AI for Performance Reviews: Guide for Small Business
AI for performance reviews: 8 use cases (what works, what doesn't), 7-step responsible process, prompt examples, risks, and common mistakes managers make.
AI for Performance Reviews
An honest guide for small business managers
The first time I used AI to help write a performance review, I made the mistake almost every manager makes the first time. I gave it three vague observations, asked for a complete review, and got back four paragraphs that sounded plausible but contained two completely invented metrics, three generic phrases that could have applied to anyone, and one line of subtly gendered language I almost missed.
The output looked finished. It was useless. The version I eventually sent took the same 90 minutes to produce as if I had not used AI at all, because I had to fact-check, rewrite, and rebuild from scratch. Here is how to get the time savings without that failure.
Most articles on AI for performance reviews fall into one of two camps. The vendor-written ones tell you AI is going to revolutionize performance management and you should buy their AI feature. The skeptical ones tell you AI is dangerous and you should avoid it entirely. Both are wrong about what AI actually does well and where it actually fails. The reality is more nuanced and more useful: AI is genuinely helpful for the mechanical writing work, useless for the evaluation work, and risky if you do not understand the difference.
This guide is different. It is written for small business owners and operators who are deciding whether and how to use AI in their actual review process, without an HR team to manage the rollout. You will get the 8 specific use cases where AI works (and where it does not), the 7-step responsible process, prompt examples you can adapt, the legal and ethical risks, and the common mistakes that turn AI from a productivity tool into a problem. I built FirstHR for this audience because most performance management content assumes either enterprise sophistication or AI hype.
What AI Can Actually Do for Performance Reviews
The simple working description: AI is a writing assistant that takes your observations and turns them into structured prose. It is not a manager. It cannot tell whether the employee performed well; it can only restructure or rephrase what you tell it. The quality of AI output is determined by the quality of your input. Garbage in, polished garbage out.
Three things AI does genuinely well for performance reviews. First, structural drafting: turning your bullet-point observations into well-organized prose with consistent sections (summary, strengths, development areas, goals). Second, rephrasing: converting vague phrases to specific ones, harsh language to neutral, technical observations to accessible language. Third, pattern analysis across multiple reviews: identifying whether you are consistently softer or harsher with certain employees, recycling generic phrases, or varying specificity. These three use cases save real time without creating risks if managed properly.
Three things AI does poorly or dangerously. First, evaluation: AI cannot judge whether the employee performed well; the manager must. Second, hallucinated specifics: AI tools sometimes invent plausible-sounding metrics, customer names, or project details that become legal exposure if disputed. Third, bias amplification: AI training data includes biased language patterns that can creep into output if the manager does not actively edit for them. A polished draft hides all three, which is why the edit pass matters.
AI in Performance Management Beyond the Review Document
The review is one step in a longer cycle, and AI fits the rest of that cycle the same way it fits the review: it shapes the words, never the verdict. Goal drafting, check-in summaries, and clustering open-text survey answers are all transformation work. Ratings, calibration, and promotion calls stay with the people who watched the work happen.
Four places in the year where it earns its keep at small business scale: turning a vague goal into a measurable one, pulling themes out of your own one-on-one notes before the review period closes, grouping survey comments so you read patterns instead of sixty separate paragraphs, and drafting an individual development plan from areas you have already identified yourself.
The boundary does not move as you walk through the cycle. Any step whose output is a decision about a person (who goes on a plan, who gets promoted, what rating lands on the form) sits outside what AI should touch, for the legal reasons covered below and the practical one that performance management only works when employees believe a human weighed the evidence.
Why This Looks Different at Small Business Scale
Most articles on AI for performance reviews are written for enterprise HR teams with formal AI policies, legal review of every tool, and dedicated People Operations specialists managing the rollout. None of that applies at small business scale. The owner-operator running a small business is the manager, the HR function, the AI policy, and the legal reviewer, all in one role.
Three implications for small business AI use. First, the time savings matter more. A founder doing reviews for 8-10 direct reports while running the rest of the business needs the productivity gains AI provides. According to an MIT experiment (2023), professionals using ChatGPT cut the time spent on writing tasks by 40%, a saving that counts most at small scale. Skipping AI entirely on principle ignores genuine productivity gains.
