This AI performance review survey template systematically measures how employees and managers experience AI in performance reviews — from trust and fairness to data privacy and governance. It covers 84 Likert-scale statements, 6 overall ratings (0–10), and 12 open questions. Use it to spot transparency, quality, or compliance issues early, and to set clear thresholds for when to pause, retrain, or tighten guardrails.

Why this survey matters in 2026

AI-assisted performance reviews are spreading fast — with significant legal and cultural consequences. A 2025 Resume Builder survey of more than 1,300 managers found that 91% use AI to assess performance, but only 71% felt confident in its fairness for individual decisions. Without structured measurement, that confidence gap turns into a trust and legal liability problem.

Three regulatory layers create specific obligations for employers using AI in reviews:

This survey gives you a measurable, documentable baseline across all these dimensions — and concrete inputs for your works council and data protection officer.

How to use this template

Use a 5-point Likert scale for statements (1 = Strongly disagree, 5 = Strongly agree). Numbering is for analysis and follow-up: E = Employees, M = Managers, S = Shared. Select one of the four blueprints below and load only the relevant items. For a first run, 18–22 questions is plenty; keep the total under 8 minutes.

Survey questions: Employees

Employees (E1–E6) — Awareness & transparency

Employees (E7–E12) — Quality & usefulness

Employees (E13–E18) — Fairness & bias perceptions

Employees (E19–E24) — Psychological safety & trust

Employees (E25–E30) — Data protection & consent

Employees (E31–E36) — Overall impact & preference

Survey questions: Managers

Managers (M1–M6) — Onboarding & training

Managers (M7–M12) — Workflow & time impact

Managers (M13–M18) — Quality of drafts & summaries

Managers (M19–M24) — Judgment, oversight & accountability

Managers (M25–M30) — Governance & guardrails

Managers (M31–M36) — Fairness, consistency & calibration support

Managers (M37–M42) — Overall confidence & willingness to continue

Survey questions: Shared (Employees + Managers)

Overall ratings (0–10) and open questions

Overall ratings (0–10)

Open questions (12 total)

Survey blueprints: which questions to use when

Pick one blueprint per cycle. Simple flow: select blueprint → load items → agree anonymity threshold → send 3–10 days after reviews close → publish results and actions within 21 days. If you already run a classic post-cycle survey, align wording to keep trends comparable even as AI changes the workflow. For tool selection guidance, see our enterprise performance management software guide.

Scoring & thresholds

Use the 5-point scale for statements and 0–10 for overall ratings. Tie decisions to thresholds — not individual anecdotes. Three bands: Low (Avg <3.0) — fix before scaling. Medium (3.0–3.9) — improve next cycle. High (≥4.0) — standardise and share examples. For 0–10 ratings, treat <6/10 as a stop-and-fix signal in pilots.

Four-step analysis routine: (1) compute averages per dimension, (2) check dispersion and "disagree" rates, (3) compare groups, (4) map to actions in the decision table.

Action plan: what to do when scores are low

AI in reviews fails most often not because of the tool, but because nobody owns the messy parts: corrections, escalations, and governance. Set routing rules upfront. Default response times for pilots: ≤24 h for severe privacy concerns, ≤7 days to acknowledge low trust scores, ≤21 days to publish actions. Align calibration follow-ups with your internal calibration process to keep AI from becoming a backdoor to inconsistent standards.

Fairness & bias checks

Fairness is not a single average score. Break results down by relevant groups and compare both perception and process signals. Research on AI-based performance prediction shows employees judge AI features as fair when they closely align with actual job performance — generic or opaque systems undermine that perception even when output quality is adequate. Practical red flags: group gap ≥0.4, or ≥15 percentage points difference in favorable rates, or repeated open-text mentions of "generic," "copy-paste," "unfair," or "privacy." Align language checks with your internal bias review process for consistent standards.

Practical scenarios

Scenario 1: Employees distrust AI because disclosure is unclear

You see E1–E6 Avg 2.8 and comments like "I only noticed AI wording after the meeting." HR adds disclosure at the point of use: the review form shows "AI-assisted draft: yes/no" plus a one-paragraph explanation. Managers get a 60-second script for the review conversation. You re-run the combined pulse (Blueprint C) in 60 days to confirm trust moved upward.

