Your employer is grading your AI use. What to do before the next review

If your employer has started grading your AI use, the review will count one thing and reward another. It will count usage: logins, prompts, tokens, seats. It will reward impact: work that got faster, cheaper or better because you changed how it is done. The gap between the two is where this year's ratings are decided, and you can close it before your next review with a record that takes twenty minutes a week to keep.

This is now written policy at large employers. Meta made "AI-driven impact" a formal part of 2026 reviews for every role. KPMG is rating staff on its AI objectives in the 2026 cycle. Accenture requires "consistent use" of AI tools for senior promotions. A General Assembly survey of more than 500 senior leaders in the US and the UK, reported by Newsweek in May 2026, found that 47% already factor AI usage into performance evaluations.

The same survey carries the warning. Daniele Grassi, General Assembly's chief executive, said leaders "were more likely to consider how much employees were using AI tools than any actual impact that usage had". So the measurement is crude, and the reward goes to the person who hands their manager a better one.

This guide is the method: find out what is measured, build the evidence, redesign one workflow, write the review entry, and know what the law says about the systems doing the grading. The cluster of War Room pieces under it carries the detail; this page carries the whole.

Who is grading, and on what

| Employer | What counts | Reported by | |---|---|---| | Meta | "AI-driven impact", every role, formal from 2026; dashboards track adoption by team | eWeek, Fortune | | KPMG | Progress against the firm's AI objectives, in 2026 reviews; usage data from tools such as Microsoft Copilot | Bloomberg, International Accounting Bulletin | | Accenture | "Consistent use" of AI tools to qualify for senior promotions; weekly logins monitored | Financial Times, Fortune | | Microsoft | AI use as part of "holistic reflections on an individual's performance and impact" | Business Insider, via Entrepreneur | | Amazon | A target of more than 80% of developers using AI tools each week, tracked on leaderboards; Amazon says usage does not feed reviews | Financial Times, via The Decoder |

Meta's phasing matters. Employees documented AI-driven impact through 2025, and it became a formal component in 2026. Meta's head of people, Janelle Gale, put it as "we want to recognize people who are helping us get there faster", and Fortune reports she told staff that AI-driven impact would be a "core expectation" in 2026, under a review system that pays top performers bonuses of up to 200%. Meta also gave staff an "AI Performance Assistant" to help draft self-reviews. The grading and the drafting are both done with the tool being graded. What Meta is actually scoring is set out in our briefing.

KPMG's global head of risk services, Samantha Gloede, framed the monitoring directly: "Monitoring is not for policing's sake, we need to make sure that all staff are using these tools because that is the best way to do the jobs."

Microsoft's version arrived as an email from Julia Liuson, president of its developer division, reported by Business Insider in June 2025: "using AI is no longer optional — it's core to every role and every level."

Accenture's policy has an exemption worth noting. Fortune reports that staff in 12 European countries, on US federal contracts and in certain joint ventures do not have AI usage factored into promotions. In those places the metric does not apply.

The other direction matters too. Your manager is probably using AI on you. A 2025 Resume Builder survey of more than 1,300 managers, reported by Reworked, found that 91% use AI to assess performance and 88% to write performance improvement plans.

Usage is what gets counted. Impact is what gets rewarded

Usage metrics produce usage. In April 2026 The Information reported an internal Meta leaderboard, built by an employee and called "Claudeonomics", that ranked token consumption across more than 85,000 staff. Fortune reported that employees burned through more than 60 trillion tokens in 30 days, that the top user averaged 281 billion a month, and that some left agents running for hours to climb the table. The dashboard came down two days after the story ran.

At Amazon, the Financial Times reported in May 2026 that some developers were using MeshClaw, an in-house agent platform, to inflate their numbers. One employee told the paper: "There is just so much pressure to use these tools. Some people are just using MeshClaw to maximise their token usage." Another: "Managers are looking at it. When they track usage it creates perverse incentives."

Read those two stories as a warning about where the reward goes. A leaderboard position lasts until finance reads the invoice. At the next calibration meeting, a manager has to defend your rating to peers and to people above them. A token count is hard to defend in that room. A before-and-after on a process the team relies on defends itself.

So the operator move is to satisfy the usage metric at the level your employer expects, and to spend the real effort on the impact record. Clear the bar on usage, then compete on evidence.

Step 1: find out what is measured (thirty minutes)

Most people guess. Ask instead, in writing, so the answer can be quoted back in the review.

