The Adoption Gap is a fortnightly newsletter for leaders who want AI that actually works. Each issue cuts to the research, data and trends that matter, not the hype, not the doom, just what closes the gap between AI deployed and AI adopted. Curated by Mark Phelps and Drew Horton. A Mark Phelps production.
Last issue we found that access to AI tools is solved and breadth is not, that the return comes from using them across enough of the work, and that almost nobody does. This issue is about the skill that decides whether that breadth pays off at all. It is not a technical skill, and most firms are not hiring for it or training it.
THE ONE TO READ
If you only read one AI article before your next leadership meeting, make it this one.
The workers who get the most out of an AI tool are not the most skilled. They are the ones with an accurate read on their own limits, and that read turns out to be the whole game.
The ABCs of Who Benefits from Working with AI: Ability, Beliefs, and Calibration — Andrew Caplin, David Deming and colleagues, NBER Working Paper 33021 | FREE
Here is the finding that should change how you staff an AI rollout. In a controlled experiment with 732 people, the researchers separated two things we usually blur together: how good someone actually is at a task, and how well they know how good they are. The second one is called calibration, and it did more work than anything else in the study.
The AI tool helped almost everyone, and it helped the lower-ability people most, which is the equalizing effect we have seen before. But hold ability flat and a sharper pattern appears. Among people of the same skill, the ones with an accurate read on their own performance gained significantly more from the tool than the ones who were overconfident or underconfident. The people who know they are weak at something gained the most of anyone, because they know exactly when to lean on it.
That makes intuitive sense once you think about it like this. Working well with an AI tool is a series of small judgment calls, trust this output or check it, take that draft or redo it, and you can only make those calls if you know where your own judgment is reliable and where it is not. Miscalibration is a tax on every one of those decisions. The researchers found that erasing it would nearly double the degree to which the assistance narrows the performance gap between workers.
The hopeful part is the actionable part: Ability is hard to move and calibration is a skill, and skills can be trained. So, the single highest-impact thing a leader can do to raise the return on an AI rollout may have nothing to do with the tool itself. It is helping people build an honest picture of what they are good and bad at, which is not something most performance systems reward or even measure.
Which lands where this whole newsletter keeps landing: the bottleneck was never the technology. It is the human judgment around it, and that judgment can be designed for. Buy the licenses and then teach people to know their own limits and edges. One of those is a purchase order, the other is the actual work.
LEADERSHIP LOAD OUT: BRING THIS TO THE MEETING
Ask whether anything in your AI rollout helps people learn what they are good and bad at. Not training on the tool, training on themselves. If the answer is nothing, you are spending on access and skipping the one input the research says decides the return. Calibration is trainable and almost free. It is also the thing hardly anyone owns.
ALSO ON OUR RADAR
Revisiting the ABCs of Working with AI: A Replication with Radiologists — Daniel Martin, arXiv preprint (Jun 10, 2026) | FREE
The follow-up that makes the lead hard to wave away. One of the original authors took the same ability-and-calibration test into a real high-stakes setting, 68 radiologists reading chest X-rays with AI support, 11,420 paired reads. Same result: lower baseline ability and better calibration each predict larger gains from the tool. If it holds for radiologists reading scans, the excuse that it is a lab curiosity does not survive. Preprint, so we flag it as not yet peer-reviewed.
PARTING THOUGHT
The most useful thing you can know at work is the shape of your own blind spots. An AI tool does not remove that requirement, it raises the price of ignoring it. The tool will do what you ask, confidently, whether or not you should have asked. Knowing the difference was always the job. Now it is the whole job.
Know someone working on AI at their company? Forward this to them.
Hit reply and tell me what you think. We read every response.
And if there is something you think we should cover, send it over. We plan these issues as one connected argument, but the best suggestions come from readers, and we will give you props when yours lands.
Published for leaders putting AI to work in the US and UK
Mark Phelps is a product and design leader with more than 20 years of Fortune 100 experience at GE, Fidelity, TIAA, Bank of America and Marsh McLennan. He works as a Fractional VP of Product Design with a focus on AI-driven environments.
For more than three years he has worked as a fractional product design leader, first at Rockefeller Capital Management, where he was engaged to provide critical process improvements, and then on an AI-powered healthcare software product, brought in for urgent user interface deliverables. Both engagements were extended well beyond their original scope. He is a named inventor on a pending design patent for the product's interface.
He holds a Stanford AI Certificate earned with Distinction and an MBA in Digital Strategy from Suffolk University (Beta Gamma Sigma). He is a United States Marine Corps veteran.
20 Portsmouth Ave., Suite One #161, Stratham, NH 03885