The same AI tool, handed to your best performer and your newest hire, doesn't produce the same lift for both of them. Here's why, what the research actually shows, and a scored way to tell who on your team needs skill-building versus more access.
Handing a team the same AI tool feels like it should level the playing field. Often it doesn’t. AI amplifies whatever judgment, pattern recognition, and taste someone already has, and someone who hasn’t built those yet gets a faster version of their current output, not a better one, because the tool has no judgment of its own to lend them. Four separate field studies, in consulting, software development, and customer support, back this up, though not in a single direction: on well-defined, playbook-driven tasks, AI can genuinely narrow the gap between strong and weak performers, while on judgment-heavy work it tends to widen it. This piece works through why, what that means specifically for a team rather than an individual, and a scored check a manager can run this week to tell which situation they’re actually looking at.
Priya has run vendor compliance reviews for nine years. Jordan started in January. On a Thursday afternoon at a regional bank, both get the same task: turn a 40-page vendor file into a one-page risk memo for the credit committee, due by 5pm, using the same AI drafting tool the bank rolled out that spring.
Both start from the same AI-generated first draft. Jordan reads it, tightens a couple of sentences, and sends it up almost as written. It’s organized, the tone is right, nothing looks obviously wrong. Priya reads the identical draft, deletes the second paragraph, and adds four sentences about a subcontractor relationship buried on page 31 that the AI’s summary had folded into one generic line about “downstream vendor exposure.” That subcontractor is the actual reason the committee should be cautious. Nothing else in the memo comes close to mattering as much.
Same tool. Same file. Same deadline. A genuinely different memo.
That’s the plainest way to see what’s happening when AI meets real work: it isn’t replacing the judgment either of them brings to the file, it’s amplifying it. Jordan’s memo is roughly what Jordan would have produced without AI, just faster and better formatted. Priya’s memo is nine years of knowing which sentence on page 31 matters, produced at a speed she didn’t have last year. The tool did the same thing for both. What came out depended on what went in.
| Jordan (first year) | Priya (nine years) | |
|---|---|---|
| What changed before it shipped | A few sentences tightened for tone | Second paragraph deleted; four new sentences added about a subcontractor on page 31 |
| Time spent editing | About 10 minutes | About 25 minutes |
| What the committee needed to know | Present, buried in one generic line the AI wrote | Front and center, in Priya’s own words |
An illustration of a genuinely common pattern: identical AI output, very different value depending on what the reviewer already knew to look for.
AI scales the judgment you already have. It has no judgment of its own to lend you.
This distinction matters because “multiplies” and “replaces” point a team in opposite directions. If AI replaces expertise, you can worry less about who on the team already knows the work cold, since the tool is meant to close that gap on its own. If AI multiplies expertise, the people who already know the work well get more out of every tool you hand them than the people who don’t, and that gap has to be actively managed, not assumed away.
What this piece works through is that whether AI narrows or widens the gap between strong and weak performers depends heavily on the kind of task in front of you, not on a single universal rule. That’s exactly why a manager needs a way to tell those two situations apart.
There’s a specific reason Priya’s read of that draft beat Jordan’s, and it isn’t that she’s a better writer. Producing a genuinely useful AI output depends on two skills the tool doesn’t have on its own: knowing what to ask for, and knowing whether what came back is actually right. Priya has both. Jordan is still building them, exactly what a first year is supposed to be doing.
Knowing what to ask for is the quieter of the two. Priya got a better memo not because she used a cleverer prompt, but because nine years of reading vendor files taught her that a subcontractor relationship hides inside a generic sentence, so she knew to go looking for it even in a draft that read as complete. Jordan had no reason to suspect page 31 mattered more than page 12, because nothing in the AI’s confident, well-organized draft signaled that anything had been glossed over. A draft that’s wrong in an important way and a draft that’s simply complete read almost identically on the page. The tool doesn’t flag which one it gave you.
