AI questions in performance reviews: how to introduce them without oversight and pressure

Shopify’s CEO admitted that the answers to the AI question in performance reviews did not matter. Meta removed AI metrics from reviews, and Amazon called token counting a deeply flawed measure. We show how to talk about AI in the review cycle so that the conversation actually benefits the team.

Who this guide is for and how long implementation will take

This guide is for managers who conduct performance reviews and want to include a conversation about AI tools in them. It describes two scales. The first is a small team in a small or medium-sized company, where a review is usually a 1:1 conversation, meaning a meeting between a manager and one employee. The second is a team in a corporation with a formal HR questionnaire and a set review-cycle calendar.

After reading, you will know what AI questions to ask, which metrics not to include in a review, how to separate learning tools from evaluating results, and how to document the team’s good practices. Preparation will take, by our estimates, a few days in a small team and a few weeks in an organization where changes to the questionnaire must be approved by HR and often also by the legal department.

Table of contents

You can jump straight to the steps. However, the first three chapters explain why the steps look the way they do.

1. What the Shopify case teaches
2. Metric, gaming, withdrawal: the lesson from Meta and Amazon
3. What the AI Act says about AI in employee evaluation
4. Step 1: decide why you are asking about AI
5. Step 2: ask about outcomes and ways of working, not usage
6. Step 3: separate learning from evaluation
7. Step 4: document the team’s good practices
8. Small team and corporation: two paths
9. Typical mistakes
10. Checklist and next steps

What the Shopify case teaches

In April 2025, Tobi Lütke, CEO of Shopify, wrote in an internal memo that reflexive use of AI is a basic expectation for the company’s employees. He announced the addition of questions about AI use to the performance review and peer review questionnaire. Teams were also supposed to show why AI would not do a given job before asking for new people or resources.

A year and a half later, on the Sources podcast in a conversation with Alex Heath (episode of October 28), Lütke said that the answers to that question “didn’t matter.” In his view, the question itself was meant to give the company a reputation it still had to grow into. He described it as a rating given by colleagues: how reflexively someone comes in with a prototype made with the help of AI. He admitted that with this question, “sometimes people game it a bit.” He did not know whether the question was still in the current review, and he did not say whether it would appear in the next cycle. He described both the question and the rule concerning hiring as something that already works on its own:

In other words: I don’t have to enforce these kinds of things. For a manager, that is an important clue. At Shopify, the question in the questionnaire worked as a cultural signal, not as a measurement. If you transfer it to your own team as a hard review element, you will get something different from what Shopify got.

The second lesson comes from an earlier, September statement by Lütke on the Knowledge Project podcast. He said there that AI had worsened a certain kind of sloppy work in the company. Shopify employees call unchecked AI work passed on to a colleague a “slop grenade.” Search Engine Journal pointed out that a question about how eagerly someone reaches for AI will not tell a leader whether anyone checks AI drafts before they reach an editor or a client. These are two different problems and require two different questions.

The third issue is the stance toward skeptics. Lütke believes that a skeptic who “tries not to stick their head above the parapet” will not become the best employee in a company where everyone around them works with AI agents. In his view, a company should clearly say who it is for and help people find a place where they can be among the best. That is a decision about a specific company’s culture, not a universal rule. Matt Poepsel of The Predictive Index, quoted by SHRM (the American association of HR professionals), believes leaders should not force people to use AI.

AI

Metric, gaming, withdrawal: the lesson from Meta and Amazon

Large technology companies have gone through a similar path in recent months. First they measured and rewarded AI use, then employees began gaming those indicators, and finally the companies withdrew them or limited them.

At Meta, HR chief Janelle Gale announced at the beginning of February 2026 that “AI-driven impact,” meaning the impact achieved thanks to AI, was becoming a basic expectation for all employees, from engineers to marketers. The company used a system called Checkpoint, which collected more than 200 data points per engineer, including the amount of code generated by AI and errors related to that person’s work. Some teams were expected by mid-2026 to generate more than 75 percent of approved code with the help of AI. The highest of the four rating levels entitled employees to a 200 percent bonus multiplier.

In September 2026, after a lawsuit, Meta reportedly removed Checkpoint from reviews. The reason described in the media: linking bonuses to AI impact encouraged engineers to prioritize the quantity of AI-generated code over its quality. The new rules made AI use optional, and work is evaluated by the project’s final impact.