Second, the risks are also real and personal. At enterprise scale, an AI policy gone wrong creates a compliance issue. At small business scale, an AI policy gone wrong becomes the founder's problem directly, since the founder is usually the decision-maker, the writer, and the point of legal exposure all at once. Small businesses get less protection from organizational structure, so the discipline matters more.
Third, transparency is easier. A founder telling a small team how they use AI is a single 5-minute conversation. An enterprise rolling out an AI usage policy across every department is a quarterly initiative. Small businesses can move faster on disclosure and policy, which is exactly the right move because trust scales poorly when discovered after the fact.
8 Specific Use Cases: What Works, What Does Not
AI is not uniformly good or bad for performance reviews; specific use cases work well, others fail consistently. Below are the 8 most common applications and where each one fits.
The pattern across these use cases: AI is good at mechanical work (drafting, rephrasing, translating, analyzing), neutral at preparation work (self-review drafts, examples), and bad at decision work (compensation, ratings, evaluation). The line that matters: anything where the AI is restructuring or rephrasing your input is generally safe; anything where the AI is making or substantially informing decisions about employees is legally and ethically problematic.
A 7-Step Responsible Process for Using AI in Reviews
The process below produces AI-assisted reviews that save time without creating risks. The time AI saves comes out of the drafting, not the thinking. The steps are not optional; skipping any one introduces a specific risk that the others do not address.
The claim rows are the part that does the work. Copying each specific out of the draft and naming where you checked it takes a few minutes, and it is the only reliable way to catch an invented metric before it becomes performance documentation. The attestation line at the bottom is what you would produce if someone later argued the review was written by a machine.
Two failure modes to avoid. First, do not skip Step 1 (evidence collection). The temptation is to ask AI to write a review based on vague impressions, hoping it will fill in the specifics. AI cannot create specifics it does not have; it will either invent them (Step 3 violation) or produce generic content. Collect evidence first, always.
Second, do not skip Step 4 (bias check). AI training data includes biased language patterns from the entire internet; output without bias review can include phrases that create legal exposure even if the manager would never use them in their own writing.
Specific Prompts You Can Adapt
The prompts below are templates you can adapt to your specific situation. Each one includes the structural elements (role, observations, output format, constraint) that produce useful AI output rather than generic content.
Three rules for prompt writing. First, always include the constraint "do not invent any details I have not provided." AI tools have a tendency to hallucinate plausible-sounding specifics; explicitly constraining this reduces the rate. Second, give specific structural requirements (length, sections, format) rather than asking AI to choose them; the output is more predictable and easier to edit. Third, include observations as bullet points rather than narrative paragraphs; AI processes structured input more reliably and is less likely to alter your facts.
Tool Categories: Which AI to Use
The right AI tool depends on your context: privacy needs, integration with existing systems, budget, and team policies. Below are the main categories, with what each is best for and where to be careful.
| Tool category | Best for | Caution |
|---|---|---|
| General AI assistants (ChatGPT, Claude, Gemini) | Drafting from notes, rephrasing, bias checks, pattern analysis | Public versions log conversations. Use anonymized inputs or paid privacy-protected tiers |
| Built-in AI in performance management platforms | Teams already on a performance platform; integration with existing review workflow | Locked into vendor; data lives in vendor systems; quality varies widely between platforms |
| AI review generators (free tools) | Single review draft when starting from scratch | Often produce generic templates; content may resemble other users' reviews; data privacy varies |
| Privacy-focused AI tools | Sensitive reviews where data must not be logged | Often more expensive; verify data handling claims with vendor; check enterprise data agreements |
| Custom prompts on company-approved AI | Standardized review process across the team with consistent prompts | Requires upfront prompt development; train managers on consistent use |
The pattern: there is no universal best tool. Public general AI assistants are flexible and accessible but have privacy trade-offs. Built-in performance platform AI offers integration but vendor lock-in. Privacy-focused tools cost more but protect sensitive content. The right choice depends on what you optimize for. For most small businesses, a privacy-protected tier of a general AI assistant with anonymized inputs (Employee A, Project X) provides the best balance of flexibility and privacy.
