AI for performance reviews: how to use it well, and what California SB 947 changes
Use AI for performance reviews to draft, structure and summarize, never to decide. From July 1, 2027, California's SB 947 bars relying solely on an automated decision system to discipline or fire, and requires a human to corroborate decisions that rely primarily on one. The safe pattern: AI drafts from work evidence, the manager judges, and the employee sees the data.
This guide is for managers and HR teams. It covers what AI does well and badly in reviews, what SB 947 requires, a 5-step workflow, 10 prompts with examples of the output to expect, the mistakes to avoid, and how Intelogos supports evidence-based reviews. If you're an employee using AI on your own self-review, the FAQ covers that too.
Table of Contents
- Should you use AI for performance reviews?
- What does California SB 947 require for performance reviews?
- What is a safe workflow for an AI performance review?
- Which AI performance review prompts work? 10 examples
- What mistakes should you avoid?
- How does Intelogos support evidence-based reviews?
- Frequently Asked Questions
Should you use AI for performance reviews?
Yes for the writing and organizing, which is where reviews eat a manager's time; no for the judgment. AI is good at turning notes and evidence into a clear draft and at checking language, but it can't see work that isn't in front of it, weigh context or take responsibility for a rating.
| What AI does well | What AI can't do |
|---|---|
| Turn a period's notes, goals and results into a structured first draft | Know about work that isn't in your notes or data, such as mentoring, calls or thinking time |
| Group evidence by goal or competency | Weigh context, such as a reorganization, an outage or a leave of absence |
| Rewrite vague or personal comments as specific, behavior-based feedback | Decide a rating fairly on its own, or be accountable for it |
| Check a draft for unsupported claims and inconsistent language | Guarantee accuracy: it can state things your evidence doesn't support |
| Prepare questions and an agenda for the review conversation | Replace the conversation |
The biggest gain comes from changing what goes in, not just how fast the text comes out. In a May 2026 Harvard Business Review article, Boston University professor Chrysanthos Dellarocas argues that most companies use generative AI to write the same narrative reviews faster and more smoothly, when the bigger opportunity is to pull evidence of how people actually performed out of everyday work, such as their decisions, mentoring and problem solving (HBR, May 15, 2026). That is the approach in this guide: evidence first, AI second, judgment last.
What does California SB 947 require for performance reviews?
From July 1, 2027, SB 947 bars California employers from relying solely on an automated decision system (ADS) for a disciplinary or termination decision, requires a human to corroborate any such decision that relies primarily on ADS output, gives employees a right to a description of their data, and requires a written notice after use. It also bans using an ADS to infer an employee's protected status.
SB 947 (Chapter 859, Statutes of 2026) was approved on September 30, 2026 and adds sections 1520 to 1526.7 to the California Labor Code (bill text):
| Rule | What the law says | Section |
|---|---|---|
| What counts as an ADS | A computational process derived from machine learning, statistical modeling, data analytics or artificial intelligence that issues simplified output, such as a score, classification or recommendation, used to assist or replace human discretionary decision-making and that materially impacts people. Exclusions include spam filters, firewalls, antivirus software, calculators and databases. | 1520(c) |
| Banned uses | An employer may not use an ADS to violate, or prevent compliance with, labor, health and safety, employment or civil rights laws, to infer an employee's protected status under Government Code section 12940, or to predict and take adverse action against a worker for exercising legal rights. | 1522(a) |
| No sole reliance | An employer may not rely solely on an ADS when making a disciplinary or termination decision. | 1522(b)(1) |
| Human corroboration | If an employer relies primarily on ADS output for such a decision, a human must corroborate it using the data behind the output or other information, which may include supervisory evaluations, personnel files, work product, peer reviews and witness interviews. Output that can't be corroborated, or that the reviewer finds inaccurate, incomplete or misleading, can't be used. | 1522(b)(2), 1522(c) |
| Data access | When an employer primarily used an ADS for such a decision, the employee can request a meaningful, objective description of their own data the ADS used. | 1522(d) |
| Post-use notice | An employer that primarily relied on an ADS must give a written, plain-language, stand-alone notice when it tells the employee the decision: that it relied primarily on an ADS, that a human reviewed the decision and corroborated the output, who to contact, the right to the data description, and that retaliation is prohibited. | 1524 |
| Enforcement | The Labor Commissioner or a public prosecutor enforces it, with a civil penalty of $500 per violation. In a claim under the reliance and corroboration rules, once ADS use is shown, the employer must show it didn't primarily rely on it, or that it complied. | 1526.1 |
The text doesn't mention performance reviews. Its corroboration, data access and notice duties attach to disciplinary and termination decisions, so they matter for a review when an AI-generated rating, score or recommendation feeds a decision to discipline or fire someone. The ban on using an ADS to infer protected status applies whatever the decision. Businesses covered by the California Consumer Privacy Act are also subject to the California Privacy Protection Agency's automated decision-making rules (section 1526.4). Whether a particular AI tool or process counts as an ADS is a question for counsel.
