TL;DR: AI performance reviews work best when AI summarises feedback, spots patterns across a cycle, and drafts a first version. The manager rewrites that draft and decides the rating, and employee feedback never goes into a public chatbot. Set the rules before the cycle starts, because managers are already experimenting.

Key takeaways

• In UNSW research with 242 managers, about 60 per cent followed a high algorithmic rating, but only about 42 per cent followed a low one.
• Letting managers adjust how the algorithm weighs its inputs lifted use of low ratings to 63 per cent. Adjusting the final score did not help.
• The Australian privacy regulator recommends keeping personal information out of public AI tools, which rules out pasting reviews into a free chatbot.
• According to AHRI, 68 per cent of Australian organisations now have formal AI policies. Few of those policies say anything specific about performance reviews.
• AI built into a performance system can summarise without sending employee names to the model, so ask whether yours does.

Can AI write a performance review?

AI can draft a performance review, and it does so quickly. However, the manager must own the final words and the rating, because pay and promotion ride on the judgement. Use AI to reach a first draft, then rewrite it in your own voice with your own evidence.

The question comes up in almost every review cycle now, and for good reason. Writing reviews is slow, and most managers write several at once. A first draft that pulls together self-review comments, peer feedback and goal progress can save an hour per person.

The risk is in what the draft leaves out. For instance, an AI draft cannot know that a restructure caused a missed target, or that a quiet employee carried a project. So treat the draft as a structure to argue with, not a conclusion to sign.

Employees notice the difference, too, and a review that reads as though nobody wrote it lands badly, however accurate it is. The manager’s own examples, in the manager’s own words, are what make feedback believable.

Where does AI genuinely help in performance reviews?

AI helps most with the reading, sorting, and drafting around a review, rather than the judgement at its centre. It can summarise long feedback into themes, spot patterns across a whole cycle, and suggest which conversations to have next. Those jobs are slow for people and quick for AI, and a manager can check the output easily.

In practice, four uses hold up well:

• Summarising feedback: AI condenses self-reviews, peer comments, and one-on-one notes into themes.
• Spotting patterns across a cycle: AI can flag a team where ratings cluster at the top, or feedback that repeats.
• Drafting a first version: AI structures the evidence under each goal, and the manager rewrites, corrects, and adds judgement.
• Suggesting next steps: AI can propose development conversations, check-ins, or training based on the feedback.

Suppose an HR lead closes a cycle of 120 reviews. Reading every one to find where to follow up takes days. A cycle summary that surfaces teams, patterns, and people worth checking in with turns that into an hour. The HR lead still decides what to do.

What can go wrong when managers use AI for reviews?

Three things go wrong most often: bias survives, personal information leaks into the wrong tools, and the review stops sounding like the manager. Each has a practical fix that needs no ban on AI, only a rule written down before the cycle opens.

Bias does not disappear

Many organisations adopt algorithmic tools to take bias out of ratings. However, research from UNSW Business School, published in The Accounting Review in 2026, shows managers use algorithms selectively. In an experiment with 242 experienced managers, about 60 per cent used the algorithm when it recommended a high rating. When it recommended a low rating, only about 42 per cent did.

The researchers linked that reluctance to managers protecting relationships, rather than any view that the algorithm was wrong. As a result, the leniency the tool was meant to remove comes back, now with a look of objectivity. The article also notes that 60 to 70 per cent of employees typically receive ratings in the top two levels.

Personal information ends up in the wrong place

A manager who pastes a review into a free chatbot sends personal information outside your systems. As best practice, the OAIC recommends not entering personal information into publicly available generative AI tools. Performance feedback is personal information by any measure, and often sensitive.

The review stops sounding like the manager

Generic phrasing, overlong praise, and the same structure for every employee all signal machine drafting, and trust in the process drops when employees notice. So keep AI on the first draft and the summary, and keep the manager on the final words.

Should AI decide performance ratings?

AI should inform performance ratings, not decide them. A rating affects pay, promotion, and development, so a person must make it and be able to explain it. Where AI suggests a rating, managers need to see how it was reached. Otherwise they ignore it when it is unwelcome, which is exactly when it matters.

The UNSW research tested two ways of giving managers control. Letting managers adjust the final score did not make them any likelier to use a low rating. By contrast, adjusting how the algorithm weighed its inputs lifted use of low ratings to 63 per cent.

The lesson for HR is about transparency, not technology. If AI contributes to a rating, show managers what went in, explain how it was weighed, and let them question it. Otherwise, the tool quietly becomes a way to confirm the rating a manager already wanted to give.

What rules should HR set before the next review cycle?

