How to Use AI for Performance Reviews in a Small Business

Performance reviews are one of those jobs that small business owners quietly dread. You know they matter — people leave managers, not companies — but writing six or twelve thoughtful reviews means pulling together a year of scattered evidence, finding language that is honest without being brutal, and doing it in the same fortnight as everything else. So reviews get postponed, then rushed, then written from memory of the last three weeks rather than the last twelve months.

This is exactly the kind of work AI is good at: structuring messy evidence, drafting, and pressure-testing your own wording. It is also work where AI can do real damage if you hand it the wrong job. This guide covers both — the four-step workflow we recommend, the prompts to copy, and the firm lines not to cross.

Why AI fits performance reviews — and where it stops

A performance review has four distinct tasks inside it, and they are not equally suited to AI.

• Gathering evidence — collating what someone actually delivered over the period. AI is genuinely useful here if the evidence exists in writing.
• Judging performance — deciding whether that adds up to "exceeding" or "needs support". This stays with the manager. Always.
• Drafting the write-up — turning your judgement into clear, specific, non-waffly prose. AI is very good at this.
• Preparing the conversation — anticipating objections, rehearsing difficult phrasing. AI is a useful sparring partner.

The failure mode in almost every bad AI review is the same: the manager outsources task two along with the rest. The output reads fine and means nothing, because no human actually formed a view. Reviews written that way are worse than no review at all — staff can tell, and the trust cost is expensive to repair.

Realistically, a manager who spends four hours on six reviews can expect to spend around 90 minutes with this workflow, and the reviews will be more specific rather than less, because the evidence-gathering step stops being the thing you skip.

What AI should never do in a review

Four hard lines, and they are not just ethical — three of them are legal exposure.

1. Never let AI assign the rating or score. A model guessing at "3 out of 5" from a pile of Slack messages is not an assessment; it is a plausible-sounding number with no accountability behind it.
2. Never feed in an employee's personal or health information. Sick leave, a disability adjustment, a grievance, counselling — none of it belongs in a prompt, and most of it does not belong in a performance review at all.
3. Never use AI to monitor sentiment, tone or emotion. Under the EU AI Act, emotion-inference systems in the workplace are a prohibited practice, not merely high-risk. Tools that score how "engaged" someone sounds in meetings are the wrong side of that line.
4. Never paste in one employee's review to generate another's. It is a data-protection problem and it produces templated reviews that read as interchangeable, which they are.

Worth knowing: the EU AI Act classifies AI used to evaluate employees for promotion, termination or task allocation as high-risk, which brings transparency and human-oversight duties with it. Using a general assistant to help you draft your own assessment is not the same as deploying an automated evaluation system — but the distinction rests on you making the decision, and being able to show that you did. Our EU AI Act guide for small businesses walks through which obligations actually land on a company of your size.

The four-step AI review workflow

Step 1: Build an evidence file before you open the AI tool

This is the step that makes everything else work, and it takes about 15 minutes per person. Pull together, in one document: completed projects and their outcomes; the person's own self-assessment if you run one; any client or colleague feedback in writing; agreed objectives from the last review; and two or three concrete moments you remember — good and bad — with dates.

Numbers help more than adjectives. "Cut the quote turnaround from four days to one" is evidence. "Very proactive" is a feeling you will struggle to defend if challenged.

Step 2: Form your own view, in rough notes, before drafting

Write three bullet points: what this person did well, what needs to change, and what you are asking of them next. Ugly, unpolished, five minutes. This is the judgement step, and doing it before you prompt anything is what keeps the AI from quietly forming the view for you.

Step 3: Have AI draft from your evidence and your view

Now the drafting prompt does what it is good at — turning your rough notes and raw evidence into a structured, specific review. Expect to rewrite perhaps a quarter of it. The parts you rewrite are usually the parts where your own view was vaguer than you thought, which is useful information.