A second California law matters for reviews too. From January 1, 2027, AB 1883 (Chapter 853, Statutes of 2026) bars employers from using a workplace surveillance tool that uses artificial intelligence to recognize, or make inferences or predictions about, an employee's emotional state, or to collect neural data (Labor Code section 1581). Our California AB 1883, AB 1331 and SB 947 guide covers all three laws, how they apply to monitoring software and a 30-day plan.
This is general information, not legal advice. Laws vary by location and change often; consult your own counsel before acting on it.
What is a safe workflow for an AI performance review?
Gather evidence first, let AI draft from that evidence only, check the draft for bias, add your own judgment, then review it with the employee. Each step keeps a person responsible for the outcome.
- Gather evidence for the whole period. Goals and results, examples of work product, feedback from peers and customers, your one-to-one notes, and work-time data the employee already knows about, such as hours worked against expected hours and time in core tools. Cover every month of the period, not just the last few weeks.
- Draft with AI, from the evidence only. Use an AI tool your company has approved, leave out details the tool doesn't need, and replace names with initials if your policy asks you to. Tell the tool to use only the evidence you provide and to flag anything it can't support. An AI acceptable use policy sets these rules for everyone.
- Check for bias. Ask the AI, then yourself: does every claim trace to evidence? Does the language describe behavior or personality? Are similar results described the same way across the team? Is the last month over-weighted? Never ask the AI to guess anyone's emotions, mood or personal characteristics.
- Add your judgment. Decide the rating yourself. Add the context the evidence can't show, such as a reorganization, an outage, mentoring or a leave of absence, rewrite the draft in your own words and delete anything you couldn't explain in the conversation. If the review could lead to discipline or termination in California, record who reviewed the evidence and what corroborated it: once ADS use is shown, SB 947 puts it on the employer to show that it complied.
- Review it with the employee. Share the evidence behind each point, ask for their view and correct anything that's wrong. Say how AI helped with the draft, in line with your policy, and agree goals for the next period.
The free employee monitoring policy kit helps with the groundwork: telling employees what work data is collected and why before it is used in any review, with a state notice checklist updated for California's 2026 laws.
Which AI performance review prompts work? 10 examples
These 10 prompts are written for managers drafting reviews in a general AI assistant your company has approved, such as ChatGPT, Claude, Gemini or Copilot. Each asks the AI to work only from evidence you provide. The "What to expect" lines show the kind of output each prompt returns; the names, dates and figures in them are made up.
1. Organize the evidence
Here are my notes, the goals we agreed for [period] and the results. Group the evidence by goal. For each goal, list what was delivered, what slipped and the evidence for it. Use only what I've given you, and list anything that's missing.
What to expect: "Goal 1, cut first-response time: delivered. Evidence: median first response fell from 6 hours to 3 (Q3 queue report). Goal 2, launch the help center: delivered two weeks late (launch notes, August 14). Missing: no evidence yet for goal 3, mentoring the new hire."
2. Turn work data into questions
Here is a summary of J.S.'s work-time data for [period]: hours worked against expected hours, time in core tools and after-hours work, by month. Describe the patterns you see, and phrase each one as a question I could ask in the review, not as a conclusion.
What to expect: "Hours worked were above the expected 40 a week in 7 of 12 weeks, mostly in September. Question: 'September looked heavy. What drove the extra hours, and is it still the case?' Time in core tools fell in July while meeting time rose. Question: 'Was July mostly the vendor migration?'"
3. Draft from the evidence only
Using only the evidence above, draft a 200-word summary of this person's performance for the period. After each sentence, name the piece of evidence it comes from. Mark any sentence you can't support with [needs evidence].
What to expect: "This period, J.S. delivered three of four agreed goals (goal tracker) and halved the team's first-response time (Q3 queue report)... J.S. is a natural leader [needs evidence]." A marked sentence is your cue to find an example or delete it.
4. Describe behavior, not personality
Rewrite each point in this draft so it states what happened, the impact, and a suggestion for next time. Replace adjectives about personality with specific behavior.
What to expect: "Instead of 'disorganized': 'Two of five handoffs to QA in August were missing the release checklist, which delayed testing by a day each time. Next quarter, use the handoff template for every release.'"