Send your manager, or HR if the policy is company-wide, four questions:

1. Which AI tools count, and is usage from other approved tools included?
2. Is usage measured automatically, and if so, what data is collected about me and who sees it?
3. Is the review scoring usage, impact, or both, and how is impact evidenced?
4. Is any part of my rating produced or suggested by an AI system?

The first three tell you what to optimise. The fourth tells you which rules apply, because the law treats a decision a system makes about you differently from a decision a person makes with a system's help. If the answer to the fourth is yes, keep the reply. Step 5 explains why.

If there is no written policy, that is also an answer. It means the standard is being set by whoever submits strong entries first. Meta ran exactly this sequence: a year in which employees documented their AI wins, followed by a year in which they were formally scored.

Step 2: keep an evidence file (twenty minutes a week)

Reviews are written from memory, and memory keeps the last month. An evidence file keeps the year. One line per change, five fields. Two example lines:

| Date | Task | Before | After | Who noticed | |---|---|---|---|---| | 12 Mar | Weekly client status report | 3 hours, drafted by hand from four trackers | 40 minutes: model drafts from exports, I check figures and write the risk section | Account lead used it unchanged for two clients | | 2 Apr | First-pass contract review | Read in full, about 90 minutes each | Model flags clauses against our playbook in 10 minutes; I read the flagged sections and the schedules | Legal adopted the checklist for the team |

Three rules make the file usable in a review.

Measure time and quality, not effort. "Saved two hours a week" is a claim. "Three hours to forty minutes, same template, no corrections from the account lead in six weeks" is evidence.

Record what you checked. Every line should say what the model did and what you verified. That is the part a manager can defend, and it protects you when an output goes wrong. The accountability for an AI-assisted decision lands on the person who signed it.

Name a witness. A result someone else relied on outweighs a result you describe. The last column turns your note into a reference.

Step 3: redesign one workflow (one month)

Using a model to draft faster produces a cheaper version of the old process. Removing steps from the process produces a new one, and a new process is what "AI-driven impact" means in practice.

Pick one workflow that meets three tests. It recurs weekly or more. Other people depend on its output. And you can measure it before you change it. Then run a month:

• Week 1. Measure the current process: steps, hours, error rate, handoffs. Write it down before touching anything.
• Week 2. Remove or merge steps. Let the model do the assembly; keep the judgement calls with a person, and say which ones they are.
• Week 3. Run old and new side by side on real work. Log where the new version was wrong and what caught it.
• Week 4. Write it up in one page: the before, the after, the checks, and who now uses it.

A one-page write-up of a changed process is the strongest entry you can bring to a review this year. It also travels. It is the same artefact a hiring panel asks for when it wants evidence that you use AI well, which is where the hiring filter now sits.

If you are unsure which of your workflows to pick, the task ledger in our role guide sorts a week of work into what a system can do, what it can help with, and what stays yours. Choose from the middle pile.

Step 4: write the review entry

Self-review forms reward a fixed shape: the change, the measure, the check, the effect on others. Keep each entry to four sentences. An example, built from the first line of the file above:

I rebuilt the weekly client status report so a model assembles the draft from our four tracker exports. Preparation fell from about three hours to forty minutes a week, measured over eight weeks. I verify every figure against the source export and write the risk section myself; two errors were caught this way and corrected before sending. The account lead now uses the format for two further clients.

That entry clears a usage metric, because it names the tool in routine use. It also clears an impact metric, because it has a baseline, a result, a control and an adopter. And it holds up in calibration, because every claim can be checked.

Avoid three things. Tool lists ("I use Copilot, ChatGPT and Gemini daily") describe inputs. Unmeasured claims ("significantly more efficient") invite the question you cannot answer. And percentages without a baseline read as invented, even when they are not.

Step 5: know what the law says about the grading

This section describes the rules as they stood in September 2026. It is general information, not legal advice. If a decision about you turns on it, speak to your union, works council or an employment lawyer.

In the EU, three rules apply now.

