Should AI Usage Be Part of Performance Reviews?

Yes, as a question the manager asks and the person answers, and never as a score. Pressure to look like an AI user is already common: in a survey by Visier of 1,000 US full-time employees, published September 22, 2026, 48% said they had exaggerated their AI use or expertise to colleagues or leadership. Read beside what the role’s work calls for and what the person produced, a usage figure gives a manager a fair question to ask.

Companies have tried both answers. Shopify and Duolingo wrote AI use into their reviews in 2025, and in April 2026 Duolingo’s chief executive said the company no longer considers it. Below is who said what, and then how to raise AI use in a review without turning it into a score.

What have companies said about AI use in performance reviews?

Three rows report a company’s own words: the Shopify memo, which its chief executive posted, the Duolingo email, which the company posted, and Amazon’s published guidance. We read the first two through BetaKit and Snopes, and Amazon’s guidance directly. Duolingo’s 2026 change came from its chief executive on a podcast. The Microsoft and Meta rows rest on internal documents that reporters saw, and the Amazon leaderboard account is Financial Times reporting that we read only as relayed by Fortune. The companies did not publish any of those four themselves.

Read in date order, the statements move toward output. The 2025 memos made AI use an expectation. The September 2026 Meta memo, as reported, tells managers to look at “output quality, velocity, problem complexity and scope taken on.” Amazon keeps its usage measures. Its guidance says they “correlate well with the outputs that matter” and that using them with care is appropriate as staff think about their careers.

Why does a token count make a poor performance score?

A token is the unit AI vendors bill by, roughly a piece of a word, so a token count measures how much AI a person used. Amazon’s guidance says such input measures are flawed “primarily because they are disconnected from what matters,” and that used “as a target instead of a measure” they could drive negative outcomes.

People respond to a count once they know it is watched. According to Financial Times reporting relayed by Fortune, Amazon employees ran an internal AI tool “on trivial tasks to inflate their token counts” and climb a leaderboard. In the Visier survey, 45% of employees said they feel pressure to use AI at work, and 48% said they had exaggerated their AI use or expertise. A score built on the count rewards both behaviors.

The count also has gaps that have nothing to do with the person. Microsoft’s per-user Copilot usage report lists the date each user was last active in each app and has no token field, so it cannot give a token count for a person who works mainly in Microsoft 365 Copilot. Anthropic’s analytics documentation gives per-user token usage and cost on usage-based Enterprise plans and only usage credits on seat-based plans. It does not return Claude Code used through Amazon Bedrock, and with customer-managed encryption keys some Claude Code and Cowork activity figures, such as lines of code and commits, “can be zero or lower than actual use.” A low number can be a gap in the report.

The same count means different things in different roles. A software developer and a field inspector will not use the same amount of AI, and the gap says nothing about either one’s effort. Amazon’s guidance makes the point from the employee’s side: when a person’s work does not benefit from current AI tools, it calls that “a finding, not a failing.”

How should AI use come into a performance review?

The count comes into the review as a question, with the figures that give it meaning. The manager brings the role’s figures: how much of the role’s work AI can do, judged task by task, what the team used and what the team produced. Beside them sits the person’s own usage from the company’s AI tools. When that usage is far from what the role’s work calls for and the person’s output has not moved, the manager asks why and the person answers.

The question runs in both directions. Heavy use with flat output prompts a question about what the AI was used for. Low use in a role where much of the work suits AI deserves the same question, and the answer is often practical, such as missing access to a suitable tool or missing training. Sometimes the person’s share of the work is the part current tools do not suit, which is the case Amazon’s guidance calls “a finding, not a failing.”

What follows from the answer is an action, such as training or a different tool, or a note that the work does not call for more AI. The rating itself rests on the work the person produced, judged by the standard the role already uses. A person makes the decision.

JobRoute supplies the role’s figures: the share of the role’s work AI can do, the team’s use and the team’s output, as described on AI spend by role. The company holds each person’s usage in its own AI tools, and JobRoute does not score or rank people.

What should employees be told and shown?

The question is only fair if the person has seen the figures first.

1. Before the review cycle, tell people in writing that AI use will come up as a question, which tools the figures come from, that the role’s figures are reported for groups, and that nobody is scored or ranked on usage.
2. Before the meeting, show each person their own usage as the company’s tools record it, beside the role’s figures.
3. Let them correct it. Leave, shared projects, a tool that reports only activity and a different mix of work all change what a count means.
4. In the meeting, ask the question, and write down the answer and what follows from it, such as training or access to a tool.
5. Offer a second look by someone with the authority to change the outcome.

Which rules may apply to AI usage data in reviews?

The rules tend to turn on two things: whether a system monitors or evaluates individual workers, and whether its output decides or substantially shapes pay, promotion or discipline. Some turn the steps above, such as notice before use, into duties. Four that bear on reviews:

• European Union. The AI Act lists AI systems intended to be used “to monitor and evaluate the performance and behaviour” of workers among the high-risk uses in Annex III, point 4(b). The EU’s political agreement of May 2026 set December 2, 2027 as the date the rules for high-risk systems in employment apply, and the Council gave final approval in June. Article 26(7) then requires employers to inform workers’ representatives and the affected workers before using such a system.
• Germany. Section 87(1) number 6 of the Works Constitution Act gives the works council a right of co-determination over the introduction and use of technical systems designed to monitor employees’ behavior or performance, where no statute or collective agreement already settles the matter.
• California. The privacy agency’s regulations cover technology that replaces or substantially replaces human decisionmaking in significant decisions, which include pay, bonuses, promotion and demotion. Under section 7001(e), human involvement means a reviewer who knows how to interpret the output, weighs other relevant information and has the authority to change the decision. A business already using such technology must meet these rules by January 1, 2027 (section 7200(b)), and section 7150(b)(4) requires a risk assessment for automated processing used to infer an employee’s performance at work from systematic observation.
• Colorado. SB 26-189 covers technology whose scores or rankings materially influence a decision about an individual, including compensation in employment, and its definitions include employees. For decisions from January 1, 2027, the deployer must give notice and, after an adverse outcome, an explanation within 30 days. The person can request meaningful human review and the correction of factually incorrect personal data.

Our AI and employment compliance calendar lists these and other rules with dates, including Illinois and New York, and our paper on measuring AI use without monitoring individuals covers each in more depth. This is general information, not legal advice.