The Risks of AI-Assisted Reviews
AI for performance reviews carries seven main risks, from bias and invented details to privacy leaks and legal exposure. Understanding each one is the first step to managing it.
| Risk | What it means | Mitigation |
|---|---|---|
| Bias amplification | AI training data includes existing biased language patterns. AI-generated reviews can perpetuate stereotypes (gendered descriptors, age-coded language, 'cultural fit' phrasing) that managers might not notice if they accept output without editing | Always read AI output specifically for protected-characteristic adjacent language. Replace any flagged phrases with observable behavior |
| Hallucinated specifics | AI tools invent plausible-sounding specifics: invented metrics, fictional customer names, fabricated quotes. These can appear authoritative and become legal exposure if disputed | Verify every metric, name, date, and quote against your actual evidence before saving the review |
| Generic content masking poor evaluation | AI produces structurally correct reviews even when the manager has not done the actual evaluation work. The review looks complete but contains no real assessment | If you cannot describe specifics in your own words, you are not ready to write the review yet, regardless of what AI produces |
| Privacy and data leakage | Public AI tools log conversations. Pasting employee names, sensitive feedback, and performance issues into public tools creates data exposure | Use anonymous descriptors (Employee A, Project X). Use privacy-focused or company-approved AI tools. Document your AI use policy |
| Legal exposure under disparate impact analysis | A tool used as a basis for an employment decision is a selection procedure under the Uniform Guidelines, which puts it inside disparate impact analysis under Title VII of the Civil Rights Act. Heavy AI reliance in reviews could expose the employer to claims if patterns emerge | Document your independent judgment and editing. AI assists; the manager decides. Retain notes showing human review of every AI output |
| Loss of manager learning | Writing reviews is one of the highest-leverage management practices. Managers who delegate to AI miss the development that comes from writing carefully about each direct report | Use AI to reduce mechanical writing time, not to skip the thinking work. The thinking is the development; the writing is just the artifact |
| Employee perception issues | A review that feels AI-written damages trust, even if the manager had genuine input | Personalize aggressively. Add specifics only you would know. Read aloud and edit until it sounds like you, not like a template |
The pattern across these risks: each one is mitigated by deliberate process, not by avoiding AI entirely. Bias is mitigated by editing for protected-characteristic adjacent language. Hallucinations are mitigated by verifying specifics. Privacy is mitigated by anonymized inputs or privacy-focused tools. Legal exposure is mitigated by documented human judgment. Manager learning is preserved by using AI for mechanical work, not thinking work. Employee perception is managed by personalization and disclosure.
Legal Considerations Small Business Owners Miss
The legal landscape around AI in employment decisions moves fast, but the rule underneath it is old. The Uniform Guidelines on Employee Selection Procedures define a selection procedure as any measure, combination of measures, or procedure used as a basis for any employment decision, language broad enough to cover an algorithmic model as squarely as a paper test.
That definition puts AI inside disparate impact analysis under Title VII, whose text reaches any particular employment practice. Once a procedure screens out a protected group at a materially lower rate, guidance on employment tests from the Equal Employment Opportunity Commission (EEOC) says the employer must show it is job-related and consistent with business necessity, after which a challenger may still point to a less discriminatory alternative.
Federal guidance on the point has been unstable. The EEOC's 2023 technical assistance document on AI and Title VII is no longer published, and the agency's artificial intelligence landing page returns a 404 (page not found) error.
The Uniform Guidelines themselves are in flux. On June 9, 2026, the Justice Department announced an Office of Legal Counsel opinion finding the EEOC disparate impact guidelines unconstitutional.
The EEOC regulatory agenda targets November 2026 for a final rule rescinding the disparate impact guidelines. Until a final rule takes effect, the definition above stays in the Code of Federal Regulations.
The disparate impact rule itself is written into the Title VII statute, which covers employers with 15 or more employees. A private plaintiff can still bring a disparate impact claim whatever happens to the Guidelines or to an agency's current enforcement priorities.
State and local law is filling the gap. New York City regulates automated tools used to screen candidates for hiring and employees for promotion. California's rules on automated-decision systems, in force since October 1, 2025, make clear that a tool harming applicants or employees on a protected basis can violate state law, and Colorado's rewritten AI Act follows on January 1, 2027.
