How to use AI to draft performance reviews without losing your team's trust

The temptation of the empty text box

It is that time of the quarter again. You have a spreadsheet of ten direct reports, a looming deadline from HR, and a completely blank document staring back at you. Your mind is mush, but you want to do right by your team. Naturally, the temptation to open up /platforms/claude or /platforms/openai, dump in a few bullet points, and ask it to "write a professional, constructive performance review" is incredibly strong.

And why shouldn't you? It saves hours of agonizing over phrasing. It smooths out rough edges. It makes you look like a polished, eloquent leader.

Except, it usually doesn't. When you hand the keyboard over to an LLM to write feedback about a human being's actual career, you run a massive risk. People can spot synthetic praise from a mile away. Nothing says "I do not actually care about your career" quite like a performance review laden with telltale AI adjectives like testament, proactive synergy, and pivotal contributions.

If your team members realise you have outsourced your basic management duties to an algorithm, you will lose their trust instantly. Here is how to use AI to help you write reviews ethically, constructively, and without losing the human connection that makes your team tick.

Why generic AI feedback is a betrayal of trust

Performance reviews are not just HR tick-box exercises (though they often feel like it). For your team, they are high-stakes evaluations that dictate promotions, raises, and professional self-worth.

When you use an AI tool to write these reviews without a strict set of ethical guardrails, several things go wrong:

1. You erase the nuance: LLMs love to normalise. They take raw, specific human behaviour and sand down the edges until it sounds like a corporate press release. Your employee's unique quirks and actual achievements get lost in translation.
2. You risk hallucinations: If you ask a model to "make this sound more substantial," it might start inventing specific projects or outcomes the employee had nothing to do with. If you do not catch this, you look completely out of touch.
3. You abdicate your responsibility: Part of being a good manager is doing the hard mental work of thinking deeply about your team's performance. If you let the AI do the thinking for you, you miss out on the genuine insights that help you guide them.

To avoid these traps, we need to treat the AI as an editor and structural assistant, never as the author of your opinions. Check out our /glossary if you want to understand more about how these models tend to generalise text.

The "scaffolding" rule: How to write the prompt

The golden rule of ethical AI assistance is simple: You must provide the substance; the AI only provides the structure.

Never ask an LLM to "evaluate" an employee based on a vague prompt like: "Write a review for Sarah, she is a senior developer who did well on the migration project but needs to work on communication." This forces the model to invent details to fill the gaps.

Instead, use a structured prompting approach. Feed the model highly specific, raw notes, and explicitly tell it what not to do.

Here is a template you can adapt for your own prompts:

`text Act as a professional developmental editor. I am writing a performance review for a team member. I will provide raw, unedited bullet points detailing their achievements and areas for growth.

Your task is to: 1. Organise these points into a clear, professional structure (e.g., Strengths, Areas for Development, Next Steps). 2. Improve the clarity of my writing without adding any new facts, achievements, or criticisms that I have not provided. 3. Avoid generic corporate buzzwords. Keep the tone warm, direct, and constructive. 4. Highlight any areas where my raw notes are too vague or need more concrete examples before we finish.

Here are my raw notes: [Insert your specific bullet points here] `

By framing the AI as an editor rather than a ghostwriter, you remain the source of truth. You are still the one doing the hard work of observing and evaluating; the AI is simply helping you articulate those thoughts clearly.

When to shut the tab and write it yourself

There are certain performance scenarios where you should absolutely not use AI, even as an editor. Using an LLM in these situations is not just lazy—it is ethically indefensible.

1. Delivering tough, corrective feedback If an employee is struggling, failing to meet expectations, or on the verge of a Performance Improvement Plan (PIP), do not touch AI. This feedback must be incredibly precise, deeply empathetic, and entirely human. If you use an LLM to soften the blow, you might end up obscuring the severity of the issue, leaving the employee confused about where they stand. Write this yourself, word by painful word.

2. Resolving interpersonal conflicts If you are writing about a team member's challenging behaviour or conflicts with other colleagues, keep the AI out of it. LLMs do not understand team dynamics, office politics, or the delicate history of your workplace. They will almost certainly suggest platitudes that fail to address the root cause of the tension.

3. Discussing personal struggles If an employee's performance has dipped due to personal issues, health problems, or burnout, this requires extreme sensitivity. An AI cannot feel empathy; it can only simulate it. Writing this review yourself shows the respect and care your team member deserves during a difficult time.

The final human sanity check

Once the AI output is generated, your job is only half done. Do not simply copy, paste, and submit. You must review the text with a critical eye.

Read the draft aloud. Does it sound like you? If you sat down with this employee for a coffee, would you actually use these words? If the draft contains phrases you would never say in real life, delete them. Replace them with your actual voice. Your team knows how you talk; if your written feedback sounds like an HR manual written by a silicon chip, they will feel the disconnect.

Using AI to help with the heavy lifting of writing is sensible. But remember: you manage people, not data points. Keep your reviews honest, keep them specific, and keep them human.