How to Use AI to Draft Hard Professional Feedback (Without Sounding Like an Unfeeling HR Bot)

The Temptation of the Emotional Buffer

Giving constructive feedback is, without a doubt, one of the most taxing parts of running a team. It requires tact, empathy, clarity, and courage. Naturally, when faced with the prospect of telling a colleague that their recent work has missed the mark, the temptation to open a chat interface and say, "Rewrite this bulleted list of grievances into a gentle but firm performance review" is incredibly high.

But there is a trap here. Left to their own devices, most large language models suffer from a severe case of corporate sanitisation. They default to a weirdly clinical, passive-aggressive dialect—let's call it 'HR-ese'—that feels deeply alienating to receive. If your team member reads your feedback and immediately senses it was spat out by a machine, you have not just failed to deliver the message; you have actively damaged the trust between you.

Using AI to help draft difficult feedback isn't inherently unethical. What is unethical is outsourcing your personal judgement and emotional responsibility to a statistical model. Here is how to use LLMs as a sounding board to find the right words, whilst keeping your humanity—and your colleague's dignity—fully intact.

1. The Golden Rule: Protect Privacy First

Before you type a single prompt, you must address the data privacy elephant in the room. Dumping a colleague's full name, specific performance history, and personal struggles into a public model training loop is a massive breach of trust (and, depending on where you operate, potentially illegal).

Never feed identifiable details into an LLM. Use placeholders for names, switch the industry if it is highly niche, and generalise the specific mistakes. Instead of writing, "Dave messed up the client pitch for the Acme account on Tuesday because he was late and forgot the deck," write, "An employee arrived late to an important external presentation and was unprepared with the presentation materials."

If you are using enterprise-grade models where you have explicitly opted out of data sharing, you have more leeway, but maintaining anonymisation is still a excellent mental discipline. It forces you to focus on the behavioural patterns rather than the personal frustrations. If you are unsure about your platform's privacy settings, refer to the official documentation on data usage, such as OpenAI Support or Claude Support, to ensure your prompts aren't being used for training.

2. Stop Asking for 'Polite'—Ask for 'Candid and Kind'

When we ask an AI to make feedback "gentle," it interprets this as a license to bury the lead in a mountain of corporate fluff. It will construct a "compliment sandwich" so thick with platitudes that the actual critique becomes invisible. This isn't kind; it's confusing.

To get better results from platforms like /platforms/claude or OpenAI, you need to change your prompting vocabulary. Avoid asking for "professional" or "softened" language. Instead, prompt for candour, clarity, and constructive action.

Try a system prompt like this:

> "I need to deliver constructive feedback to a team member. I will provide the raw, unedited points of concern. Your job is to help me structure this into a written message that is direct, fair, and focuses entirely on behaviours and outcomes, not personal attributes. Do not use corporate jargon, cliches, or overly formal language. The tone should be conversational, warm, and focused on helping them grow."

By setting these boundaries in your /prompts library, you bypass the default robot-speak and get drafts that actually sound like they came from a supportive human leader.

3. The 'Raw Dump' Method

One of the best ways to use an AI ethically here is as an emotional translator. When we are frustrated, our initial thoughts about a colleague's performance are often too sharp, emotional, or unfair.

Instead of trying to write a polished draft yourself, write down your raw, unfiltered thoughts. No filtering, no self-censoring—just get the frustration out on the page (again, keeping it anonymous). Then, ask the AI to help you identify the underlying objective issue.

For example, if your raw dump is: "Sarah is completely checked out. She hasn't contributed a single good idea in the last three sprint planning sessions and just sits there looking bored. It’s dragging the whole team down."

Ask the LLM: "Based on this raw feedback, what are the objective, observable performance issues, and what are the emotional assumptions I am making?"

An LLM is brilliant at pointing out that "looking bored" is an assumption, whereas "not actively contributing ideas during sprint planning sessions" is an observable behaviour. By using the AI to separate your emotional reaction from the facts, you can draft feedback that is fair, actionable, and free of personal attacks.

4. Knowing When to Shut the Tab

AI is a drafting tool, not a decision-maker. There are moments when using an LLM to communicate is a moral cop-out.

If you are delivering feedback that could lead to disciplinary action, a PIP, or termination, you cannot rely on AI-generated templates. These situations require absolute personal ownership. Every word must come directly from your own brain, matched to what you know of the person and the context.

Similarly, if the feedback is highly sensitive, skip the written message entirely. Use the AI to help you prepare your talking points for a face-to-face conversation, but do not send a cold, synthesised email. What makes a team relationship tick is the presence of real, human empathy—something that cannot be simulated, no matter how clever your prompt engineering is.

Use LLMs to clarify your thoughts, to strip away your biases, and to find the courage to be direct. But when it comes to the actual human-to-human connection, close the browser window and do the talking yourself.