Knowing whether the output is right is the second piece, the same skill applied after the fact. This is where a VP of HR and a newly promoted HR coordinator can be handed the identical AI-drafted performance improvement plan and land on two very different outcomes. The coordinator reads it, likes the structure, and sends it. The VP catches one clause that, worded that way, would let the employee argue the goals were never made clear, exactly the kind of gap that can turn a routine PIP into a legal problem eighteen months later. Neither of them wrote that clause. The AI did. Only one had the hard-won knowledge to recognize it as a problem rather than reasonable-sounding language.
Two questions tend to separate someone who can evaluate an AI draft from someone who can only tidy one.
Field research backs this up, and it’s more nuanced than a simple “AI levels the field” story. A pre-registered study of 758 Boston Consulting Group consultants tested tasks inside what GPT-4 could reliably handle, and one designed to sit outside it, requiring the same contextual, connect-the-dots judgment Priya used on page 31. On the task outside the model’s reliable range, consultants using AI were 19 percentage points less likely to land on the correct recommendation than consultants working without it.[1] Same people, same tool, and nothing about the AI’s output looked visibly different between the task it handled well and the one it didn’t. This pattern has a name, the jagged frontier: an uneven boundary between what AI reliably handles and what it doesn’t, one that gives no signal when you’ve crossed it. (If you want a practical way to sort your own team’s tasks into which side of that boundary they fall on, this workflow diagnostic walks through exactly that.)
That’s the honest, complicated version of the leveling claim. Well inside AI’s reliable range, it helped weaker performers more than stronger ones, a real equalizing effect. On work that calls for the kind of judgment Priya and the HR VP were applying, it doesn’t level anything, because recognizing where that boundary sits is itself a form of expertise. A team that hasn’t built that judgment yet isn’t protected by the tool. It’s exposed by it, just less visibly, because the draft always looks finished either way.
“Judgment” is doing a lot of work in this article, so it’s worth being specific about what actually gets amplified when AI meets someone who has it, because it isn’t one single trait. Three things tend to show up together: judgment about what matters in a specific situation (Priya’s subcontractor), pattern recognition built from having seen a version of this problem many times before (the HR VP recognizing a clause that reads fine but isn’t), and taste, a working sense of what “good” looks like in this specific context that lets someone spot a mediocre AI draft even when nothing in it is technically wrong.
None of those three can be handed to someone who hasn’t built them yet, and that’s the whole limit of the multiplier idea. A multiplier applied to zero is still zero. If a first-year analyst hasn’t yet learned what a normal variance looks like for a specific client’s numbers, no amount of AI-drafted commentary teaches them that in the moment, it just produces commentary that sounds informed either way. The fundamentals still have to be built the slow way: by doing the work, getting it wrong sometimes, and having someone more experienced explain what was missing.
A multiplier applied to zero is still zero. Build the fundamentals before you lean on the multiplier.
A field study of 244 management consultants gives this some useful shape. Researchers tracked how people actually worked with AI across a full project, not just a single prompt, and found three distinct patterns rather than one generic “uses AI” behavior.[2]
| Mode | What it looks like | What it builds |
|---|---|---|
| Deeply fused | Works with AI continuously across the whole task, shaping the problem and the answer together | New skill at working with AI itself |
| Clearly directed | Keeps their own thinking separate, hands AI specific bounded pieces, stays in charge of the judgment calls | Deeper expertise in their actual domain |
| Fully handed off | Lets AI decide both what to do and how to do it, with minimal engagement | Neither. No new domain expertise, no new AI skill |
Based on a Harvard Business School field study of 244 management consultants. [2]
The group that handed everything off matters most here: visible proof that “using AI” and “getting more out of your expertise” aren’t the same behavior. Someone can run a tool constantly and still not be multiplying anything, because there’s nothing underneath being amplified. That’s a different problem from Jordan’s, who is still building the fundamentals and, given time and support, will likely move toward the group that stays in charge and deepens its own expertise. (Why your AI skills are worth more than a master’s degree covers why that underlying skill holds its value even as AI access becomes common.) Someone in the fully-handed-off group needs the opposite of more access: less reliance on the tool for the parts of the job meant to build their judgment in the first place.
