Analysis · 2026-08-04 · 8 min read

AI's Leadership Trap: Why Token Counts Make Terrible Performance Reviews

Some companies now tie promotions to AI token usage — the newest 'tokenmaxxing' failure mode. Why measuring developers by token consumption repeats the lines-of-code mistake, what Goodhart's Law predicts, and the outcome-based KPIs that actually correlate with business value.

TL;DR

A Forbes report describes companies rewarding developers for maximizing AI token usage, linking it to promotions and job security.
It's the lines-of-code mistake reborn: rewarding resource consumption instead of business value, creativity, or customer impact.
Goodhart's Law guarantees the outcome — once token counts become the target, developers burn tokens to game the score, not to ship better products.
Research from Nature Machine Intelligence and the NBER backs the same conclusion: AI inflates activity far faster than it inflates delivered value.
The companies winning right now measure cost per merged PR, cost per resolved ticket, and cost per shipped feature — not tokens consumed.

The Metric That Rewards Burning Money

Forbes contributor Michael Edmondson reported a management practice that should set off alarms in any engineering org: a manager instructed developers to maximize AI token usage because promotions and long-term job security would depend on it.

The policy misunderstands what AI tokens are. A token is a unit of resource consumption — like gallons of fuel, kilowatt-hours, or cloud compute minutes. Rewarding engineers for burning more of them is like rewarding delivery drivers for using the most gasoline. Fuel is a *cost* of creating value. It is not the value itself.

!OpenAI logo

The Lines-of-Code Mistake, Repeated

This isn't a new failure mode. Business history is full of leaders who measured the wrong thing:

Manufacturers rewarded production volume before quality mattered.
Call centers celebrated shorter conversations before first-call resolution won.
Software companies judged programmers by lines of code until elegant solutions — which usually had *fewer* lines — proved better.

Token counts are lines of code with a cloud bill attached. They reveal nothing about creativity, engineering judgment, collaboration, security, or customer impact. Measuring a developer by token consumption makes about as much sense as judging a carpenter by the number of nails in the table.

Goodhart's Law Predicts the Damage

Economist Charles Goodhart framed the principle decades ago: once a measure becomes a target, it ceases to be a good measure. People change their behavior to improve the score, and the score stops reflecting what you actually wanted.

Developers rewarded for AI consumption will predictably consume more AI. Longer prompts, unnecessary experimentation, and frontier models for tasks a $0.27/M model could handle quickly become rational career decisions. Every extra interaction increases operating costs while providing little assurance of stronger products or happier customers. The organization begins rewarding computational activity instead of engineering excellence.

!Anthropic Claude logo

The Research Says the Same Thing

Two pieces of evidence point the same direction:

Nature Machine Intelligence: deploy with purpose, not consumption

An editorial in *Nature Machine Intelligence* warns organizations against treating token usage as a measure of productivity. AI, the authors argue, should be deployed with purpose rather than measured through consumption. They describe "tokenmaxxing" as another example of confusing technology usage with meaningful organizational performance.

NBER: activity inflates faster than delivered value

Researchers at the National Bureau of Economic Research found AI dramatically increases coding activity, yet far smaller gains appear once projects move through testing, integration, deployment, and customer delivery. Great software still depends on architecture, review, collaboration, security, and thoughtful decision-making. AI accelerates pieces of the journey; people remain responsible for reaching the destination.

The Irony: Leaders Are Already Cutting Tokens

The most ironic part of the token-as-KPI trend is that the most sophisticated AI users are moving in the opposite direction. Companies increasingly route routine requests through smaller, less expensive models while reserving frontier systems for genuinely complex work.

EY reported significant AI operating-cost reductions after deploying intelligent routing that matches each request with the smallest model capable of high-quality output — lower token consumption, same business performance.
Coinbase cut its AI bill roughly in half with the same pattern: cheap open-weight models for the 80% of tasks, frontier calls reserved for the 20% that need them.
AP News reported companies shifting focus from maximizing AI *usage* toward maximizing AI *value* as compute costs rise.

As Tim Keary noted in *Forbes*: *"In an economic climate with greater board scrutiny, token usage without demonstrable business impact runs the risk of coming across as conspicuous consumption."*

!Meta logo

What to Measure Instead

The organizations gaining the greatest advantage from AI share one characteristic: they measure AI by the business value it creates, not the amount employees use it. One asks whether AI was *used*. The other asks whether AI made the organization *better*.

The KPIs that actually correlate with outcomes:

Gallup found employees are "substantially more likely to be engaged when leaders introduce AI with clear expectations, a thoughtful implementation plan and active manager support." Employees create the greatest value when AI removes repetitive work, freeing time for creativity, collaboration, problem-solving, and customer engagement. Those outcomes strengthen organizations. Token counts just increase cloud bills.

The Self-Test Every Leader Should Run

Edmondson closes with a question worth pinning to every engineering dashboard:

> If artificial intelligence disappeared tomorrow, would your current performance measures still identify your highest-performing employees?

If the answer is no — if your metric only "works" because a language model is in the loop — you're measuring the technology, not the results the technology makes possible. Organizations optimizing dashboards will spend years chasing tokens. Organizations optimizing outcomes will build stronger products, stronger cultures, and stronger businesses.

Customers never purchase AI tokens. They purchase the value people create with them.