5. Check the language for bias
Review this draft for words that describe personality rather than work, for feedback that is vague or not tied to evidence, and for any reference to age, gender, health, family or other personal characteristics. List each issue with a suggested rewrite.
What to expect: "'Abrasive in meetings' describes personality. Try: 'In two planning meetings, cut colleagues off before they finished their updates.' 'Surprisingly quick with new tools for someone her age' refers to age: remove it and name the result instead, such as 'set up the new reporting tool for the team in a week'."
6. Check for recency bias
Here are my notes for [period], by month. Does my draft give more weight to the last six weeks than to the rest of the period? Show which months are under-represented and what evidence from them is missing from the draft.
What to expect: "January to April hold 9 of your 20 notes but appear in one sentence of the draft. The Q1 data migration, which your notes call the team's biggest delivery, isn't mentioned."
7. Compare consistency across your team
Here are draft reviews for my four direct reports, with names replaced by initials. Compare how I describe similar results. Point out where I've used stronger or weaker language, or different ratings, for comparable outcomes.
What to expect: "A.B. and C.D. both met all four goals. A.B.'s review calls the results 'exceptional'; C.D.'s says 'met expectations'. Either state the reason the ratings differ, or make the language match."
8. Suggest development goals
Based on the gaps in this review, suggest three development goals for the next six months. Make each one specific and measurable, and add one way I can support each.
What to expect: "1. Run the weekly release review on your own by March. Support: co-run the first two with you. 2. Write the handoff checklist for the new QA process by February. Support: a review of the first draft. 3. Present one customer insight at each monthly all-hands. Support: time in the agenda."
9. Prepare the conversation
Turn this review into an agenda for a 30-minute conversation: an opening question, two strengths with examples, one area to improve with an example, questions to ask the employee, and next steps to agree.
What to expect: "Opening: 'What are you proudest of from this half?' Strengths: the help center launch (August) and first-response time (Q3 report). Area to improve: QA handoffs (two incomplete in August). Question: 'What got in the way of the checklist?' Next steps: agree the handoff template and review it together in six weeks."
10. Compare with the self-review
Here is the employee's self-review and my draft. List where we agree, where we differ and what evidence each of us cites. Suggest questions to resolve the differences in the conversation.
What to expect: "Agree: the launch and the response-time results. Differ: the self-review rates stakeholder management as a strength; your draft lists two escalations from the finance team in June. Question: 'Walk me through the two escalations in June. What would you do differently?'"
What mistakes should you avoid?
The three that do the most damage are letting the AI decide, asking it to read emotions or sentiment, and drafting without evidence. Each makes a review less fair, and in California the first two can also create legal risk.
- Using AI as the judge. A rating, score or recommendation from an AI tool is something to check, not a decision. From July 1, 2027, SB 947 bars relying solely on an automated decision system for discipline or termination, and output that can't be corroborated can't be used.
- Inferring emotions or sentiment. Don't use tools that claim to score mood, stress or sentiment from faces, voices or messages, and don't ask a chatbot how someone feels. Feelings aren't evidence of performance, and from January 1, 2027, AB 1883 bars California employers from using workplace surveillance tools that use artificial intelligence to recognize or infer an employee's emotional state.
- Reviews without data. A fluent draft from a two-line prompt is mostly the AI's invention. Every claim should trace to evidence you can show the employee.
- Inferring protected characteristics. SB 947 bars using an ADS to infer an employee's protected status. Keep age, health, family, ethnicity and similar details out of prompts and drafts.
- Pasting confidential data into unapproved tools. Use the tools your company has approved, with the data its policy allows.
- Copy-paste sameness. Identical phrases across a team's reviews tell people nobody wrote them. Edit each draft in your own words.
- Hiding AI use. Say how AI is used in reviews, in your policy and in the conversation, so employees don't find out by surprise.
How does Intelogos support evidence-based reviews?
Intelogos gives managers work-time evidence for the whole review period: KPIs with published definitions, the activity record behind them, and AI reports and answers that summarize it. Reviews then start from data the employee can also see, not from memory.
Intelogos is a workforce analytics platform that shows how work time is spent across apps and websites, without screenshots or keystroke logging, and lets managers ask questions about it in plain language. Intelogos doesn't take screenshots. Intelogos doesn't record what people type; it records only whether keyboard or mouse input happened. Its agents record the applications and websites used, window titles, active and idle time and working hours, and don't capture webcam images or audio recordings (privacy policy).
