AI Assisted Performance Review

AI-Assisted Performance Review Skill

The uncomfortable review question of the decade: when a report ships twice the output with AI, what did they do? Volume stopped measuring effort; polish stopped measuring skill. Punishing AI use is as wrong as crediting the model's work to the human. This skill separates the signals — and gives managers the conversation, not just the theory.

What This Skill Produces

• A what-measures-whom analysis of the role's current evaluation criteria
• Rewritten criteria that measure the human: judgment, verification, outcomes, leverage
• Calibration rules for teams with uneven AI adoption
• Conversation scripts for the three hard cases

Required Inputs

Ask for (if not already provided):

• The role and current review criteria (the rubric, or how it really works)
• How AI shows up in the work — which tasks, how much of the output it drafts, what the tooling reality is
• The specific situation, if any: one person's review? team calibration? criteria rewrite?
• The org's AI stance — encouraged? tolerated? policy exists? (Reviews must not punish sanctioned behaviour)

Method

1. Sort every criterion: human, tool, or hybrid. Walk the current rubric. Volume of drafts, formatting quality, speed to first version → now mostly tool signals (evaluating them evaluates prompt luck and subscription tier). Decision quality, stakeholder trust, error catch rate, what they chose to build → still human. Output quality overall → hybrid: credit belongs to the pair, and the review's job is to see the human's contribution inside it.
2. Rewrite around the four durable human signals:
   • Judgment — what they decided to do, what they declined, how they scoped; the quality of taste applied to AI output (what they kept, cut, and corrected)
   • Verification — do errors get caught before shipping? A person whose AI-assisted work is reliably right is demonstrating skill; one who forwards unverified fluency is a risk wearing productivity's clothes
   • Outcomes — did the work move what it was for (the metric, the decision, the customer), independent of how it was produced
   • Leverage — do they make AI multiply the team (shared prompts, workflows, teaching) or only their own count
3. Set the calibration rules for mixed adoption. In one team you'll have a 2×-output adopter and a careful non-adopter. Rules that keep it fair: evaluate against the role's outcomes, not each other's volume · where AI use is sanctioned, not adopting is a development conversation (not a values one) · where someone's edge is invisible verification labour, surface it explicitly before comparing. Never let the review become a proxy war about the tools.
4. Demand evidence that sees the human. Volume anecdotes are out. In: a sample of shipped work walked backwards (what did the AI draft, what did you change, why) · error/rework history · decisions log · peer signals about trust and leverage. The walk-backwards exercise is the single highest-signal artifact — put it in the review prep.
5. Script the three hard cases:
   • The volume star with thin judgment — "Your output doubled; let's walk three pieces backwards" (the conversation is about the delta between draft and shipped)
   • The careful sceptic being out-shipped — outcomes-first framing; adoption raised as growth, not deficiency; their verification strength named as a strength
   • The launderer — unverified AI work shipped as their own, errors reaching others: this is a reliability conversation with the accountability rule from the org's AI policy, not an AI conversation