• Solely automated decisions. Article 22 of the GDPR gives you "the right not to be subject to a decision based solely on automated processing" that significantly affects you. Where such a decision is allowed, the employer must provide "at least the right to obtain human intervention", to "express his or her point of view and to contest the decision". A rating, a promotion refusal or a performance plan produced by a system with no real human review falls inside this.
• Your data. Article 15 gives you the right to a copy of the personal data processed about you, including "the existence of automated decision-making" and "meaningful information about the logic involved". Usage logs and dashboard entries about you are personal data. You can ask for them.
• Emotion recognition. Since 2 February 2025, Article 5(1)(f) of the AI Act bans AI systems that infer the emotions of people at work, except for medical or safety reasons. Commission guidance, as summarised by the Future of Privacy Forum, treats inferred attitudes the same as emotions, so a tool that reads your mood from your face or voice is prohibited in the EU whatever the vendor calls the score.

One more EU rule is coming, later than planned. The AI Act lists AI used for performance evaluation, task allocation, monitoring and promotion or termination decisions as high-risk. For those systems, Article 26(7) requires employers to inform workers and their representatives before deployment, and Article 26(11) requires them to tell people that a high-risk system is being used in decisions about them. The Digital Omnibus, Regulation (EU) 2026/1744, moved those obligations from August 2026 to 2 December 2027. Until then the documentation and oversight duties do not apply to the systems that grade you, though the GDPR rules above do. What moved and what did not is in our briefing.

In the UK, the Information Commissioner's Office guidance on the UK GDPR sets out the same safeguards for significant decisions based solely on automated processing: meaningful information about the logic, human intervention, the right to express your point of view, and the right to obtain an explanation and challenge the decision.

In the US, there is no federal rule, and state law is patchy.

• Illinois. Since 1 January 2026, an amendment to the Illinois Human Rights Act (HB 3773) bars employers from using AI that has a discriminatory effect in decisions on promotion, discipline, discharge and the terms of employment, and requires them to notify employees when AI is used in those decisions.
• Colorado. SB 26-189, signed on 14 May 2026, applies from 1 January 2027 to automated systems used in consequential decisions, including employment. After an adverse outcome, a worker is owed notice, a plain-language description of the system's role within 30 days, a chance to correct the data and a human review.
• Everywhere else, existing anti-discrimination law still applies to a decision made with AI. The former head of the Federal Trade Commission, Lina Khan, argues that existing laws already reach the people who deploy these systems.

If the grade looks wrong

Work through it in order, in writing, and keep copies.

1. Ask what the rating was based on. Which data, which period, and whether a system produced or suggested any part of it.
2. Ask for the data. In the EU and the UK, make a subject access request for the usage data and any automated scores held about you. Elsewhere, ask HR for the same information under the company's own policy.
3. Put your evidence file next to their data. Most disputed AI ratings are usage counts read as performance. Your before-and-after lines are the correction.
4. Ask for human review. Where a system decided, the GDPR, the UK GDPR and, from 2027, Colorado give you a route to a person. Elsewhere, ask for it anyway, and ask who that person is.
5. Write your point of view down. Short, factual, dated. It becomes part of the record the next reviewer reads.

If you manage people

The same method works from the other side of the table. Grade the evidence file, not the dashboard. Ask each person for one redesigned workflow a year, with a baseline and a check. Tell your team in writing which tools count and whether usage is logged. And if a system suggests a rating, read the inputs before you sign it, because your name is on the review.

Where you stand, by stage

• Early. You use AI occasionally and have no record. Send the four questions this week and start the evidence file with whatever you did last month.
• Aware. You use AI routinely and can describe the gains. Measure one workflow before your next change, so the gain has a baseline.
• Operational. You have a redesigned workflow and a log. Get a colleague to adopt it, and write the one-page account.
• Command Level. Others use your process. Offer it to your manager as a team standard, and help write the measure your team is graded on.

To place yourself properly, take the seven-question AI readiness assessment. It takes about two minutes.

Questions readers ask

Should I refuse to use AI if I think the metric is unfair? Refusing is visible, and it concedes the standard to others. A better route is to meet the expected usage level and argue about the measure with evidence in hand. People who can show impact get listened to when they say usage is the wrong metric.

My employer says usage is not in reviews. Should I believe it? Amazon said the same, and its employees told the Financial Times that managers were looking. Treat any data that is collected as data that can be read. Keep usage at the expected level and let your evidence file carry the case.

Does using AI to write my self-review count against me? Meta built a tool for exactly that. What counts against you is a self-review that could have been written without knowing your work. Draft with the tool if it helps; the measures and the checks have to be yours.

What if my job has little that AI can help with? Then the grading question is mostly about the handful of tasks where it can: reporting, research, first drafts, scheduling. Pick the one that recurs most and run Step 3 on it. One documented change answers the question for the year.