Two people on the same team, with the same licence, getting completely different value out of it. The difference is smaller and more learnable than it looks.
Almost everyone has access to the tools now, and the results people get from them vary enormously. The instinct is to explain that with aptitude, and the evidence doesn’t really support it. What heavy users do differently is mostly procedural: they decide task by task what AI should touch, they keep some work deliberately manual, they build on what they already ran last week rather than starting fresh each time, and they compare notes with colleagues in the open. This piece covers those habits, an honest look at how much of the gap is individual versus environmental, how the advice changes depending on your role, and a two-week starting point.
Two analysts on the same finance team. Same Copilot licence, issued the same week, same manager, similar experience.
One of them uses it maybe twice a week, mostly to tidy up an email, and privately thinks the whole thing has been oversold. The other has stopped building the monthly variance commentary from scratch. She feeds it last month’s version, this month’s numbers, and a note about what she wants flagged, then spends her time on the two lines that look odd instead of on assembling the document. She’d struggle to tell you when that changed. It happened over about six weeks.
Nobody trained her differently. That’s the part that makes this interesting rather than just unfair.
The usual explanation is aptitude, and I don’t think that holds up. What’s actually happened is that one of them has built a handful of habits and the other hasn’t, and the habits are specific enough to write down. Before getting to them, it’s worth being honest about how much of this is really about the individual at all, because a lot of the writing on this topic quietly assumes everyone is starting from the same place.
They aren’t. A New York Fed survey supplement from November 2025 found 37% of employed respondents say their workplace doesn’t offer AI tools, and another 11% say their employer actively prohibits them. [1] Almost half the workforce, in a nationally representative US sample, is not in this conversation yet. Among those who are, 19% said they were still in the phase where tasks take longer, because they’re learning.
That last group is the most reassuring thing here. The awkward stage where AI makes your work slower is not a sign you’re bad at this. It’s the normal middle of the process, and roughly one in five people who use AI at work are in it right now.
Access is not proficiency. Having the licence is the start of the work, not the end of it.
Microsoft’s 2026 Work Trend Index surveyed 20,000 knowledge workers who already use AI at work, and separated out the most advanced 16% of them. The behavioural differences it found are less exciting than you’d expect, which is what makes them useful.
The clearest one: 53% of the advanced group say they pause before starting a piece of work to decide what should be done by AI and what by a human, against 33% of everyone else. [2] Not a better prompt. A few seconds of deciding, before anything is typed.
The second clearest is stranger. The advanced group are more likely to say they deliberately do some work without AI to keep their own skills sharp, 43% against 30%. The people using it most are also the people most deliberate about where they don’t.
Both of those are self-reported, so what’s being measured is how people describe their own working style rather than what they actually do. Worth holding that in mind. Still, it’s a consistent picture, and it matches what the habits look like from the outside:
| Habit | What it looks like on a Tuesday | What it replaces |
|---|---|---|
| Deciding first | Ten seconds before starting: which part of this is judgement, which part is assembly? | Pasting the whole task in and hoping |
| Bringing context | Attaching last month’s version, the brief, the actual numbers, before asking for anything | Describing the situation in a paragraph and getting something generic back |
| Reusing, not restarting | Keeping the prompt that worked and running it again next month with new inputs | Writing a fresh request every single time |
| Keeping a manual lane | Writing the client-facing summary themselves, on purpose, every time | Handing over everything and slowly losing the feel for it |
| Comparing in the open | Posting what worked in a team channel, including the attempts that didn’t | Everyone privately solving the same problem five times |
A synthesis of the behavioural contrasts reported in the Microsoft Work Trend Index 2026 plus common practice. The habits are our framing, not Microsoft’s categories.
The first habit is the one worth actually building, because the other four tend to follow from it. In practice it’s a note, not a meditation. This is what one looks like filled in, for a recurring task:
| Question | Monthly variance commentary |
|---|---|
| What am I actually producing? | Two pages explaining why the numbers moved, for the leadership pack |
| Which part is assembly? | Pulling the figures into the standard structure, restating last month’s context, formatting |
| Which part is judgement? | Deciding which two variances matter, and what I think caused them |
| What does it need from me before it can help? | Last month’s document, this month’s actuals, and the three things the CFO always asks about |
| How will I know the draft is wrong? | I check every number against the source. I have been caught once by a plausible-looking total |
The five questions take under a minute once the habit is formed. Fill this in once per recurring task, not once per session.
The skill is deciding, not prompting.
Here’s where I’d push back on the version of this story that gets told most often, including in some of Microsoft’s own framing of it.
