Where we are
You now know the rating instrument is noisy and that bias travels in language. AI can genuinely help with both problems - it never forgets what happened in March, and it is very good at spotting that one person's review is all adjectives while another's is all evidence. But the same tool is one prompt away from producing a confident verdict on a human being from thin data, and a confident verdict is exactly what a biased system produces most easily. So the rule for this session is a single line, repeated: AI works on the evidence; the human makes the judgment. Everything else here is machinery to keep that line intact - guardrails, disclosure, a never-automate list, and a legal floor that already makes parts of it mandatory.
What AI may touch, and what it must not 6 min live
Do not think about AI in performance measurement as a capability question. Think about it as two lanes with a hard wall between them: work on the evidence, and judgment about a person. AI belongs entirely in the first.
LiveThree uses that genuinely reduce bias3 min▶
These three are not productivity tricks. Each one attacks a specific bias you met in b7:
- Evidence assembly beats recency. Recency bias wins because the last month is the easiest to remember. Ask AI to pull the full period's notes, tickets, and check-in records into a dated list, and the whole year is back on the table.
- Language checking beats the specificity gap. Have it flag every personality label, every unfalsifiable claim, and every sentence with no evidence attached - then compare across your whole team. Who got adjectives and who got numbers becomes visible in one pass.
- A first draft beats the blank page. The blank page is where managers fall back on impressions. A draft built strictly from your evidence list gives you something to correct rather than something to invent. You then rewrite it, and every word is yours.
Notice what all three have in common: the AI never sees a conclusion it is asked to reach. It handles the material; you carry the judgment.
Self-studySurveillance is not measurement2 min read▶
The fastest way to wreck a measurement system with AI is to point it at activity streams because they are easy to collect. Two well-documented cautionary cases:
- Microsoft's Productivity Score (2020) drew a public backlash for surfacing per-employee activity data to managers. Microsoft changed the product to remove individual-level identification. The lesson: the moment a metric can be pointed at a named person's keystrokes, it stops being a steering signal and becomes monitoring.
- Amazon warehouse "time off task" reporting, as reported by MIT Technology Review and others, drove automated productivity tracking that could feed termination processes. The lesson: automating the pipeline from activity data to a consequence removes exactly the human judgment that made it defensible.
Activity streams measure presence, not contribution - they are the bottom rung of the b1 ladder, dressed up in better data. If you would not grade someone on it manually, do not let a tool grade them on it automatically.
Four guardrails, and the law underneath them 5 min live
The four guardrails are good practice on their own. What makes them non-negotiable is that regulators have already turned parts of them into requirements for exactly this use case - AI used to manage and evaluate workers.
LiveHuman in the loop means a real decision, not a signature3 min▶
The guardrail that gets faked most often is human oversight. A manager clicking "approve" on twelve AI-generated ratings in four minutes is not a human in the loop; that is automation with a signature attached, and it is the shape GDPR Article 22 is aimed at.
- The reviewer must be able to disagree - and there must be evidence that sometimes they do. If nothing is ever overridden, oversight is decorative.
- The reviewer must see the inputs, not just the output. A conclusion with no visible evidence cannot be meaningfully checked.
- The reviewer must have time. Oversight that is impossible within the time allotted is oversight on paper only.
- The reviewer must be accountable. Whoever signs owns the judgment - "the tool said so" is not a defence you would want to give to the person affected, or to a regulator.
The four-minute review round. A Cartwheel manager with fourteen reports uses AI to draft all fourteen reviews from her notes, skims them, and submits. Two are subtly wrong: one credits work another person did, one repeats a personality label from her own old notes. Both go into the record unchallenged. The tool did not cause the failure - the missing step did. Draft with AI, then read every line against your evidence as if you were the person receiving it.
Self-studyWhat the bias audit actually looks like2 min read▶
"Bias audit" sounds like a specialist exercise. At a team level it is closer to arithmetic, and you can do a useful version yourself:
- Compare outcomes by group - ratings, promotion rates, and stretch-project assignments split by gender, tenure, team, and working pattern (part-time and remote people are a common blind spot). Small numbers will not be conclusive; patterns repeating across cycles are.
- Compare language by group - the b7 audit, counted: evidence-backed statements versus adjectives, per person.
- Compare the tool's suggestions to your decisions - where you overrode the draft, and for whom. Systematic overrides in one direction tell you something about the tool; none at all tell you something about the oversight.
- Write down what you found and what you changed. An audit nobody acts on is documentation of a known problem, which is worse than not looking.
For anything vendor-supplied and consequential, ask the vendor for their bias-audit results and the data behind them. Under NYC Local Law 144 that expectation is a legal requirement for covered tools, and it is a reasonable question anywhere.
The floor is not the same everywhere, and the dates move - check before you rely on any of this. EU: worker-management AI is still classified high-risk under Annex III, but the AI Omnibus (in force 27 July 2026) moved the compliance duties for standalone Annex III systems from 2 August 2026 to 2 December 2027. Treat that as build time, not a pass. New York City: Local Law 144 has been in effect since January 2023 and enforced since July 2023, so the annual audit, the published summary, and the ten business days' notice are live obligations today. UK: no AI Act, no high-risk tier - the duties come through UK GDPR and ICO guidance, and the Data (Use and Access) Act 2025 replaced the old near-prohibition on solely automated significant decisions with a safeguard-led regime. A team spanning the EU and the UK can therefore run one tool under two different rulebooks. Summarised for awareness, not legal advice.