Scenario 2: Managers save time, but employees find feedback generic

M7 and M8 score high (≥4.0), but E7–E12 score low (Avg <3.0) and E33 rises. Decision: keep AI for structuring, but require evidence. Each review section must include 2 proof points — project, metric, observable behaviour. HR shares "good vs bad" examples, updates prompts, and spot-checks 10% of reviews for specificity before they reach employees in the next cycle.

Scenario 3: Fairness concerns cluster in one group

E16 is stable overall, but one location shows a gap of 0.5 points and higher E14. HR and the local works council check whether different prompts, rubrics, or data sources were used there. Next steps: calibration refresher, sharpen rubric anchors, re-brief managers on bias patterns. You compare the same group slice in the follow-up trend survey (Blueprint D) to confirm the gap narrowed.

Implementation guide

Start small, then scale. Involve your works council early whenever AI touches performance processes, monitoring questions, or decision support — the EU AI Act and national labor law both require prior consultation. Keep the survey clearly separate from individual outcomes: responses must not influence ratings, pay, or performance labels — or you will kill trust and participation rates.

Recommended rollout rhythm: pilot (6–10 weeks) → first review cycle → survey + fixes (≤30 days) → scale to next area → trend check at 6–12 months. For broader talent and skills planning alongside your AI rollout, our talent management software guide for DACH covers the GDPR and works council checklist in detail.

For broader skills and competency tracking alongside your AI governance work, a structured skill management software comparison helps you separate "tool problems" from "skill gaps" — and plan targeted fixes faster.

FAQ

How often should we run these AI performance review survey questions?

During a pilot, run it after every AI-assisted review cycle for at least the first 2 cycles — while memories are fresh. After that, a deep-dive once a year plus a short pulse after major AI feature changes works well. Keep 6–8 core items stable across runs (trust, fairness, psychological safety, privacy) so you can track trends reliably over time.

What should we do if scores are very low (Avg <3.0) or comments are harsh?

Start with containment and clarity — not defensiveness. Acknowledge results within ≤7 days and say what will happen next. Use focus groups to identify root causes: disclosure gaps, generic wording, privacy fear, or manager behaviour? Then commit to 3 fixes maximum, with owners and deadlines. Close the loop publicly within ≤21 days. For severe privacy concerns, route to DPO/IT with a response time of ≤24 hours.

How do we stop the survey from feeling like monitoring or performance control?

Be explicit about purpose and separation: survey responses must not influence ratings, pay, or individual outcomes. Report aggregated results only, and apply anonymity thresholds (no reporting for groups smaller than 5–7). Involve your works council or employee representatives early and document guardrails clearly — data sources, retention periods, access rights. The EDPB guidelines on automated decision-making are a useful reference point for framing this internally.

What do GDPR Art. 22 and the EU AI Act require in practice?

GDPR Art. 22 prohibits decisions with significant effects that are based solely on automated processing. AI-generated rating suggestions or calibration inputs that managers accept without substantive review potentially trigger this protection. Employees have the right to request human intervention, express their view, and contest the outcome — and genuine review requires actual decision-making authority, not rubber-stamping. The EU AI Act adds a second layer: performance evaluation systems are high-risk AI (Annex III), requiring risk management documentation, logging, human oversight, and a fundamental rights impact assessment. From August 2026 (or later if the Digital Omnibus proposal passes), employees may also request a meaningful explanation of any high-risk AI output under Art. 86 of the AI Act.

Should we tell employees when AI was used in their review?

Yes — if you want trust. Lack of transparency tends to inflate fairness concerns even when output quality is fine. Keep it simple: where was AI used (drafting, summarising, calibration support), what data did it use and not use, and who owns the final content. Add a correction right: employees can challenge wording and request edits without needing to "prove" the AI was wrong.

How do we keep the question bank current as tools and policies change?

Run an annual review with HR, a few managers, People Analytics, and your works council where applicable. Keep your core trend items stable (trust, fairness, psychological safety, privacy) and rotate a small set of feature-specific questions based on what changed — new summarisation features, new rating suggestions, new data sources. Pilot new items with one team first, then scale once wording is unambiguous and action-ready.