Illinois reaches the review cycle most directly. Since January 1, 2026, it has been a civil rights violation for an employer to use AI, generative AI included, in a way that discriminates in promotion, discipline, discharge, or other terms of employment. Using AI for those decisions without notifying the employee is a violation too.
The table below sets out what to track across the whole landscape.
| Legal area | What it means | Practical action |
|---|---|---|
| Disparate impact analysis | If AI-assisted reviews produce different outcomes for protected groups, the employer can be liable under Title VII even without intentional discrimination | Document independent human judgment. Audit review patterns periodically. Apply consistent standards across all employees |
| Algorithmic decision-making | The Uniform Guidelines define a selection procedure broadly enough to cover an algorithmic model. Heavy AI reliance in compensation, promotion, or termination decisions creates exposure | Use AI for writing assistance only. Documentation, decisions, and judgment stay with the manager |
| Vendor liability | The Uniform Guidelines put the validation burden on the user of a selection procedure rather than its developer, so 'the vendor said it was safe' is not a defense | Ask vendors specifically what they have done to test for adverse impact. Document the answer. Audit periodically |
| Data privacy | Performance content typically includes sensitive personal information. Public AI tools that log conversations create data privacy exposure | Use anonymized inputs in public tools. Use privacy-focused tools for sensitive content. Document your data handling policy |
| Disclosure to employees | New York City requires a bias audit and 10 business days notice before an automated employment decision tool is used on someone, promotions included. Illinois requires notice whenever AI is used for decisions such as promotion, discipline, or discharge. Colorado's rewritten law adds notice and post-decision explanation duties from January 1, 2027 | Check applicable state and local laws. Default to transparency with employees regardless of legal requirement |
| Reasonable accommodation | AI tools may inadvertently disadvantage employees with disabilities. The Americans with Disabilities Act (ADA) requires employers with 15 or more employees to provide reasonable accommodation, whether or not AI is involved | Audit AI output for adverse effects on employees with known accommodations. Provide alternatives where AI creates barriers |
Three rules for legal protection when using AI. First, document your independent judgment. The strongest defense in any algorithmic decision claim is showing that the manager exercised judgment and the AI assisted writing only. Second, apply standards consistently across all employees regardless of AI use.
Consistency answers a disparate treatment claim (that you treated someone differently because of a protected trait), but a disparate impact claim (that a neutral practice hits a protected group harder) turns on whether the practice is job-related and consistent with business necessity. Third, audit periodically. Look at review patterns across your team to identify whether AI use is producing different outcomes for different groups.
Detecting Bias in AI-Generated Content
AI training data includes biased language patterns from the entire internet, including stereotyped descriptions of different demographic groups. AI-generated reviews can perpetuate these biases without the manager noticing if output is accepted without editing. Bias detection is one of the most valuable edits you can make.
| Biased language pattern | Common AI output | Replace with |
|---|---|---|
| Gendered descriptors for similar behavior | 'Bossy,' 'aggressive,' 'soft-spoken,' 'emotional,' 'mature' | 'Direct,' 'assertive,' 'measured,' 'thoughtful,' 'experienced' |
| Age-coded language | 'Energetic,' 'mature,' 'fresh perspective,' 'set in their ways' | 'Maintains consistent productivity,' 'demonstrates sound judgment,' 'brings new approaches,' 'consistent in their approach' |
| Family-status references | 'Despite family commitments,' 'works late despite obligations' | Remove family references entirely. Focus on work output and quality |
| 'Cultural fit' phrasing | 'Strong cultural fit,' 'fits well with our culture' | 'Collaborates effectively with team,' 'contributes to team norms,' 'supports team goals' |
| National origin/accent references | 'Speaks well for someone with their background' | 'Communicates clearly in written and verbal formats' |
| Disability adjacent | 'Reliable despite their condition,' 'overcomes their challenges' | 'Delivers on commitments consistently,' 'achieves results' |
The rule: read every AI-generated review specifically for these patterns before saving. The 60 seconds of bias review per review is the difference between AI-assisted reviews that protect against legal exposure and AI-assisted reviews that create it.