Six months after that Thursday memo, the gap between Priya and Jordan hasn’t closed. It’s wider, in a specific, predictable way once you know what to look for. Priya now drafts a first pass in AI for almost every memo, catches the same kind of buried detail she always caught, and does it in a third of the time, freeing her to carry two more accounts. Jordan uses the same tool just as often. Jordan’s memos are faster and better formatted than in January. They also still miss the thing that matters roughly as often as before, because the tool never taught Jordan what to look for. It just made Jordan’s existing work arrive faster.
This catches leaders off guard, because it runs against the instinct that handing everyone the same tool should level a team out. Most leaders hoping for exactly that get it on tasks that are more about applying an existing best practice quickly than about judgment under ambiguity. Assuming it happens automatically everywhere is the mistake. On judgment-heavy work, the senior person’s output improves by more, not less, because they have more existing skill, pattern recognition, taste, for the tool to amplify. The floor rises a little. The ceiling rises a lot. The distance between them grows.
A study of roughly 160,000 software developers on GitHub found the same shape outside of finance and HR entirely. Less experienced programmers used AI coding tools in 37% of their code, more often than the 27% figure for experienced developers, yet only the experienced developers turned that assistance into measurable gains in output and in the range of things they could build.[3] As the researchers put it directly: generative AI does not automatically level the playing field, it can widen an existing one.[3]
For a manager, that reframes what “AI adoption is going well” actually means. High usage numbers tell you people are trying the tool. They don’t tell you whether the gap between your strongest and weakest performers is closing or opening, and those are opposite outcomes that can produce the exact same usage dashboard. The honest check isn’t who’s using AI. It’s whose output has genuinely changed shape since they started, versus whose output just arrived faster.
A few signals tend to show up before the gap becomes obvious in a formal review:
| Priya (senior) | Jordan (first year) | |
|---|---|---|
| Time per memo | Down roughly two-thirds | Down roughly half |
| Accounts carried | Up by two | Unchanged |
| Buried-detail catch rate | Still consistently high | Roughly where it was in January |
| What actually multiplied | Nine years of knowing where the risk hides | Not much yet. The fundamentals are still being built |
An extension of the scene earlier in this piece, illustrating the shape researchers have documented in real teams: usage stays roughly equal, output diverges.
None of this means restricting AI access for anyone unproven, or handing out more licenses and hoping the gap sorts itself out. Both are guesses. What works is diagnosing, for each person, whether what’s missing is the underlying skill AI is supposed to multiply, or just access and practice with the tool, because the fix for each is different, and the wrong one wastes a quarter.
More access fixes an access problem. It does nothing for a skill problem.
Here’s a short test to run against each person on your team, not a formal review, just a five-minute check against what you’ve actually seen in their work.
| Question | Score |
|---|---|
| 1. Could they do this task acceptably well without AI, just slower? | 1 = yes, 0 = no |
| 2. Do they know which parts of an AI draft they changed, and why? | 1 = yes, 0 = no |
| 3. Have they ever flagged an AI output as wrong, or caught an error before it shipped? | 1 = yes, 0 = no |
| 4. Are their questions about the tool mostly about trust and judgment, not just speed? | 1 = yes, 0 = no |
| 5. Has their output changed shape in the last month, not just gotten faster? | 1 = yes, 0 = no |
Score 0-5. Fill this in from what you’ve actually seen in their work, not from a self-assessment.
| Score | What it usually means | What to do next |
|---|---|---|
| 0-1 | Fundamentals aren’t there yet | Pair with someone senior on real work; keep AI use closely reviewed |
| 2-3 | Building well, judgment inconsistent | Keep supporting, review their reasoning, be specific about what “good” looks like |
| 4-5 | Ready to lean on the tool more | Widen access and autonomy; put the time saved toward harder work |
Jordan scored roughly 1 out of 5 in January and closer to 3 out of 5 by month six. Moving in the right direction. Not there yet.
