Where we are
The hiring wave is half done and the people you hired in b4 to b6 are now inside the performance cycle with everyone else. You already know the discipline this session needs, because you built it for interviews in b5: a defined scale, evidence against every rating, and a named person who decides. What changes here is the stakes on the other side - a review follows someone through promotion, pay and, eventually, whether they stay. It is also, like hiring, explicitly inside the regulated category.
Three jobs inside a review, and the line between them 6 min live
A performance review looks like one task and is really three: gathering the evidence, writing the words, and making the judgment. AI is genuinely useful at the first, useful with supervision at the second, and must not touch the third. Drawing that line is not caution for its own sake - performance evaluation, promotion and termination sit inside the same high-risk employment category as hiring under the EU AI Act, with the same expectation of real human oversight.
LiveWhy the third job is not negotiable3 min▶
The AI Act's employment high-risk scope is wider than most HR teams assume. It covers recruitment and selection, and it also covers performance evaluation, task allocation, monitoring of workers, promotion and termination. A tool that rates or ranks your employees is in the same regulated category as a screening tool - with obligations around record-keeping, transparency and human oversight, and with oversight meaning someone who could actually have decided differently. As of mid-2026 those obligations apply from 2 December 2027 after the Digital Omnibus postponement, and that date has already moved once, so check the current position rather than trusting this page.
- A rating is a consequence, not a summary. It moves pay, promotion and, in a downturn, who is on a list. That is why it needs an accountable human name on it.
- "The system suggested it" is the worst possible answer in an appeal, a grievance, or a tribunal - and it is exactly what a manager will say if you let a draft arrive with a number already in it.
- Rail 1 does the work here. A person decides, the record shows who, and the reason is written the same day - not reconstructed after somebody objects.
The number that arrived first. One team let a tool pre-populate a suggested rating at the top of each review form, on the reasonable theory that managers could change it. Almost nobody did. The suggestion became the anchor, the written evidence got reverse-engineered to fit it, and the cycle produced a distribution nobody could explain when asked. The fix was one line of process: the evidence gets written before any number exists, and the number is typed by the manager.
Self-studyWhat the data rail means in a performance cycle2 min read▶
Rail 3 does not soften because these are your own employees. Performance material, grievance and disciplinary records, health and accommodation information, and named pay data stay out of a general AI workspace. Under GDPR and PDPA you need a lawful basis, the minimum necessary data, a retention limit, and meaningful human involvement in anything that decides something about a person.
The practical version for this fortnight: managers work from their own notes about their own reports, de-identified where a tool is involved, and nobody pastes a review of one person into a workspace that another manager can read. If your organisation has an approved HR-specific tool with the right terms, that is where this belongs. If nobody can tell you what is approved, that is a finding worth escalating - leader session a5 is the conversation.
Evidence over vibes 5 min live
A defensible review contains behaviours, artifacts, dates and impact. An indefensible one contains adjectives doing the work that examples should be doing. The test is simple and unforgiving: for every claim, can you point at the thing? If not, it is an impression, and impressions are where everything unfair in a performance cycle hides.
LiveThe rewrite table - three vague lines, fixed3 min▶
| Vague, and unusable | Evidenced, and actionable |
|---|---|
| "Strong team player, well liked." | "Ran the March incident bridge for four hours, kept two customer teams informed on a 30-minute cadence, and wrote the postmortem that support now uses as a template." |
| "Needs to be more strategic." | "Escalated three recurring ticket patterns as individual issues rather than proposing the underlying fix. Next quarter: bring one root-cause proposal per month to the ops review." |
| "Communication could improve." | "Two customer updates in Q2 went out without a next step or an owner, and both generated a follow-up call. Use the update template from the support playbook." |
- Behaviour, not character. "Is disorganised" describes a person. "Missed three of eight agreed follow-up dates" describes something that can change.
- No unnamed comparisons. "Not as strong as others on the team" is unappealable and tells the person nothing.
- Developmental advice as concrete as the praise. This is where reviews quietly become unequal - specific praise and vague growth advice is a pattern, not an accident.
The review that could not survive being read aloud. A manager wrote that an engineer "sometimes lacks executive presence". Asked in a calibration session what that meant, he could not produce a single incident, and eventually said she was quiet in large meetings. The rewrite - "in three architecture reviews she raised her objection afterwards in writing rather than in the room; agree in advance which meetings she will speak in" - was something she could act on, and something he could defend. Same observation, one of them fair.