If you find these patterns appearing in AI output, the AI tool is doing what AI tools do: surfacing biases from training data. The fix is editing, not avoiding AI; the same patterns appear in human writing that has not been bias-checked.
A review does not land in a vacuum; it lands inside the working relationship between employee and manager. Gallup (2015) reports that one in two US adults have left a job at some point to get away from their manager, and that employees whose managers hold regular meetings with them are almost three times as likely to be engaged.
The manager relationship, in other words, is the real driver of engagement, and how a manager handles AI inside that relationship matters as much as which tool they picked.
Common Mistakes Small Business Owners Make
The mistakes below appear consistently across small businesses adopting AI for performance reviews. All are avoidable once you understand the underlying patterns.
The pattern across these mistakes: treating AI as a substitute for management work rather than as an assistant for the writing portion of management work. AI cannot create context, judgment, or relationship; it can only restructure what the manager has already developed. Managers who use AI to skip the development part of management produce thin reviews regardless of how polished the output looks.
The fix for most AI mistakes is reasserting the manager's primary role: collect evidence, evaluate, edit, decide. AI accelerates the writing; humans do everything else.
The payoff comes from what the manager does around the draft, and feedback frequency is where it shows. Gallup (2022) finds employees are 3.6 times more likely to strongly agree they are motivated to do outstanding work when their manager gives daily rather than annual feedback, so the quality of feedback matters more than the polish of the document carrying it.
What Stays Human in the AI-Assisted Review
The clearest framework for AI use in performance reviews: identify what AI does and what humans do, and never confuse the two. Below is the practical division.
| What AI does | What humans do |
|---|---|
| Structures bullet points into prose | Collects bullet-point observations from the period |
| Rephrases vague language to specific | Determines whether observations are accurate and complete |
| Translates technical to accessible language | Decides what observations matter and which to include |
| Suggests alternative framings for difficult feedback | Decides what feedback to give and how to deliver it |
| Analyzes patterns across multiple reviews | Decides what the patterns mean and what to do about them |
| Drafts goals based on stated development areas | Decides which development areas matter and which goals to pursue |
| Flags potentially biased language | Decides whether language is biased in this specific context |
| Translates to other languages | Decides what should be communicated in the conversation |
The pattern: AI handles transformation and analysis of existing material; humans handle judgment, evaluation, and decision. Reviews that respect this division are AI-assisted. Reviews that violate it are AI-substituted, and they produce the problems described throughout this article: generic content, hallucinated specifics, biased language, legal exposure, and broken trust with the team.
Telling Your Team You Use AI
Disclosure is one of the most underrated practices in AI use for performance reviews. Most managers either skip it (assuming no one will notice) or over-engineer it (writing formal AI policies before they are needed). The right approach is direct conversation: tell the team you use AI to assist with structure and writing, that the substance and judgment are yours.
| Approach | How it lands | When to use |
|---|---|---|
| No disclosure | Damages trust if discovered | Never. The risk of discovery exceeds any benefit |
| Brief verbal disclosure | 'I use AI to help structure first drafts; the observations and decisions are mine' | Most small businesses; covers the essential ethical ground |
| Written AI policy | Formal document covering allowed use, privacy, decision-making boundaries | Larger teams; companies in regulated industries; teams that have raised concerns |
| Policy plus training | Document plus session on responsible AI use across the team | Teams scaling AI use to multiple managers |
| Periodic transparency updates | Quarterly note on what AI use is producing, what is not working, what is changing | Teams adopting AI broadly; useful for building trust over time |
The single most important thing about disclosure: do it before AI use becomes invisible. Teams that learn after the fact that their manager used AI without telling them experience the discovery as a trust violation, regardless of how reasonable the AI use was. Teams that are told upfront usually accept AI use as a normal management practice. The disclosure is essentially free; the secrecy is genuinely expensive.
Where AI for Performance Reviews Is Going
The AI landscape for performance management is evolving rapidly. Three trends matter for small business owners thinking about AI use over the next 1-3 years.
First, AI capabilities are improving fast but not uniformly. Drafting and rephrasing have improved dramatically with each model generation; evaluation and judgment have not. The pattern is likely to continue: AI gets better at the mechanical work, stays bad at the judgment work. The framework in this article (AI assists writing, humans make decisions) should remain stable even as the tools improve.
