Everybody Can Do Anything. Nobody Needs That: Why Individual AI Competence Doesn’t Add Up

I get a surprising range of responses when I tell people that I work as an "AI Transition Consultant."

Maybe it shouldn't be surprising. The advent of Artificial Intelligence is the kind of cultural sea-change that impacts all of us - it makes sense that everyone will have notes. Even so, I'm regularly taken aback by the variety, depth, and passion of opinion I hear from the people I meet who are nearly-universally motivated to give me their take on the topic whether they have any experience with the technology or not.

It's the kind of surprise I've come to delight in: no matter where I am or who I'm talking to I can always find a unique take. And frequently a challenging one.

"They Work Faster, But They're Working At Odds"

Theresa is a people manager at a mid-sized company, and she is fed up.

I met her at a conference where she’d spent the day in sessions about AI-enabled productivity, and she had reached the point in her day where politeness gives way to candor. Her team had adopted AI enthusiastically. Enterprise licenses, a lunch-and-learn, a Slack channel for sharing prompts - by every measure her organization was tracking, the rollout was a success, and they had spent the last month patting themselves on the back for that.

But in her estimation, her people had gotten worse at their jobs.

Her example was two analysts on her team who had each spent the better part of a week building a solution to the same recurring reporting problem. One had written a script. The other had assembled a rather elegant prompt workflow. Neither had any idea the other was working on it. Both solutions functioned. Neither produced output in the format Finance had already asked for, which nobody had checked, because checking hadn't been anybody's job when the report took four hours and one person to produce.

Two weeks of skilled labor. Two competent people. One unusable deliverable, and a third week spent reconciling it.

"They work faster now," she told me. "But they're working at odds. And working at odds faster just gets you to the wrong place sooner."

Competence Doesn't Aggregate

I’ve actually been hearing different versions of this same problem a lot lately, usually from people managers on the front line of the AI Transition. The tools are working, and the people are using them, but somehow the work is not getting better. A team of capable people, each using AI well on their own terms, can produce an organization that is measurably worse off.

This is the part of the transition that I think gets consistently underestimated: individual competence with AI does not automatically aggregate into organizational value.

What struck me about Theresa's story is that nobody in it did anything wrong.

Both analysts identified a real inefficiency. Both applied a capable tool to it. Both delivered something that worked. If you evaluated either of them in isolation - which is, notably, how we evaluate almost everybody - you'd conclude the AI rollout was going great.

The failure wasn't in either person. It was in the space between them. A decision can make perfect sense to the person making it while creating cost, risk, or duplication for the organization around them. Fifty people making locally sensible choices do not necessarily add up to a sensible operating model, and nothing about the individual choices will tell you that.

You can see the same pattern everywhere once you start looking for it. A lot of very smart people look at a tool that can do anything and conclude they should do everything with it. But organizations don't need everybody to do everything. 

They need the things that people do to fit together.

It's Not The Quarter You'd Think

There's research that sharpens this considerably.

KPMG and the University of Texas at Austin ran a field study with 523 early-career professionals, published in Harvard Business Review this July. Participants completed realistic business tasks using an AI agent, and their work was graded against a baseline established by the AI agent working alone. 

The question was direct: who actually adds value on top of the machine?

Half did. The other half split into two groups. About a quarter produced work roughly equivalent to what the AI produced by itself, and about a quarter produced work that was worse than the AI working alone.

The researchers named the groups Amplifiers, Delegators, and Apprentices.

And the findings about the underperformers should stop every HR leader cold: it's not the quarter you'd think. The Apprentices - the group that did worse than the machine alone - scored higher on critical thinking, domain knowledge, and AI literacy than the Delegators did. On those foundational measures they were essentially indistinguishable from the Amplifiers. Meanwhile the Delegators, who scored lowest on every foundational skill assessed, quietly matched the baseline by accepting whatever the AI handed them with minimal interrogation. They looked productive. They contributed close to nothing.

The things we currently use to identify high performers didn't predict performance. Neither did AI literacy, which is the skill everyone assumes will replace them.

What actually predicted performance was how people worked. Amplifiers framed the problem before handing it over, anchored the work in a real domain framework, defined what a good answer would look like, and then argued with the output until it got there. Apprentices had the same raw capability and spent their interactions reorganizing information and chasing questions that didn't matter.