At Amazon, in May 2026, KiroRank was shut down, an informal ranking created by a group of employees that sorted people by their AI token consumption (the units used to bill the work of language models). It turned out that employees were inventing artificial tasks, including with the help of the internal agent MeshClaw, to boost their score. Dave Treadwell, Amazon’s senior vice president, then told employees: “Please don't use AI just for the sake of using AI”, meaning they should not use AI only for the sake of using it.

In a memo dated October 27, 2026, Amazon called token consumption, the pace of AI tool adoption, and the number of code changes deeply flawed measures of AI usage because they lead to higher costs and unnecessary complexity. The company did not discard this data. It instructed managers to read it together with other indicators and to look at both extremes: very low use may mean missed opportunities, while very high use raises the question of whether it is justified by growth in what the team actually delivers.

SHRM described a similar reversal in Duolingo’s case. The company backed away from its 2025 decision to treat AI as a performance metric, and CEO Luis von Ahn said, in SHRM’s paraphrase, that in evaluation the most important thing is doing your job as well as possible. Our editorial conclusion: any AI usage indicator that translates into evaluation and money will, sooner or later, start being gamed by someone.

AI

What the AI Act says about AI in employee evaluation

EU Regulation 2024/1689, the AI Act, classifies AI systems used in employment as high-risk systems. Annex III, point 4 includes, among other things, systems for task allocation, employee monitoring, and performance evaluation. An important distinction: the rules apply to systems in which AI evaluates or monitors. A manager who asks an employee in a conversation how they use AI tools is not deploying a high-risk system. According to legal analyses, a tool that only collects opinions without an AI-issued rating is not one either. It is different with software that independently analyzes people’s activity and assigns them a score.

If your company is considering such a tool, remember three provisions. Article 26(7) requires the employer to inform workers’ representatives and the employees themselves that they are subject to an AI system before it is deployed. Article 86 gives an employee the right to a clear explanation of how the system affected a decision concerning them. Article 5 bans emotion-recognition systems in the workplace, with narrow medical and safety-related exceptions; that ban is already in force.

Also remember that employee monitoring is already regulated today by labor law and the GDPR (the EU General Data Protection Regulation). This guide does not replace legal advice: for any tool that collects data about the work of specific people, involve a lawyer or a data protection officer.

Step 1: decide why you are asking about AI

Before you add anything to the form, define the purpose. It determines whether a question about AI should go into the review at all.

1. Write down in one sentence what you want to learn. For example: “I want to know which tasks the team has already sped up with AI and where it needs training.”
2. Assign the purpose to one of three categories: skills development, sharing practices in the team, or evaluating results.
3. If the purpose is development or sharing practices, do not put the question into the part that affects evaluation and bonus. Move it to the development conversation.
4. If the purpose is performance evaluation, evaluate results, and treat AI as one of the ways to achieve them.
5. Check whether someone will read the answers and do something with them. Remove any question that leads to nothing.

What you should have after this step: a one-sentence purpose and a decision about which part of the review cycle the AI topic will appear in.

Step 2: ask about outcomes and ways of working, not usage

SHRM distinguishes between two types of questions: “does the employee use AI” is a question about adoption, while “how many tokens did they use in the quarter” is an attempt to measure productivity. Use alone does not tell you whether someone is doing it well, and rewarding volume rewards waste. Dustin Snyder of consulting firm Wayforward Associates believes AI use should not be a separate measure in evaluation, because the company would then be rewarding the tool, not the outcome. Jackie Stinnett of O.C. Tanner emphasizes that evaluation should have multiple dimensions rather than rely on a single AI metric.

On that basis, we have prepared a set of questions for the conversation. Choose two or three:

1. In which task during the last period did AI genuinely shorten your work, and by approximately how much?
2. How do you check the result of AI’s work before passing it on to me, a colleague, or a client?
3. Where did AI let you down, and how did you catch it?
4. Which task would you like to try doing with AI in the next period, and what do you need to do it?
5. What AI work practice is worth showing to the rest of the team?
6. In what situations do you consciously not use AI, and why?

The second question directly addresses Shopify’s “slop grenade” problem. The sixth makes space for skeptics: caution toward AI when handling client data or in tasks requiring precision can be a good professional decision rather than a lack of competence.

It is equally important what not to include in the evaluation. Each of the indicators below measures usage, not outcome. Some of them were used by large companies and later abandoned when they encouraged people to game the metric:

• number of prompts (commands entered into the tool),
• time spent in the AI tool,
• token consumption,
• percentage of code or text generated by AI,
• number of code changes or reviews,
• position in the ranking of AI users.