Set four rules before the cycle opens. Name the approved tools and permitted data, limit AI to drafting and summarising, decide when to disclose AI use, and mark generated content. Write them into your AI policy and brief managers at cycle launch.

Timing matters, because managers are already experimenting. According to AHRI’s December quarter 2025 Work Outlook, 68 per cent of organisations have formal AI policies. Three-quarters are also training staff to use AI. Most of those policies were written for AI in general, though, and a review cycle raises specific questions they rarely answer.

Four rules cover most of the risk:

• Approved tools only: name the tools managers may use, and ban pasting employee feedback into public chatbots.
• Draft and summarise, never rate: the manager decides, records, and explains every rating.
• Disclose where it matters: decide whether employees are told when AI helped draft their review, and apply it consistently.
• Mark generated content: anything AI produced is visibly marked, so nobody mistakes a summary for a manager’s assessment.

What does AI built into a performance system do differently?

Built-in AI works on review data inside the system, under its permissions, rather than on text pasted elsewhere. Done well, it can summarise without sending employee names to the model, check output before anyone sees it, and mark every summary as generated. Ask any vendor to show you each of those.

Worknice’s AI cycle summaries are one example of what to look for. Completing a review cycle in Worknice generates a plain-language summary for account owners only. It covers cycle health, patterns, people worth checking in with, earlier cycles, and suggested next steps. Names are never sent to the model: input is aggregated against internal IDs, and names are restored on display. Leave and integration data stay out of the input entirely.

Each summary must also pass an automated check for protected-attribute terms, leaked names and invented people before it appears. It then carries a footer recording that it was AI generated, with the model and prompt version. Cycles with fewer than four eligible reviewees produce no summary at all. Worknice also offers an AI packet summary, which the employee, their manager and account owners can see.

So Worknice suits mid-sized Australian organisations that want AI summaries in their reviews without employee names reaching the model. An organisation wanting AI-suggested ratings or enterprise calibration analytics may prefer a platform built around those. The guide to Australia’s favourite performance review software compares the wider field.

Frequently asked questions

Can AI write a performance review?

AI can write a first draft of a performance review, pulling together self-review comments, peer feedback and goal progress. The manager should then rewrite it with their own examples and make the rating themselves. A review affects pay and promotion, so a person must own the judgement and be able to explain it.

Is it OK to use ChatGPT for performance reviews?

Not with real employee information in a public version. The Australian privacy regulator recommends not entering personal information into publicly available AI tools. If your organisation approves an AI tool for reviews, use that. Otherwise, use AI only for general help, such as structuring feedback, with no names or identifying details.

What are good AI prompts for performance reviews?

Useful prompts contain no personal information at all. For example: “Suggest a structure for a mid-year review covering goals, strengths and development.” Or: “Rewrite this feedback to be specific and actionable”, with names removed. Or: “List questions a manager could ask in a development conversation.” Then add your own evidence and judgement.

Should employees be told if AI was used in their review?

Decide as an organisation and apply the answer consistently. Many employers disclose where AI drafted or summarised content, because employees tend to find out anyway and trust suffers when they do. At a minimum, AI-generated content should be marked as generated, and the rating should always come from a person.

Does AI remove bias from performance reviews?

Not on its own, because bias returns through selective use. UNSW research found managers followed a high algorithmic rating about 60 per cent of the time, but a low one only about 42 per cent. Transparency about how AI reached its view, plus clear rules, does more than the tool itself.

About the author

Graham Martin is Co-founder of Worknice, an Australian HRIS built for mid-to-large organisations. He has spent more than a decade working with Australian People and Culture teams on HR systems, compliance and payroll integration, including many software evaluations from the buyer’s side of the table.

Related reading

• What is a performance review platform, and how does it work?
• Performance management vs talent management
• Best AI HR software in Australia
• Questions to ask HR software vendors about AI

Sources

1. Humphreys, Kerry, and Mandy Cheng. “What’s behind the hidden manager bias risk in AI performance reviews?” UNSW BusinessThink, 24 May 2026, reporting Lin, Cheng and Humphreys, “Tough Ratings, Tougher Sell”, The Accounting Review. https://www.businessthink.unsw.edu.au/articles/ai-hr-algorithmic-performance-management-bias
2. Office of the Australian Information Commissioner. “Guidance on privacy and the use of commercially available AI products.” OAIC, 21 October 2024, updated 17 January 2025. https://www.oaic.gov.au/privacy/privacy-guidance-for-organisations-and-government-agencies/guidance-on-privacy-and-the-use-of-commercially-available-ai-products
3. Australian HR Institute. “Quarterly Australian Work Outlook, December Quarter 2025.” AHRI, November 2025. https://www.ahri.com.au/resources/hr-research/ahri-quarterly-australian-work-outlook-december-2025