What managers can bring to a review:
- KPIs for the period. Workday, Availability, Activity, Engagement, Productivity, Contribution and an Overall Rating, each with a definition in the help center. Productivity, for example, is "the portion of active work minutes spent in primary and secondary tools, out of the total possible time based on how much the user is expected to work" (Productivity KPI). They show how time was spent, not the quality of the work, so read them as trends next to outcomes.
- Chronicle for specific questions. A chronological record of the tools and windows a person used, available to Owners and to Admins and Managers with Chronicle permission (Chronicle). Use it to answer a specific question that came up, not to build a case.
- Generate Insights. From a person's Profile or Employee page, an AI-written Individual Performance Insights report for the date range you choose, covering KPIs and trends, tool usage, patterns such as after-hours work, and recommendations. The help center lists one-to-ones and performance reviews among its uses (Generate Insights). The Analytics plan includes 5 a month; AI Intelligence and Enterprise include an extended allowance (pricing).
- AI Performance Summaries and AI Coaching Recommendations, included in the AI Intelligence and Enterprise plans (pricing).
- Ask AI. On the AI Intelligence and Enterprise plans, owners, admins and managers can ask about their team's performance in plain language, for example "How has productivity changed over the last 30 days?", and connect tools such as GitHub, Jira, Slack and Google Calendar so answers can draw on that data too (Ask AI). The help center's own advice: use AI insights as one input to your decisions, not the sole basis.
- The employee's view. Employees can see their own time data, or their full activity and KPIs, depending on an account-wide setting (roles and access), so the evidence in a review is evidence they can check.
Intelogos KPIs and AI reports are inputs to a manager's judgment. If you use any of them in a disciplinary or termination decision in California, SB 947's corroboration, data access and notice duties apply to your process.
Frequently Asked Questions
How do you use AI for a review?
Gather the evidence for the whole period first: goals and results, work product, feedback and work-time data the employee knows about. Then ask an AI tool your company has approved to organize that evidence and draft from it only, check the draft for unsupported claims and biased language, decide the rating yourself and discuss it with the employee. Use AI for the writing, not the judgment.
Is it legal to use AI for performance reviews?
SB 947 doesn't ban AI-assisted performance reviews. It governs how California employers use automated decision systems in disciplinary and termination decisions: from July 1, 2027, no sole reliance, human corroboration when such a system is the primary basis, a description of the employee's data on request and a written post-use notice. It also bans using one to infer protected status. AB 1883 separately bars AI workplace surveillance tools that infer emotional state, from January 1, 2027. Other states and cities have their own rules, so check with counsel before relying on AI output in any employment decision.
This is general information, not legal advice. Laws vary by location and change often; consult your own counsel before acting on it.
Can employees see AI-written reviews?
Often, yes. In California, current and former employees have the right to inspect and receive a copy of the personnel records their employer keeps relating to their performance (Labor Code section 1198.5), and that doesn't depend on who or what drafted them. From July 1, 2027, an employer that primarily relied on an automated decision system for a disciplinary or termination decision must also give a post-use notice and, on request, a description of the employee's data it used. Rules differ by state, and sharing the evidence behind a review is good practice everywhere.
This is general information, not legal advice. Laws vary by location and change often; consult your own counsel before acting on it.
Can employees use AI to write their self-review?
Check your employer's AI policy first, since it may limit which tools you can use and what information can go into them. Use AI to organize your own evidence, such as results, examples and feedback you received, and to tighten the wording, then edit the draft so every claim is accurate and it sounds like you. Don't paste confidential company or customer information into a tool your employer hasn't approved.
Sources
Every legal statement in this guide was checked on the official source below on October 23, 2026.
- California SB 947 bill text: Chapter 859, Statutes of 2026, approved by the Governor and filed with the Secretary of State on September 30, 2026; Labor Code sections 1520 (definitions), 1522 (banned uses, no sole reliance, human corroboration, data access), 1524 (post-use notice), 1526.1 (enforcement and the $500 civil penalty), 1526.4 (CCPA-covered businesses) and 1526.7 (operative July 1, 2027).
- California AB 1883 bill text: Chapter 853, Statutes of 2026; Labor Code section 1581, AI workplace surveillance tools that infer emotional state or collect neural data. Its January 1, 2027 effective date is explained in our California employer guide.
- California Labor Code section 1198.5: the right to inspect and copy personnel records relating to performance.
Background, checked October 23, 2026:
- Harvard Business Review, "Gen AI Could Fix Performance Reviews, or Make Them Even Worse" (Chrysanthos Dellarocas, May 15, 2026).
Intelogos facts come from our help center (Ask AI, Generate Insights, Chronicle, Productivity KPI, roles and access), our pricing page and our privacy policy.