Instead, look at three things HR advisors point to in their recommendations: time or effort saved on a specific task, the quality of the output (AI work requiring major corrections does not count), and the final result rather than the frequency of tool use.

Step 3: separate learning from evaluation

People learn new tools through trial and error. If every attempt can lower their rating, they will stop trying or start pretending. Daniel Zhao, a Glassdoor economist quoted by SHRM, points out that employees connect AI metrics in reviews with fear of job loss. That fear affects what you will hear in the conversation.

1. Plan the AI conversation as a separate agenda item or a separate meeting, outside the part where you discuss evaluation and pay.
2. Say explicitly at the beginning that the answers do not affect evaluation or bonus.
3. Ask about experiments, including unsuccessful ones. An unsuccessful attempt with a lesson learned is a valuable outcome in this conversation.
4. Evaluate work results as before: deadlines, quality, impact on the client or project.
5. Write down what support the employee needs: training, access to a tool, time to learn.
6. Return to those needs in a month and check whether they have been met.

What you should have after this step: two separate conversation notes, a performance evaluation and a development plan in the area of AI.

Step 4: document the team’s good practices

You will gain the most from AI conversations when the conclusions from individual discussions make it into a shared document. The team learns from one another, and you do not have to monitor each person separately.

1. Create one shared document or a page in the team wiki titled “How we work with AI”.
2. For each practice, record: the task, the tool, the method of checking the result, the approximate time saved, and the risks.
3. Add a handoff rule: whoever passes on the result of AI work has read and checked it beforehand.
4. Add a list of data that must not be pasted into external AI tools, agreed with the person responsible for data protection.
5. Record practices without names or with the author’s consent. The document is meant to support learning, not accountability.
6. Review the document once a quarter and remove practices that have stopped working.

Such a document largely replaces what Shopify tried to achieve with the questionnaire question: it shows what expectations the team has, but it is based on examples rather than declarations.

AI

Small team and corporation: two paths

In the materials we used, there is no study or case study showing how AI questions work in small companies. The differences below are the editorial team’s conclusion from comparing the formalities of HR processes in both types of organizations.

In a small team, a review is usually a conversation rather than a form. Two or three questions from step 2 asked in a 1:1 conversation and a short note are enough. The legal risk is lower, because there are usually no tools here that independently analyze people’s work. The risk of personal pressure is higher: in a small team, the manager sees who uses AI and who does not, and it is easy to carry that over into the evaluation. Anonymity is also lower, so keep the practices document in a way that does not make anyone feel exposed.

In a corporation, the review questionnaire is changed by HR, and every new question reaches hundreds or thousands of people. Before you propose a change, establish with HR whether the company uses or plans tools that independently collect data on AI use. If so, a lawyer, a data protection officer, and the workers’ representatives referred to in Article 26(7) of the AI Act should be involved in the process. If such data already exists, do not enter it directly into the evaluation. A more reasonable model is the one Amazon described in its October memo: usage data read together with other indicators, as a starting point for a conversation rather than a score.

In both cases, the same principle applies: evaluation concerns what someone has achieved, and the conversation about AI concerns how they can work better.

Typical mistakes

These mistakes recur in the examples described above and in the recommendations of HR advisors:

1. Putting the volume of AI use into evaluation or bonus. The effect can be seen at Meta and Amazon: quantity crowds out quality, and people invent tasks to fit the metric.
2. Tolerating AI usage rankings in the team, even informal ones. KiroRank emerged from the bottom up, and the company still had to shut it down.
3. Asking only about adoption. The answer “I use it every day” says nothing about whether someone checks the results.
4. Punishing caution. An employee who does not use AI with client data may simply be doing their job well.
5. Copying solutions from huge companies without their context. The question at Shopify was part of a broader cultural shift announced by the CEO.
6. Deploying a tool that analyzes people’s work without informing the team and without legal consultation.
7. Asking a question whose answers no one analyzes.

Checklist and next steps

Before the next review cycle, go through the list below. Skip the items marked as corporate if you manage a small team without analytics tools.

Start with a small step. In the coming week, write down the purpose and choose two questions from step 2. In the next 1:1 conversation, ask them outside the evaluative part and note down the answers. After the first round of conversations, create the practices document and enter the team’s first three examples into it.

If you work in a large organization, schedule a conversation with HR about the questionnaire before planning the next cycle and take this guide and the checklist with you. After the first cycle, check whether the practices document is alive and whether people add examples to it without reminders. If they do, you have implemented the AI topic in your team better than with any metric.