The Most Valuable Work I Do for My Employer Uses Zero Tokens.

The most valuable work I do for my employer would show nothing at all on a dashboard. Every morning, when I first get to my desk, I spend anywhere from ten to thirty minutes on a yellow legal pad. I write stream-of-consciousness sentences, real ink on real paper, and I keep writing through the day as topics come to mind: from work, from what I read on LinkedIn, from ideas that come into my head. I started this long before AI existed; it is so much better for the brain.

That pad is where I work through my thinking for my day job, and it is what keeps things coordinated and moving toward long-term objectives. None of it goes into an AI tool directly; parts of it may end up in a prompt later, or become the basis from which I write one. Nobody is monitoring my AI usage yet, and I imagine a usage rating would completely miss that my AI usage is efficient because I already know how to describe exactly what I want before I ever touch a keyboard. Yesterday's piece was about the budget line that prices every seat by its output. This one goes a step further back, to the rating that prices every person by what they consume.

I am sure a usage rating is coming. I have watched this repeat throughout my career, and every year there is something new. I need to be ahead of this problem before it arrives, with better metrics I can answer with. None of this is an argument against the tools; I run my publishing on them, and I use them for my day job as well.

This year's new thing was AI usage in performance reviews, and the most-reported case has already backed away from it. Business Insider reported early this year that Meta tracked the code each engineer produced with AI among more than 200 data points, and that one Meta organization set a target for 65% of its engineers to have more than 75% of their code written with AI, though the coverage said it was unclear whether that target was tied to formal reviews. Coverage described "tokenmaxxing," people burning tokens to look good on the numbers. In September, WIRED reported that Meta's new review guidance no longer lists AI usage, and according to The Information, a memo told engineers Meta "will not use AI adoption dashboards or token counts to evaluate impact."

Very few people reading this are at a Meta level. That is great for Meta, which can afford to measure 200 things and then change its mind; what are the rest of us going to do? What travels is the headline, not the 200 data points or the reversal. JPMorgan, per Business Insider, labels engineers as light or heavy users of AI tools, and that use may influence their reviews. Usage is all most companies are likely to count, so usage is what lands on the form.

A rating tells people what gets rewarded under pressure, and calibration is the pressure, so people move toward the number. That is not cynicism; it is what a rating is for, and tokenmaxxing is what it looks like when the number is consumption. Usage measures input, and generated work is not finished work: LinearB's 2026 benchmarks, drawn from 8.1 million pull requests, show AI pull requests merging within 30 days 32.7% of the time, against 84.4% for unassisted work.

The highest-judgment work on any team uses zero tokens. My morning on the pad, the decision not to build something, the conversation that kills a bad idea before it reaches a sprint, and the careful review that sends a change back all score as nothing. The person doing that work scores low. Think about the engineer on your team you would trust most to kill a bad idea: how much AI did they use last month?

Under a usage rating, the risk is that the team learns to avoid zero-token work: thinking moves into the tool or disappears, fewer ideas are killed, and reviewers approve faster to get back to generating. If calibration then promotes the heaviest consumers, next year's leads are selected for consumption rather than judgment. That is what I named the Rewards-and-Penalties Inversion, one level down: the rating that picks next year's leads rewards the behavior that damages the team, and the structural work leaves no evidence at all.

The leaders who come through this with their best people intact do not out-argue the rating; they propose the measure before one is handed to them. Before your company's 2027 review form gains an AI line, put forward the measures you can answer for: what the work was aimed at and what it delivered against that aim, the review load your people absorbed, and the work they decided not to build. Keep adoption where it belongs, as a team-level rollout number with an end date, never as a line on anyone's individual rating. This cycle, you walk into calibration with measures you chose, or you inherit the one that was easiest to count.

Yesterday's two lines were for a seat. This week, write them for people: one page for your team, three lines per person: what their work was aimed at this quarter, what it delivered against that aim, and one thing they decided not to build and why. Take it to your manager as the measure you want your team held to, before anyone asks for a usage number. Then pull out the last review form your team was rated on and ask of each line: can someone score well on this without making a single good decision? Every line where the answer is yes is the line your team is learning from.

The LeadershipOS™ Inner Circle is for leaders who would rather propose the measure than inherit it: this depth and deeper, every month, in print: https://LeadershipOSInnerCircle.com