Second, the rules keep moving, and state and local law is where the movement is. New York City is the clearest example. Before an automated employment decision tool is used on someone, the city requires a bias audit, public posting of the results, and 10 business days notice.
Colorado rewrote its AI Act in May 2026. The replacement applies to decisions made on or after January 1, 2027 and drops the impact assessments, but it still requires clear notice at the point of interaction with an automated tool that materially influences an employment decision, a plain-language explanation within 30 days of an adverse outcome, and human review on request.
The compliance requirements around AI in employment decisions will keep shifting from here. Small businesses that adopt deliberate AI processes now have less retrofitting work later, and the investment in process discipline pays back as the rules change.
Third, employee expectations are shifting. Employees have opinions about AI's use in evaluating their work. Disclosure norms are tightening; teams that handle AI use transparently build trust, teams that handle it opaquely lose it. The trajectory favors the small businesses that establish good practices now over the ones that defer the conversation.
How FirstHR Fits
FirstHR covers the operational HR foundations most small businesses need, starting with an AI Onboarding Wizard that builds a role-specific onboarding plan from the new hire's position.
Around it sit employee profiles, document management, org charts, and onboarding workflows. Your AI performance review drafting, when you adopt it, lives in the AI tool you pick from the categories above, with your own judgment on top, rather than in dedicated FirstHR software.
AI use for performance reviews also works better when the underlying people operations are working. A manager using AI to write reviews on top of broken onboarding will spend most of the AI-assisted writing time compensating for expectations that were never made clear to the new hire.
A manager using AI on top of consistent onboarding, clear documented roles, and structured employee profiles will produce AI-assisted reviews that actually drive growth. FirstHR exists to handle the operational HR foundation at flat, predictable pricing, so that owners can focus on the higher-impact work of evaluating employees well and writing reviews that matter.
Whichever tool you choose, start your next review the same way: write down your specific observations first, then open the AI.
Frequently Asked Questions
Can AI write performance reviews?
AI can turn your own specific observations into a structured first draft of a performance review, but it cannot judge how someone performed or take the manager's place. Used well, it handles much of the structural writing while you supply the substance: concrete examples, generic phrases cut, and an honest read of the person's results. Used carelessly, it returns boilerplate that changes nobody's behavior. Treat the draft as scaffolding rather than the finished review, because your editing is what turns a template into real feedback.
Is it ethical to use AI for performance reviews?
Yes, with three boundaries. First, be open with your team: say that AI helps with the structure of reviews while the substance and the judgment stay with you. Second, the manager makes evaluation decisions, not the AI: ratings, promotions, terminations stay human. Third, privacy: do not paste employee names and sensitive details into public AI tools that log conversations. AI as a writing assistant is similar to using a grammar checker or template; AI as the actual evaluator is ethically and legally problematic.
What are the best AI tools for performance reviews?
No tool wins in every situation, so start from what you need to protect and what you already use. General assistants such as ChatGPT, Claude, and Gemini handle drafting, rephrasing, and pattern analysis well, though their public versions keep a record of your conversations. AI inside a performance platform fits neatly into your existing review workflow, but your data stays with that vendor. Tools built around privacy cost more but give sensitive reviews better protection. Free review generators tend to return templated text. For most small businesses, a paid, privacy-protected plan of a general assistant, fed anonymized inputs, is the practical middle ground.
Can ChatGPT write a performance review?
ChatGPT can write a structured first draft of a performance review when given specific observations as input. The output requires significant editing to add specifics ChatGPT could not know, remove generic phrases that creep in from training data, and verify any specific claims. ChatGPT-only reviews without manager editing produce template content that employees recognize and that fails to drive behavior change. Use ChatGPT for the mechanical writing; use your own judgment for the substance.
Should I tell my team I used AI to write their reviews?
Yes. Trust matters more than the time savings. If team members discover undisclosed AI use, the trust damage exceeds any efficiency gain. The honest framing works: 'I used AI to help structure first drafts; the observations, evaluation, and final editing are mine.' Most teams accept this. Few accept secrecy. Disclosure also creates accountability for managing AI well, which produces better reviews than secret use.
