automatically, through the tool itself: a new agent reaching a veteran's level in two months. A company that deliberately identifies whoever already asks well — and publishes that request to the right group — does not need to wait for AI to figure it out on its own, or rely on luck when hiring. And it does not need to start from zero: Skyller, for instance, already ships with more than 170 ready-made process and policy templates to adapt to each team's routine.
Under this design, the productivity gain stops being about "who is good with AI" — a talent unevenly distributed and hard to predict at hiring time — and becomes a repeatable process instead: discover, publish, set the reach, keep it current.
A roadmap to get started
Before buying more licenses or scheduling another generic training session, five steps help turn individual talent into team capability.
- List the five most repeated routines on the team. Ask each person which AI request they make every week, almost without variation. Overdue-payment follow-ups, meeting summaries, ticket triage: whatever repeats is a candidate to become a template.
- Find whoever already asks well, even without knowing it. According to the KPMG study, that person is not necessarily the most technical, nor the one who uses the tool the most — it is whoever refines the request until it actually works. That person is probably already on the team.
- Turn their request into a published template, with the right process or policy built in — not a loose piece of text that everyone copies, pastes, and adapts incorrectly.
- Set the reach before publishing. The whole team, one department, or only a given role: the right reach avoids both permission chaos and unnecessary approval bottlenecks.
- Revise the template when the process changes, not only when someone complains the answer came out wrong.