Run the check against Jordan and the diagnosis is clear: closer to skill-building than to more access. Many managers default to another round of general AI tool training here, which mostly teaches things Jordan already knows, like how to write a clearer prompt. What’s actually missing is the domain judgment Priya has, and that comes from doing the real work, with feedback, on real files, with someone senior checking the reasoning and not just approving the draft. That’s an argument for structured support, not against it: left alone, most people build this the slow way, through years of trial and error, while structured coaching shortens that timeline and gets more of the team contributing at a senior level sooner. The tool doesn’t replace that kind of mentoring. It raises the stakes on having it.
| What AI does | What the manager still owns | How it’s checked |
|---|---|---|
| Drafts a fast first-pass memo, summary, or plan | Deciding whether their edits show real judgment or just formatting | Spot-check the AI draft against the final version monthly |
| Makes everyone’s output look equally polished | Knowing whose polish is backed by judgment and whose isn’t yet | The AI Fit Check above, filled in from real work, never self-reported |
The tool changes what work looks like on the page. It doesn’t change who’s responsible for knowing whether it’s right.
Measuring whether it worked has to go past login counts. Use tells you someone tried it once. What matters is persistence, still doing it a month later unprompted, and impact, whether the work now catches what used to slip through. High usage in week one and nothing different by week eight hasn’t worked.
For the broader picture of where AI genuinely saves time on a workflow versus where it just relocates the bottleneck, see the companion piece AI as a Workflow Accelerator. And for the day-to-day version of managing a team through this shift, not just diagnosing one person at a time, our practical guide to using AI as a manager goes further.
Pick one person on your team this week. Run the five questions above against their actual work, not your impression of them. Then do exactly one thing: either pair them with someone senior on a real task, or widen what they’re trusted to do on their own. Not both, not neither. One decision, made on purpose, this quarter.
It means the tool scales whatever judgment, pattern recognition, and standards someone already brings to a task, rather than supplying judgment they don’t have. A quick test: could this person do the task acceptably well without AI, just slower? If yes, AI is multiplying something real. If no, the polish is hiding a gap rather than closing it, since a confident AI draft looks the same either way.
Because a useful AI result depends on two things the tool can’t supply: knowing what to ask for, and knowing whether what came back is right. An experienced person knows which detail in a messy file matters and where an AI draft is likely to be thin. A beginner often doesn’t yet know which questions are worth asking, so a confident answer can look complete even when it quietly skipped the one thing that mattered.
No, it depends on the kind of task. For well-defined, playbook-driven work, AI can genuinely help newer employees more than experienced ones. A field study of customer support agents found AI assistance raised productivity substantially more for novice agents than for experienced ones, largely because the tool surfaced best practices the experienced agents already knew.[4] That’s a different kind of task from a risk memo missing the one detail that matters, where the gap runs the other way.
Both, depending on the task, which is why a blanket answer misleads more than it helps. On judgment-heavy work, ambiguous, high-stakes, requiring context a person already has to hold, AI tends to widen the gap. On narrower tasks with a known right answer, closer to looking up an established best practice, it can genuinely narrow it. The manager’s job is sorting which of a team’s tasks fall into which category, not assuming either effect applies everywhere.
Diagnose before you act, rather than defaulting to more training for everyone or more access for everyone. The fastest informal check: pull two people’s AI-assisted output on the same task and compare how much you had to change before you’d trust it. If you’re rewriting one person’s drafts every time and only skimming the other’s, you already have your answer. Someone missing the fundamentals needs closer pairing, not more access. Someone who’s clearly built the judgment is ready for more autonomy, with the time saved going toward harder work.
The scenes involving Priya, Jordan, and the HR performance improvement plan in this piece are illustrations, built to show a documented pattern, not real people or real cases. The research citations point to four separate field studies and analyses, spanning management consulting, open-source software development, and customer support, because the underlying pattern, expertise deciding how much a person gets out of the same AI tool, holds up across genuinely different kinds of work, not just one industry’s version of it. All four sources were fetched and checked directly for the specific figures used here.
Weekly tips. Real examples. Practical help for busy professionals.
We care about your data, check out our privacy policy.