Buried in the same data is something that cuts against the whole idea of an AI elite. When Microsoft ran its analysis across 29 factors, organisational conditions (culture, manager support, how the company handles talent) accounted for roughly twice as much of the AI impact people described as individual mindset and behaviour did. [2] Microsoft is careful to call that a statistical association rather than a causal effect, and so am I. But it points somewhere specific: whether you become a heavy AI user is substantially about the room you’re standing in.
That matches the New York Fed picture from the first section. If your employer hasn’t given you the tools, or has banned them, no amount of personal initiative closes that gap.
The other half of the honest answer is that this is genuinely learnable, and there’s decent evidence for it. A randomised trial with 758 consultants gave one group a chatbot and another group the same chatbot plus a short overview on how to use it well. The trained group produced measurably better work, with the difference holding up statistically across specifications. [3]
There’s a catch in the same experiment, and it’s important enough that leaving it out would be a bit dishonest. On a task deliberately chosen to sit outside what the AI could do well, both AI groups did worse than the people working without it, and the trained group did worst of all. They scored around 60% and 71% correct against 85% for the control group.
So a short course made people better at using AI on suitable work and more confidently wrong on unsuitable work. The training raised capability and it raised confidence, and confidence moved faster. This is also a 2023 experiment on a 2023 model, published in 2026, so treat the specifics loosely and the shape seriously.
Which is exactly why the fourth habit in that table is keeping a manual lane. The people who never work without it lose the reference point they’d need to notice when the output is confidently off.
So the honest summary of where the gap comes from is three things, in rough order of how much they matter:
“Get better at AI” is advice with no edges. What actually helps is knowing where the payoff sits in your particular job, because it moves around a lot.
| Role | Where heavy users pull ahead | The trap for this role |
|---|---|---|
| Content marketer | Feeding it real brand material, past performers and the actual brief, so drafts start from your voice | Volume. Producing four times as much undifferentiated copy and calling it productivity |
| Marketing ops / analyst | The recurring report. Same structure monthly, new numbers, so the prompt compounds | Trusting a total that looks right. Every figure still needs checking against source |
| HR / people ops | Preparation work: synthesising survey comments, drafting frameworks, prepping difficult conversations | Letting it near a decision about a specific person. That stays human, always |
| Finance / operations | Narrative and commentary around numbers, not the numbers themselves | Arithmetic. Ask it to reason about figures, not to calculate them |
| Manager or team lead | Reading long inputs fast: proposals, reports, board packs, before a decision | Delegating the reading and then having to defend a view you didn’t build |
Where the practical gains sit by role, based on the recurring work each one owns. Our own framing rather than a research finding.
Notice what the middle column has in common. Every one of them is a task that repeats. That’s not a coincidence, and it’s the most practical thing in this article.
A one-off task gives you one shot at getting the instructions right and no chance to improve them. A monthly task lets you do it badly in January, slightly better in February, and by April you have something that takes twenty minutes instead of three hours. The heavy users mostly aren’t better at any individual attempt. They’ve just been round the loop more times on the same piece of work.
Skill compounds where the work repeats. Start with something you do every month.
The instinct is to go and learn more about AI generally. That’s the slowest route, and it’s roughly what most corporate training does, which is part of why so much of it doesn’t stick.
The faster route is narrower than feels comfortable. Pick one recurring task. Get good at that one. Let the rest follow.
A starting sequence, not a measured method. Step 4 is the one that separates people who improve from people who stay level.
Two things to expect. The first attempt will probably take longer than doing it by hand, which is the awkward middle from earlier and is not a signal to stop. And the output will be mediocre in a specific way: generically competent, missing the things only you know. That’s information about what context you didn’t give it, not a verdict on the tool.
If you want the fuller picture of what this looks like at the level of a whole role rather than one task, we’ve written about what actually makes someone an AI-powered professional, which goes considerably deeper than five habits.
If you manage a team, the finding that organisational conditions explain roughly twice as much as individual mindset is a direct instruction, and it isn’t a comfortable one. It means most of the variance you’re seeing across your team is yours, not theirs.
Three things follow from that.
There’s one more thing worth being explicit about, because it’s where this goes wrong in a way that’s hard to reverse. When someone on your team starts handing real work to AI, the ownership question needs an answer before it comes up in a difficult moment rather than after.
| What AI does | What the person still owns | How it gets checked |
|---|---|---|
| Assembles the first draft from real inputs: last cycle’s version, current data, the house structure | Every number against its source, the interpretation, and the decision about what to flag. Their name is on it either way | Spot-check one output a week at random, not when something feels off. Random beats suspicious |
Agree this once per delegated workflow, in writing, before the first time something goes wrong.