Build a Cartwheel review pack with AI ★ 12 min · everyone tries
One Cartwheel analyst, one review cycle, a scattered pile of evidence. Use AI for the four jobs it is good at, and keep the judgment jobs in your own hands - then check the pack before it goes anywhere.
Collect the raw material yourself: check-in notes, OKR score history, the shipped work, peer comments. AI cannot fix evidence you never gathered.
Ask it to organise, not conclude: a dated list of what happened across the full period, grouped by the criteria you are rating against.
Ask it to flag the language in your own draft notes: personality labels, unfalsifiable claims, anything with no evidence behind it.
Ask for a first draft strictly from that evidence list, with a marker anywhere it had to infer. Then rewrite it in your voice.
Do the judgment yourself: the rating, the level call, the pay input. Read the final pack line by line against your evidence, as the person receiving it.
| Job | AI | You | Why the split |
|---|---|---|---|
| Assemble the year's evidence | ✓ does it | supplies + verifies | Beats recency; it never forgets March |
| Flag vague or biased wording | ✓ flags | decides each fix | Pattern-spotting at scale, judgment stays human |
| Draft the narrative | ◐ first pass | rewrites + owns | Beats the blank page where impressions live |
| Set the rating | ✗ never | ✓ decides | Judgment about a person - the decide lane |
| Pay or promotion input | ✗ never | ✓ decides | Significant effect - human required |
| Score the OKR | ✗ never | ✓ with the team | Scoring is a learning conversation, not an output |
Write the never-automate list and the disclosure ★ 10 min · do your own
Two short documents make the guardrails real: a list your team can hold you to, and a plain-English disclosure that turns transparency from an intention into something people have actually read.
Write the never-automate list: every decision AI may not make in your team, stated as decisions, not tools.
Write the disclosure: what you use AI for, what data goes in, what it never decides, and how someone contests an outcome.
Add the contest route: who to go to, what they get to see, and what happens next. A right to challenge with no named path is not a right.
Share both with the team and ask what is missing. If reading the disclosure makes anyone uneasy, the practice needs changing, not the wording.
| Document | Contents | Test it passes |
|---|---|---|
| Never-automate list | Ratings, pay and promotion, ranking, exit flags, OKR scores, anything about a named person's activity stream | Every item is a decision, not a product name |
| Disclosure | Uses (evidence assembly, language flagging, drafting), data used, what AI never decides, who decides instead | A team member could explain it back to you |
| Contest route | Who to raise it with, what evidence they see, response time, what can change | Someone could actually use it this week |
| Audit note | What you checked by group, what you found, what you changed | It has a date and an action, not just a finding |
The disclosure that changed the practice. A team lead drafting this document got as far as "I use AI to summarise your check-in notes and draft your review" and stopped - because she realised her reports had never been told that their notes were being used that way at all. The document did its job before anyone read it. That is the useful property of writing transparency down: anything you would not want to disclose is the thing to stop doing, and you find it while drafting rather than after a complaint.
Try it yourself - this week ◐ 30-45 min total
- Write your never-automate list. Keep it to decisions, and keep it short enough that people remember it.
- Draft the AI-use disclosure for your team in plain language, then read it as if you were a report receiving it. Fix whatever makes you wince.
- Take one review you have already written and ask AI only to flag vague, unfalsifiable, or personality-based lines. Decide each fix yourself.
- Check any performance tool your organisation uses against the four guardrails. For consequential vendor tools, ask for the bias-audit results.
- Confirm where performance data goes when you use an AI tool, and whether that tool is approved for personal data at your organisation.
- Run a small language audit by group across your last cycle: evidence-backed statements versus adjectives, per person. Note the pattern before b9.
Sources covered (cited)
Taught from the AI-governance sources for worker measurement. This session covers:
Three questions before you go 🎯 ◐ 90 seconds
1 · Which task belongs in the assist lane rather than the decide lane?
AI works on the evidence; the human makes the judgment. Evidence assembly directly attacks recency bias. Ratings, ranking, pay, and exit flags all change what happens to a person, so a human must decide and sign them.
2 · A manager approves twelve AI-drafted ratings in four minutes. What is wrong?
Human in the loop means a real decision. The reviewer must see the inputs, have time to check them, be able to override, and be accountable for the result. If nothing is ever overridden, the oversight is decorative - and GDPR Article 22 is aimed at exactly this shape.
3 · Under the EU AI Act, AI used to manage and evaluate workers is classified as...
Annex III places employment and worker management in the high-risk category. High-risk means obligations attached, not prohibition - and as of August 2026 those obligations apply from 2 December 2027 for standalone Annex III systems, after the AI Omnibus deferral. Together with GDPR Article 22 and, in New York City, Local Law 144's already-live bias-audit and notice duties, that is the floor - meeting it makes a system legal, not yet fair.