Step 4: Stress-test it before the conversation

Ask the model to read the draft back as the employee would. This catches the two most common problems in SMB reviews: criticism so softened that the person leaves thinking everything is fine, and praise so generic that it lands as indifference.

Prompts you can copy

Use these as written, with names and identifying details removed. Replace the bracketed parts.

The drafting prompt:

You are helping a manager at a [12-person marketing agency] draft a performance review for a team member in the role of [account manager], covering [January to December 2026].

Here is my own assessment in rough notes: [paste your three bullets].

Here is the evidence: [paste the evidence file].

Draft a review with four sections: strengths with specific examples, areas for development with specific examples, progress against last period's objectives, and three proposed objectives for the next period. Use concrete detail from the evidence — do not add achievements that are not there. Do not assign a rating or score. Where my notes are too vague to support a claim, say so instead of filling the gap.

That last sentence matters more than it looks. Without it, models smooth over thin evidence with confident-sounding filler, which is precisely what you cannot defend in a disagreement.

The honesty check:

Read this draft review as if you were the employee receiving it. Tell me: what is the single clearest piece of criticism here, and would I notice it on one read? Which praise is generic enough to feel like filler? Is any part ambiguous enough that I could reasonably think I am doing fine when the draft means the opposite?

The difficult-conversation rehearsal:

I need to tell a team member that their work quality is fine but they are missing deadlines, roughly two in five. Give me three ways to open that conversation, each in two sentences. Then list the four most likely responses and how I should answer each, bearing in mind I want this person to stay.

The objectives prompt:

Turn these three development areas into objectives for the next six months: [paste]. Each needs a measurable outcome, a deadline, and one thing I am responsible for providing — training, budget, time, or a change on my side. Flag any objective that depends on something outside this person's control.

That final instruction is the one managers thank us for. A surprising share of "performance problems" in small teams are really resourcing problems wearing a performance costume.

Privacy, and what to keep out of the prompt

Performance data is personal data under the UK GDPR and EU GDPR, and employment records attract more scrutiny than most. Three practical rules keep you on solid ground.

Use a business-tier account, not a personal one. Consumer plans may train on your inputs; business and team tiers generally do not by default. If you are reviewing staff on a free personal account, you have a data-processing problem regardless of how careful your prompts are.

De-identify by habit. Use "the account manager" rather than a name. It costs nothing and it means an accidental paste into the wrong window is an embarrassment rather than a breach.

Tell your team you do this. Not a legal necessity in every case, but a trust one — and discovering it accidentally is far worse than being told. A single line in your AI policy does the job: which tools are used for what, that no AI assigns ratings, and that a named human makes every decision. If you have not written that policy yet, our guide on how to write an AI policy for your small business has a template you can adapt in an afternoon.

How to roll this out in a team of five to fifty

Do not introduce this across the whole company in review season. A sensible sequence:

1. One manager, one cycle, as a trial. Ideally you. Note where the AI helped and where you rewrote it.
2. Write the two-page internal guide. Which tool and tier, the four hard lines above, the prompts, and a reminder that the manager owns the judgement. Two pages, not twenty.
3. Keep the evidence file running all year. The biggest gain is not the drafting — it is a short monthly note per person. Ten minutes a month removes the annual scramble entirely, and it is the same habit that makes AI-assisted employee onboarding work.
4. Review the reviews. After the cycle, read a sample and ask one question: could a stranger tell these were written about different people? If not, the AI is doing too much and the managers too little.

The bottom line

AI will not make you a better manager, and it will not tell you how your people are doing. What it will do is remove the administrative friction that makes reviews late, thin and generic — the collating, the structuring, the search for honest phrasing that does not sting more than it needs to. Keep the judgement, delegate the drafting, keep personal and health information out of the prompt entirely, and make sure every rating has a human name attached to it. Done that way, a review cycle that used to eat a week becomes two focused days, and the conversations get better rather than blander.