Most companies who use AI have two or three people who are really good at it.

Everyone else asks them, or uses AI in trivial ways and gets little from it, often without realizing they are missing anything.

That is a super user problem. It is not a company using AI well.

You probably already have one, and here is why I can say that before knowing anything about your company.

This is not new. AI is just adding to it.

Rob Cross, Reb Rebele and Adam Grant published research in Harvard Business Review in 2016 covering more than 300 organizations. In most cases, 20 to 35 percent of the value-added collaboration came from 3 to 5 percent of the people.

Margaret Mead is believed to have said "Never doubt that a small group of committed people can change the world, because it is the only thing that ever has." That reads as a celebration.

Cross and his team found that should not be a celebration. A small group carries the load, the load quietly breaks them, and the organization misses their burn out and assumes it has capacity.

AI did not create this. It is another iteration of it. Sometimes the same people, now carrying two queues. Sometimes someone junior who got curious first and now fields questions all day with none of the standing that used to come with being asked.

Why giving everyone a chatbot does not fix it

When you hand someone an answer from AI, you are asking them to judge it. Keep it, or throw it out.

But judging the answer takes the same knowledge that would have let them skip asking. So they guess. Sometimes they keep a good answer. Sometimes they throw one away.

MIT researchers gathered 106 experiments that measured all three: the person alone, the AI alone, and the two together. On average the pair did worse than whichever one was better by itself.

• Spotting fake hotel reviews. AI got 73 percent. People got 55 percent. Together, 69 percent. The person made it worse.

• Identifying birds from photos. AI got 73 percent. People got 81 percent. Together, 90 percent. The person made it better.

The difference is not the tool. It is who was better going in.

Why would an AI trainer tell you that? Because there is a version of this that works, and it is not the version most companies are doing.

What actually worked

A team at the Max Planck Institute went back through the 74 studies behind those experiments and looked at did the people ever find out whether they were right?

Most of the time, no.

When the studies did close that feedback loop, they got better results. AI explanations without the feedback loop got worse results.

I mean old school feedback loop not the trendy loop engineering here.

Loop engineering is the AI model's loop. Plan, act, check, retry. Making the machine iterate better before it hands you anything.

That is helpful, but I mean the feedback loop with the humans. A person tests something, they find out whether it worked, and they get better at knowing when to trust the output.

Who should be in the loop when?

AI knows everything written down, instantly, and belongs in the middle of the process. The junior person knows what actually happened on the floor today, runs the AI tool, and learns from the senior review at the end. The senior person knows which output is wrong and why, designs the AI tools at the start, and reviews at the end."]

[C] AI in the middle. The junior runs the tool. The senior designs at the start and reviews at the end.

Map any process you run regularly and split it that way.

The senior person is better than the model at this work, so the final review is the pairing that gains. The junior did the work, so they learned something nobody had to sit down and teach them. The review happens with the junior in the room, which is where finding out actually works.

It also changes what the senior does all day. Right now they answer questions. In this arrangement they review outcomes. Reviewing ten finished attempts costs less than being interrupted ten times, and it is the thing they are actually good at.

Nobody comes out of the loop. You are being deliberate about which person is in it, and where.

Most people aren't distilling knowledge in ways that are reusable

Designing the process is where most improvement efforts end. The process lives in a document or in the senior person's head, and it gets explained again every time somebody new arrives.

It has to become a reusable tool. A prompt, a project, a skill, or an agent. Something the next person runs without the expert present.

A woman I trained a few years ago just did this within a large financial services firm.

Three people there could run a proper root cause analysis. One had been doing it about thirty years. Most situations no one even looked at why things went wrong. People who hit a recurring problem ignored it, or guessed.

What replaced that does not ask anyone to judge an answer. You answer questions about what you observed, which is the only thing you know better than the model does. The model proposes what probably caused it. Then you go fix that and watch whether the problem stops.

The method tells you whether you were right. The feedback loop is built in.

From three people to thousands, now able to think about problems more effectively.

Nobody has quantified whether those thousands of analyses are as good as the expert team's. They are better than the nothing they replaced, and the hardest cases still go to the experts.

Her thirty years are in the question set now, not in her calendar. And the whole culture has shifted from one workflow AI pilot. Almost everyone now talks and thinks about root cause and problem solving. And they are curious to use AI for more things.

If you are trying to start an example, try methods that already tell you when you are wrong. Premortems get tested by what actually goes wrong. Dependency maps get tested by what actually blocks. A proposal process or sales follow up process tells you what actually worked.

Why building the tool is the thing that lifts the load

Cross separates what people ask you for into kinds of resource. Information and access can be shared without using them up. Time and energy cannot. Those run out.

A question to a person is time and energy.

The same question answered by an AI tool that person built is information and access.

The AI tool moves the requests for help out of the finite column.

Reading the Jerk Layers this week

The JERK layers this week are qualitative

No numbers here. This is a ladder of states, not a measurement. Find your rung.

If you do nothing, the unhappy path

Position

Two or three people are good with AI.  The rest may not even realize how bad they are at it or they are pestering the super users.

Velocity

The asking super users spreads. More people ask the super users, more often, wider range of questions. Still looks like success; your super users are spread too thin.

Acceleration

The rate of asking climbs. They start batching requests and blocking time. People pre-apologize before asking. Someone says out loud that they did not want to bother her.

Jerk

People stop asking. Not because they learned, but because the wait got long enough that guessing feels cheaper. They go around. They do it the old way and do not mention it.

The going around might look like the problem resolving and fewer questions to the super users.

What actually happened is that adoption reversed, and it is invisible because the thing you were counting went down.

If you build a new AI tool, the happy path

Position

One expert and one AI super user build one reusable tool. It handles the most common version of a recurring problem. It takes a few rounds.

Velocity.

People use it without being told to. That is the real signal, because it means the tool is easier than asking.

Acceleration.

A second one gets built, faster, and it does not need the same expert. The pattern is visible now, so somebody else's method starts getting captured.

Jerk.

The questions change character. The expert does not get fewer of them. She gets harder ones. Instead of how do I run this, she gets here is a case the tool could not handle. People start asking to build more tools and use AI more.

That is the flip, and you can read it without counting anything. Ask your expert what kind of questions she gets now compared to six months ago. If the answer is harder ones, the derivative turned.

Your five-minute practice this week

Prompt your team instead of your favorite LLM. Ask your team, via a quick poll:

The last time you got stuck on something AI might have helped with, what did you do? 
·  Figured it out with AI myself.
·  Asked someone. If you asked someone, who? How long did you wait?
·  Did it the old way.
·  Did not think of AI.

You will see who the nodes are from the names and how concentrated that load is on them, whether they might be spread too thin, how many people never thought to use AI at all.

The question worth an afternoon

Who here already knows how to do essential things well?

And what would it take to turn their questions, not their answers, into a prompt, a project, a skill, or an agent that the rest of the team can run without them in the room?

The goal was never to make everyone a super user.

It is to get one super user's method out of her calendar and into the whole team.

Best wishes,
Rose

Sources

Cross, Rob, Reb Rebele and Adam Grant. "Collaborative Overload." Harvard Business Review, January–February 2016.

Vaccaro, Michelle, Abdullah Almaatouq and Thomas W. Malone. "When combinations of humans and AI are useful: A systematic review and meta-analysis." Nature Human Behaviour 8 (2024): 2293–2303.

Berger, Julian, et al. "Fostering human learning is crucial for boosting human-AI synergy." arXiv preprint 2512.13253, December 2025. Not yet peer reviewed.

The hotel review and bird classification figures come from Cabrera, Perer and Hong, 2023, as reported in the Vaccaro analysis.