LiveThe bias check, and the smoothing trap2 min▶
Written feedback shows systematic differences by group: more personality language and less concrete achievement language for some people, vaguer developmental advice for others. The check is a re-read against three questions, not a feeling.
- Does every claim cite an artifact, a date or an incident?
- Is the developmental advice as specific as the praise?
- Would this sentence survive being read aloud to the person, in their words, in a room?
Then the trap. Ask a model to "improve" or "soften" a review and it will regress toward corporate mush - the specific incident becomes "some challenges around delivery", the actionable advice becomes "continue to develop in this area". That is the opposite of what a review needs. Use it to organise and tighten; keep the manager's specifics verbatim, and fix wording only.
The same feedback in three states 4 min live
Vague, smoothed, evidenced. The middle one is the dangerous state, because it reads better than the first and is just as useless - it has been polished into unfalsifiability, which is the same failure the status report and the screening slate both had earlier in this course.
Self-studyCalibration, from b5, applied to reviews2 min read▶
The interview calibration drill from b5 works unchanged here: several managers rate the same written sample independently, reveal, and discuss the widest gap. Rating drift across managers is the performance cycle's version of rater drift, and it produces exactly the same result - two people doing the same work getting different numbers because their managers use the scale differently.
Run it before the cycle opens, not after the ratings are in. Once numbers exist, calibration turns into negotiation, and the person with the most confident manager wins.
The review clinic ★ 14 min · everyone builds
Take a manager's rough notes - the real kind, half sentences and a couple of dates - and produce an evidenced draft with the gaps flagged rather than filled. This is the artifact you hand to Kestrel's managers this fortnight.
Collect the manager's own material: their notes, the artifacts they can point at, and the role's capabilities from the scorecard (b2). Nothing about anyone else's report.
Run the prompt below. Read the flagged list first - the claims with no artifact behind them are the real output.
Send the manager back to find the evidence, or to drop the claim. Both are acceptable; inventing an example is not.
The manager writes the rating themselves, after the evidence is settled, and records the reason the same day.
Run the three-question bias check on the finished draft before it goes to the employee.
The read-aloud test ★ 8 min · build your own
Eight minutes on one real paragraph of your own. This is the drill that changes how people write for the rest of their careers, and it needs no tooling at all.
Take one paragraph of feedback you have written about a real person. Strip the name.
Underline every claim. Next to each, write the artifact, date or incident that supports it. Any claim with a blank next to it is not feedback yet.
Find the vaguest developmental sentence. Rewrite it as something the person could do this month, with a way to tell whether it happened.
Read the paragraph aloud as if the person were in the room. The sentences you would soften while speaking are the ones to fix on the page - not by softening them, but by adding the specific that makes them fair.
The blank column. A people team ran this drill across one department before the cycle opened. Managers underlined their claims and filled in the evidence column, and roughly a third of the cells came back empty - not because the managers were careless, but because they had never been asked. The next cycle's reviews were shorter, more specific, and produced half as many appeals. Nothing about the tooling changed; one column did.
Try it yourself - this week ◐ 30-45 min total
- Run the review clinic on one real set of manager notes and keep the flagged list. Count how many claims had nothing behind them.
- Add the keep-every-specific instruction to whatever drafting prompt your managers are already using. It takes one line and it is the highest-value edit in this session.
- Run the three-question bias check across five finished reviews from different managers, and compare how concrete the developmental advice is between them.
- Check whether any tool in your cycle pre-populates or suggests a rating before the evidence is written. If one does, that is the process fix to make before the next cycle opens.
- Book the calibration session before ratings exist, not after. Bring the widest-spread sample to it.
Official sources covered
Taught from public regulation, employee-data rules, and feedback research. This is education for HR professionals, not legal advice, and obligations vary by jurisdiction. This session covers:
Three questions before you go 🎯 ◐ 90 seconds
1 · Which of the three jobs inside a review must never move to a model?
Gathering and drafting are where AI earns its place. The rating moves pay, promotion and eventually who is on a list, which is why it needs a named human who could have decided otherwise - and why performance evaluation sits in the same high-risk category as hiring.
2 · A manager asks the model to "improve" a review and gets back smoother prose. What most likely happened?
The smoothing trap. "Two Q2 updates went out with no next step" becomes "continue developing in stakeholder communication" - better prose, no information. Instruct it to keep every specific verbatim and fix wording only.
3 · You re-read a review and every positive point cites an artifact, while the growth advice is one vague sentence. What is that?
Specific praise plus vague growth advice is one of the documented patterns in how feedback differs between groups, and it leaves the person with nothing to act on. Make the developmental point as concrete as the strongest positive one.