There was a product analyst on a team I worked with — call her M. She was not the most senior person on the team. She was not the most technical. By every legible measure of the role, she was the third-best of four.
She also out-produced the other three combined. Reports that took the team days came back from her in an afternoon. Decisions that used to require a week of cross-functional asks came back the same day, with three options laid out and the trade-offs named. Her dashboards drifted toward the unreasonable: clear, well-annotated, surfacing exactly the metric you didn't know you needed.
The other analysts attended the same training. They had the same tool access. They had been told to "use AI more." None of it produced the gap.
The gap was not skill. The gap was that M had become a different kind of thinker.
The skill-badge frame misses it
The default move for adapting people to AI is training. Run a workshop. Hand out a prompt cookbook. Give everyone access to the model and check whether usage goes up.
That kind of training matters as a floor. Everyone should be at the floor.
But the difference between the floor and what M was doing is not prompt skill. It is the shape of her thinking. She did not approach a problem by figuring out what to ask the model. She approached the problem by thinking out loud, in writing, with the model already in the room. The model was not the destination of her thinking. The model was part of the medium.
You could see it in how she opened her laptop. She was not "going to write a report." She was opening a thinking session where the report would emerge as a byproduct. The first draft used to be the hard part. For M, the first draft was the cheap part. The hard part was knowing which draft was right.
AI-native people don't prompt better. They think differently.
The new instincts
What M had, that the other analysts did not, was a set of instincts about what not to delegate.
She didn't let the model do the framing of a problem. Framing was the part where the analytical taste lived, and she would not outsource it — she would generate the frame herself, then ask the model to argue against it, then revise. She didn't let the model summarize a document she was about to act on. The summary would become her version of the document, and she had learned the hard way that the summary loses the things that turn out to matter.
She did, on the other hand, let the model do almost all of the mechanical writing. The data pull, the chart, the formatting, the executive-summary boilerplate, the recommendation framing. Anything that had a queryable definition of "right" was the agent's job. Anything that depended on her read of the specific person who would receive the report stayed hers.
The instincts were not in the cookbook. The instincts were what she had built by failing at this repeatedly and noticing where the failure came from.
The next literacy is knowing what not to delegate.
Personal canon
The other thing M had — the thing none of the workshops had told her to build — was a personal canon.
Not a knowledge base. A small set of artifacts: paragraphs she had written herself, distinctions she kept coming back to, observations about her own thinking that she had bothered to articulate. She would paste them into model conversations as context, not because the model needed her bio but because she needed the model to operate against her thinking, not against its own median.
It served two purposes.
First, it gave her something to interrogate the model with. When the model proposed a frame that sounded plausible, she had something to test it against. She was not absorbing the model's framing by default; she was running it past her own.
Second, it protected her from drift. Models converge on a median voice. People who write a lot with models discover, six months in, that their own writing has started to sound like the model unless they actively defend it. M had noticed this happening to herself. The canon was her response.
Your personal canon matters more as models get better, not less. The better the model, the more its drift is rewarded by the world, and the more you have to deliberately hold your own ground if you want your output to have a center.
This happens person by person before it happens institutionally
The institutional version of all this — companies redesigning their work to assume AI in the room — is going to take years. Org charts move slowly. Hiring moves slowly. Performance reviews move slowly.
But the individual version is happening, person by person, in the companies I've been watching up close. There is an M on most teams I look at. Sometimes the team has noticed her and is trying to figure out what to do about it. More often they have not, because by every legible measure of the role she's just doing what an analyst is supposed to do — only a little faster, with a little more leverage, in a way nobody can quite name.
The transitional period — the period where being AI-native is still a competitive advantage rather than a baseline — is now. It will not last forever. Within a decade, this will be how everyone works, the way that working with a laptop is how everyone works today. Right now, the M on every team is doing roughly twice the work of the person sitting next to her, in a way the org chart can't see and the performance review can't measure. Most companies haven't noticed. The ones that have, haven't figured out what to do about it.
Closing
If you are leading a team, the move worth making is not another training program.
It is finding the M on your team and giving her air. Let her work the way she actually works. Let her externalize her loop where others can see it. Let her name the instincts she has built. The skill does not transmit through a workshop. It transmits through proximity.
And pay her accordingly. Her output isn't a senior-analyst output; it isn't a manager output either. It's a category your comp band doesn't have a row for. The market will price it before you do.
AI-native is not a skill badge. It is a way of thinking. And there is already one on your team.