Iro AIIro AI Blog
AI is good at turning your notes into a clear, specific review and has no idea what anyone actually did. Use it on your evidence, never instead of it, and check your company's policy before an HR document goes anywhere near a chatbot.
Yes, as long as the AI works from your evidence and you make every judgment: use it to organize notes, sharpen wording and catch vague or biased language, never to invent what someone did or decide how well they did it.
That rule is the same whether you're a manager writing reviews for your team or an employee writing a self-review. ChatGPT, Claude or Copilot can turn a page of scrappy notes into a clear structure in a minute. None of them knows what happened in your team this year. Anything they add beyond your notes is a guess written in a confident voice, and in a review that guess ends up attached to a real person.
Before you start, check your company's policy. Some employers ban AI for HR documents, some allow only an approved tool, and some have no rule yet. A review may end up in someone's employment record, so treat the text like any other confidential file.
| Use AI to | Never use AI to |
|---|---|
| Turn scattered notes into a structure: strengths, a growth area, goals | Invent accomplishments, numbers or quotes |
| Tighten wording and cut filler | Decide a rating or whether someone met expectations |
| Flag vague phrases like "great team player" and ask for the example | Compare people or rank a team |
| Spot comments about personality rather than work | Receive health, pay, disciplinary or investigation details on an unapproved tool |
| Find claims your notes don't support | Write feedback you haven't read and agreed with |
The right-hand column isn't only good manners. The EU's AI Act lists AI systems intended to "monitor and evaluate the performance and behaviour of persons" in work relationships among its high-risk uses. Asking a chatbot to tighten your sentences isn't the same as building a system that rates staff, but the direction is clear: the judgment belongs to a person.
Practice this, don't just read it.
Iro AI turns ideas like the ones in this post into 5-minute exercises with feedback. Free tier, Pro from $4.17 a month ($49.99 a year, 7-day free trial).
Here's the prompt that produces a review that could describe anyone:
Write a performance review for Priya, my software engineer. She's had a good year but needs to communicate more.
It has a real name in it, no evidence, and a growth area so vague the model will pad it with stock phrases. This one starts from what actually happened:
I'm writing an annual review for a mid-level backend engineer (call them R). Use only the evidence below. Don't add achievements, numbers or adjectives the notes don't support. If a section has no evidence, write [NEED EVIDENCE] instead. Evidence: led the move of billing jobs off the old scheduler, finished in March, two weeks late, no customer-facing incidents. Cut median response time on the orders API from 420 ms to 180 ms (dashboard, June vs September). Reviewed about 140 pull requests; two teammates named R's reviews as the most useful in peer feedback. Growth area: design docs arrive late, so decisions get made in chat threads; happened on the billing move and the search project. Structure: three strengths with evidence, one growth area with an example and a next step, two goals for next half. Plain language, under 350 words. Don't assign a rating.
Notice what's missing: the name, the rating, and anything personal. The [NEED EVIDENCE] rule matters most, because it stops the model papering over a gap with a confident sentence.
Once you have a draft, written by you or with AI, run an audit. This is where AI earns its place in the process:
Here is my draft review. Don't rewrite it. List: (1) any sentence that makes a claim my evidence doesn't support; (2) praise or criticism with no example, such as "great attitude" or "needs to be more proactive"; (3) comments about personality rather than work and its effects; (4) words that might read differently depending on who they describe, such as "abrasive" or "emotional". Quote each one and tell me what evidence would fix it.
If you manage several people, read the audited drafts side by side yourself. The question is whether similar wording rests on similar evidence for everyone. Reviewing AI-assisted work is a manager skill in its own right, and our AI for managers page covers the rest of it.
The blank-page version gets you a self-review full of "spearheaded" and "drove impact":
Write my self-evaluation. I'm a marketing manager and I had a great year.
Give it facts instead, including the inconvenient ones:
Help me write my self-review. I'm a marketing manager. Use only the facts below and don't inflate them; if a sentence sounds bigger than the facts, flag it. Facts: ran the spring webinar series, 6 sessions and 1,900 registrations (from the webinar platform report). Rebuilt the onboarding email sequence; trial-to-paid conversion went from 8.1% to 9.4% between Q1 and Q3, but product changes shipped in the same window, so I can't claim all of it. Missed: the case-study program, 2 published of 6 planned, because customer approvals stalled. July feedback from my manager: warn earlier when a deadline is at risk. Write a three-sentence summary, three accomplishments (what I did, the result, how we know), one honest miss with what I learned, and two goals. First person, confident but not boastful, under 300 words.
The line about product changes is what makes a self-review credible. Left alone, a model will happily turn "conversion rose while I worked on onboarding" into "drove a 16% lift in conversion". Don't let it.
Then ask one more question: Which sentences would my manager push back on, and what evidence would they want? The answer tells you what to dig out before the meeting. The same evidence-first method works for a resume, and it works best if you keep a running file of wins through the year.
HR teams writing the rules for this will find more on bias and confidentiality on our AI for HR page. If you're new to AI at work more generally, how to use AI at work covers protecting company data in everyday tasks.
Using AI on writing that matters comes down to two habits: giving it real evidence, and knowing what it shouldn't touch. Iro trains both in five-minute lessons with instant feedback, including delegate-versus-do-it-yourself calls, guardrail writing and prompt repair. There's a managers path for leading and reviewing AI-assisted work, and our AI for work page lists the other work skills Iro covers.
Practice this, don't just read it.
Iro AI turns ideas like the ones in this post into 5-minute exercises with feedback. Free tier, Pro from $4.17 a month ($49.99 a year, 7-day free trial).
Yes, if your company allows it and you use it on your own evidence. Give it your notes, metrics and examples, ask for structure and clearer wording, and tell it not to add anything your notes don't support. Keep names and confidential details out, decide the rating yourself, and read every sentence before you submit.
Paste a list of what you did, the result and how you know it, plus one honest miss and any feedback you received. Ask for a short summary, three accomplishments, a growth area and goals, in first person, without inflating anything. Then ask which sentences your manager would push back on and what evidence they would want.
It depends on your employer's policy, so check that first. Where it's allowed, AI is reasonable for structuring notes, tightening wording and flagging vague or biased language. It isn't appropriate for inventing accomplishments, deciding ratings or comparing employees, and confidential HR details shouldn't go into a tool your company hasn't approved.
They might if it reads generic: inflated verbs, no numbers, and phrases that could describe anyone. A self-review built from your own facts, with specific results and an honest miss, reads like you whatever helped with the wording. If your company asks employees to disclose AI use, disclose it.
Names, health information, family circumstances, pay, disciplinary or investigation details, and anything your HR team treats as confidential, unless your company has approved the tool for exactly that use. Use initials or a role instead. The model doesn't need to know who someone is to improve the wording of your evidence.
It can flag patterns worth a second look, such as personality-based criticism, vague praise and words that read differently depending on who they describe. It can't guarantee fairness, because it can reproduce bias too. Treat its flags as prompts for your own reread, and check that similar ratings rest on similar evidence across your team.