What are the risks of using AI for performance reviews?
Seven main risks. Bias amplification (AI training data includes biased language patterns). Hallucinated specifics (AI invents plausible-sounding details). Generic content masking poor evaluation. Privacy and data leakage (public AI tools log conversations). Legal exposure, since the Uniform Guidelines sweep any decision-basis tool into their definition of a selection procedure. Loss of manager learning (writing reviews develops management capability). Employee perception issues (AI-written content damages trust if discovered). All risks are mitigatable, but they require deliberate process: editing, verification, privacy protection, and disclosure.
Can AI determine performance ratings or recommend promotions?
No, and handing those calls to a model creates real legal exposure. The relevant regulation is 29 CFR 1607.16, part of the Uniform Guidelines on Employee Selection Procedures, and its definition of a selection procedure turns on function rather than form: whatever you actually use as the basis for an employment decision qualifies. A rating engine, a promotion recommender, and a scored termination list all land inside that definition, which means they sit inside Title VII disparate impact doctrine. Should the outputs fall unevenly across a protected group, you carry the burden of proving the procedure is job-related and consistent with business necessity, and even a successful defense leaves room for a plaintiff to propose something that works as well with less adverse effect. Agency guidance in this area has come and gone, and the EEOC has scheduled the Uniform Guidelines for rescission, but disparate impact liability sits in the text Congress added to Title VII in 1991, so it outlasts them. Keep the drafting with AI and the deciding with people, and write down the reasoning behind each decision.
How do I write a good prompt for performance review AI?
A good prompt has four parts. First, the role and review type ('annual review for a customer success manager'). Second, your specific observations as bullet points (8-15 examples with specifics). Third, the structural requirements ('2-3 sentence summary, 3-5 strengths with situation-behavior-impact, 2 development areas'). Fourth, the explicit constraint: 'Do not invent any details I have not provided.' The constraint is critical because AI tools sometimes hallucinate plausible-sounding specifics. Verify everything in the output against your actual evidence.
How much time does AI save on performance reviews?
AI saves time on the writing, not on the evaluation. The evidence comes from writing tasks in general rather than reviews specifically: in a 2023 MIT experiment, college-educated professionals who used ChatGPT on occupation-specific writing assignments cut their completion time by 40%, and the quality of their work rose by 18%. The time savings come from the mechanical writing work, not the thinking work. The manager still needs to collect evidence, evaluate the employee, verify specifics, and conduct the conversation. AI does not save time on the parts that actually matter; it saves time on the parts that are mechanical.
Can AI detect bias in my performance reviews?
Yes, partially. AI is reasonably good at flagging gendered descriptors (bossy, soft-spoken), age-coded language (mature, energetic), 'cultural fit' phrasing, and family-status references. AI can also compare several reviews you wrote and show whether you tend to go easier or harder on particular people. AI is less good at detecting subtler bias patterns (whose work gets praised vs. whose gets credited to circumstance). Use AI bias detection as a first pass, not as the final check; have a colleague review for nuanced patterns AI misses.
Should small business managers use AI for performance reviews?
Most should, with the boundaries described in this guide. Small businesses have less administrative bandwidth than enterprise HR teams; AI is a force multiplier for the writing portion. The constraints are the same as for any size: AI assists writing, manager makes decisions, edit aggressively, verify specifics, protect privacy, disclose to the team. Small businesses that follow these constraints get the time savings without the risks. Small businesses that skip them end up with boilerplate reviews that change no one's behavior and may create legal exposure.
What is the difference between AI-assisted and AI-written reviews?
AI-assisted reviews use AI for drafting, rephrasing, or analysis while the manager retains evaluation, editing, and decision-making. AI-written reviews delegate the evaluation work to the AI, with the manager simply approving or lightly editing the output. The first approach is a productivity tool; the second is a delegation of judgment that AI is not capable of exercising. Reviews that read as AI-written damage trust with the team and may create legal exposure. Reviews that are AI-assisted save time without these trade-offs.
Make HR the easy part of running your business
Onboarding, e-signatures, training and core HR in one place. Up and running in 5 minutes.
- 7-day free trial
- No credit card required
- Cancel anytime