Judge the work, not the tool use.
Teams do build these habits on their own, and plenty of the people who are now the heavy users in their company got there without any help. What structured training changes is the pace and the spread: it gets a whole team to the same place rather than producing one enthusiast and eleven people who feel behind, and it can point the practice at the work people actually do instead of at generic exercises. That’s the argument for it, and it’s also roughly the shape of moving an organisation from awareness to fluency.
For this week, though, the smallest useful move is the one from section five. Pick the task you do every month that you find most tedious, spend one minute on the five decide questions, and save whatever works. That’s it. That’s the gap, most of it.
Someone who gets substantially more out of the same AI tools than their colleagues do, usually because of a handful of procedural habits rather than any technical knowledge. Microsoft’s 2026 research classified roughly 16% of the AI users it surveyed as its most advanced group, and the behaviours that separated them were unglamorous: deciding task by task what AI should handle, bringing real source material rather than descriptions, and comparing notes with colleagues openly. Worth knowing that the classification is self-reported, so it measures how people describe their own working style rather than an observed output difference. The term also gets used loosely by vendors to mean anyone with high usage numbers, which is a different and much less interesting thing.
The behavioural gap is well documented and the career gap is much shakier than the headlines suggest. On behaviour, Microsoft found large differences in the 20-point range on things like pausing to decide what AI should do. On careers, the figure most often quoted is a wage premium of around 60% for AI skills, from PwC’s Jobs Barometer. That number is real, and PwC states in the report itself that it doesn’t control for education, experience or location, and that it compares advertised salaries in job postings that mention an AI skill against ones that don’t. It isn’t evidence that a given person’s pay rises if they learn AI, because most of that gap is probably composition: the postings asking for AI skills are disproportionately senior technical roles. We’ve left it out of this article for that reason.
The available evidence points to learnable, with one real caveat. A randomised trial with 758 consultants found that a short overview on effective use produced measurably better work than the same tool with no guidance. The caveat from the same study is that on a task deliberately designed to sit outside what the model could handle, the trained group performed worse than people working without AI at all, and worse than the untrained AI group. Training improved capability and confidence, and confidence improved faster. Practically, that means the learnable part isn’t just technique, it’s calibration: knowing where the tool stops working. That’s why deliberately keeping some work manual shows up as a habit of heavy users rather than a contradiction.
Worry about the conditions before you worry about the person. Microsoft’s own analysis of its 2026 survey found organisational factors accounted for roughly twice as much of the reported difference in AI impact as individual mindset and behaviour, which it describes as an association rather than a cause. The most common practical reason someone isn’t using AI much is that they can’t see which of their tasks it’s for, and that’s a design problem you can fix in one conversation by naming a specific recurring task. There’s also a second group worth separating out: people who tried it, hit the phase where everything takes longer, and quietly stopped. Around 19% of employed AI users in a late-2025 US survey said they were in that phase. They need cover, not encouragement.
Pick one task you do at least monthly and get good at that single task before broadening out. Spend a minute writing down which part of it is assembly and which part is judgement. Give the tool the real inputs rather than a description of them, which means attaching last cycle’s output, the current data and an example of the standard you’re aiming for. Then save the instructions that worked, because reusing them next cycle is where the compounding actually happens and it’s the step almost everyone skips. Expect the first attempt to take longer than doing it by hand. Two or three cycles on one recurring task will teach you more than any amount of general reading, largely because you get to see the same work improve rather than starting from nothing each time.
The three external figures here come from their original publishers rather than from coverage: the Microsoft Work Trend Index 2026 report itself, the New York Fed’s Liberty Street Economics post on its November 2025 Survey of Consumer Expectations supplement, and the published version of the Dell’Acqua et al. field experiment, read via the open-access PDF hosted by Harvard Business School because the journal’s own page blocks automated access. All three were checked on 29 August 2026. Three caveats belong here rather than only in the body. Microsoft’s advanced-user classification is entirely self-reported, and that sample skews towards technology roles and large firms. The New York Fed figures are self-reported and US only. The consultant experiment ran on a 2023 model in spring 2023 and was published in 2026, so the effect sizes should be read as a shape rather than as current numbers. One widely quoted statistic was deliberately excluded: PwC’s finding of a roughly 62% wage premium for AI skills. It is genuine and primary, and PwC states in the same document that it doesn’t control for education, experience or location, which makes it unsuitable for the claim this article would have used it to support. The five habits, the role table and the two-week sequence are Future Factors’ own framing rather than research findings, and are captioned that way.
Weekly tips. Real examples. Practical help for busy professionals.
We care about your data, check out our privacy policy.