DAOKO

Corporate AI usage fallacy

Scott Bagby ·

“In this new economy, three groups will have a particular advantage: those who can work well and creatively with intelligent machines, those who are the best at what they do, and those with access to capital.”

Cal Newport, Deep Work (2016)

Newport wrote that ten years ago, building on work by Erik Brynjolfsson and Andrew McAfee. Leave capital aside. The first two groups describe what is happening inside companies right now.

You would expect the people who work well with AI and the people who are best at what they do to overlap. In most companies, they have split apart. Most CEOs can’t tell which of their people sit in which group.

That matters because decisions on budgets, tools and headcount are being made now, with data that points at the wrong people.

My view is simple. As AI does more of the producing, the scarce skill becomes judging what it produces. That skill belongs to the people who know what good looks like. Many of them are not heavy AI users. A CEO who sorts people by AI usage will keep the wrong ones.

1

Two groups that look like winners

Many companies now describe a power law. A small group, often around 10% of staff, appears to drive most of the gain from AI. The problem is how that group gets found.

Most companies find it by usage. They count licences, tokens and prompts. That measure picks up two very different groups.

Lower expertise High expertise
Heavy AI use Volume producers. They top the usage dashboard. The real 10%. Same quality, in less time.
Light AI use The roles most likely to shrink. The experts. They know what good looks like.

The first group confuses more output with better results. I have received five-page emails that were clearly written by AI. They took me longer to read than they took to write. On a usage dashboard, the sender looks like a star.

Others see the same thing. In a 2025 survey by BetterUp Labs and the Stanford Social Media Lab, 40% of US desk workers said they had received AI-generated “workslop” in the past month. Each case took close to two hours to deal with.

These are often the people who confused effort with results before AI. Now they have more output to point to.

The second group is the real 10%. Their work was always good. AI lets them produce it faster. Nobody notices, because most managers judge the output, and the output looks the same as it always did. Only the time has changed.

2

Why you can’t see them

This leaves the CEO with two problems. The usage dashboard ranks the volume producers at the top. Reviewing the work won’t find the real 10% either, because their work hasn’t visibly changed.

Some of them won’t appear on the dashboard at all. When approved tools fall behind, your best people often find their own, creating a shadow IT setup within the organisation.

The only way to see them is to ask how long a good result took, before and now.

3

The work is moving to review

As models improve, machines will do more of the producing. The human job moves to deciding whether the output is right, and putting your name to it.

Reviewing doesn’t require skill with AI tools. It requires knowing what good looks like. That knowledge often sits with the people a dashboard calls laggards. Think of the estimator with 30 years in the business, or the account lead who knows which client will push back and why. They may rarely write a prompt. They can tell in a minute that a confident answer is wrong.

AI output is persuasive to someone who doesn’t know the subject. It reads well and sounds certain. The non-expert accepts it. The expert catches it.

Two things follow. Experts need a short grounding in how AI fails. It sounds sure of itself when it is wrong, and it leaves things out without saying so. And expert time becomes the constraint. If AI produces ten times the output, review is the bottleneck. That time needs protecting.

There is a deeper reason this matters. Someone has to be responsible for the output. AI can’t stand behind a forecast or be accountable to a client. A person has to. The people best placed to take that responsibility are the ones who know the work.

My bet is that this shift will move faster than most org charts. I’d watch for three signals. Review time grows as a share of knowledge work. Companies start naming accountable reviewers for AI output. Job descriptions ask for judgement in a field ahead of skill with AI tools.

4

What this means for cuts

Some roles will shrink. It doesn’t help anyone to pretend otherwise.

A cut list built from usage data is still the wrong list. It keeps the people producing the most output, and that includes the slop. It cuts the people who use AI least, and that includes the experts.

A company that does this keeps its volume and loses its judgement. The knowledge that made it good leaves with those people. The AI that remains has nobody left who can check it.

5

What to do this quarter

• Map your people on expertise and AI use together. Usage alone puts the volume producers at the top and hides the experts.

• Name an accountable expert reviewer for any AI output that leaves the building or drives a decision. Protect their time.

• Measure the time to a good result. Volume of output says nothing about value.

6

The advantage goes to the overlap

Newport’s first two groups were never meant to be rivals. The companies that do well will put them together. The people who work well with machines will produce. The people who are best at what they do will judge the work and take responsibility for it.

The hard part is seeing who is who. Usage data won’t tell you.

If you are working through this in your own company, I’d be glad to compare notes.

Contact

If you’re working through one of these questions, I’d be glad to compare notes.

scott@daoko.com