Of course, this is one study, at one firm, on simplified versions of real tasks; I wouldn't hand these percentages to your CFO as a forecast. But the shape of the finding is hard to dismiss, and it maps directly onto what Theresa was describing. Roughly three-quarters of participants had the foundational skills to succeed. A third of that group couldn't convert them into results.

Her analysts weren't short on capability. They were short on structure. And that’s an HR problem.

You're Not Recruiting Amplifiers

Here’s why that distinction matters so much:

How people work is not a personality trait. It's a product of workflow design, decision rights, incentives, and what gets measured. Which means the distribution of Amplifiers and Delegators inside a given organization is something that organization built, whether or not anybody built it deliberately.

You're not recruiting Amplifiers. You're manufacturing them. Or you’re failing to.

That reframes the whole problem. If the differentiator were foundational skill, this would be a hiring question. If it were AI literacy, it would be a training question. Because it's neither, it becomes a question about how work is coordinated and how judgment is exercised - and the instruments for that are ones HR already owns:

Decision rights. Who is permitted to accept an AI output as final? At what level of stakes, with what review? Most organizations haven't written this down, and are improvising it daily.

Assessment that examines process, not just deliverables. This is the one that follows most directly from the research. If people are evaluated solely on what they turn in, Delegators are invisible by design, and that invisibility is exactly what allows the pattern to persist. Underperformance gets to hide inside of acceptable output indefinitely.

Provenance norms. Documenting why an AI output was accepted, modified, or rejected, and on what criteria. This makes judgment visible and therefore coachable - and it produces the audit trail governance will need anyway. One practice, two problems.

An intake path for locally built tools. People are going to build things, and mostly that's good. A lightweight registry where a useful tool gets adopted, absorbed, or retired helps keep employees aware of existing (or attempted) options and turns individual initiative into shared capability instead of shadow infrastructure.

Governance that's maintained rather than published. Policy written against the capabilities of eighteen months ago is decorative. Someone needs to own the question of what changed upstream and what that affects downstream.

Five Questions Worth Asking This Week

None of this gets solved by a single policy document, and every organization's approach will ultimately look different. But you can learn a great deal about the environment you're currently running by asking five questions and listening carefully to how readily they're answered.

1. Where does our data actually go? Not where policy says it goes - where it goes when someone is behind on a deadline and the sanctioned tool is slow. This tells you whether your governance is operational or aspirational.

2. Who is allowed to ship an AI-assisted output without a human reviewing it?.This tells you whether decision rights exist, or are being improvised person by person.

3. When was any AI-touching workflow last checked against the model currently running underneath it? This tells you whether your workflows are maintained systems or set-and-forget ones.

4. How would we know if two people were solving the same problem right now? This tells you how much coordination your work design actually provides.

5. Does any part of our performance process look at how work got done, rather than only what arrived? This tells you whether you can currently distinguish an Amplifier from a Delegator at all.

The answers are often surprising, and they're considerably cheaper to gather now than after somebody's third week of reconciliation.

This Is a Work-Design Problem

The instinct when a technology arrives this quickly is to treat adoption as the goal and literacy as the solution. Get everyone trained, get everyone licensed, get everyone comfortable. Those things matter. They're just nowhere near sufficient, and Theresa's team is the proof: two well-trained people with good tools and real skill, producing three weeks of work and no usable result.

What's actually being asked of organizations right now is harder and more interesting than a training rollout. It's a rethink of how work gets coordinated, how decisions get made (and by whom), how performance gets assessed when the deliverable no longer tells you as much, and where human judgment sits in a process that can increasingly run without it.

That's management-system work. It's work design. It belongs to the people who have spent their careers building the structures that determine how an organization actually functions - which is to say, it belongs to HR. And it's arriving faster than most functions are prepared to handle.

It's also the reason Reframing HR exists, and the reason we built the HR Reframed Masterclass: not to teach the tool of the week, but to help senior HR and TA leaders build the judgment and the structures that determine whether all that individual capability adds up to anything.

Nobody has a template for this yet. What works in a 60-person firm with one shared workflow will be useless at an enterprise with twelve functions and a regulator, and most organizations will get some of it wrong on the first attempt. That's fine. That's how new work gets designed.

But somebody has to own the space between capable people. That space is where the Next Iteration of Work either comes together or falls apart.