AI in Performance Management: What It Helps With (and What It Should Not Decide)

By Johannes Sundlo, AI & Future of Work Advisor. I help leaders and HR teams turn AI into adoption that sticks, through keynotes, workshops and change programs.

Most of the work in performance management is not the decision. It is the preparation: pulling together scattered feedback, finding the right words for a hard message, and noticing the patterns that sit just below the surface. This is where AI performance management earns its place. Used well, it gives managers more time to think and a calmer way to prepare, while the judgment stays firmly with people.

The promise here is narrow on purpose. AI is good at drafting, summarizing and surfacing. It is not good at deciding who gets promoted or what someone should be paid. Keeping that line clear is what separates a useful tool from a quiet liability.

Where AI genuinely helps

The strongest use cases all share a trait: a human still owns the outcome. AI handles the heavy lifting on the inputs.

Drafting and de-biasing feedback

Managers often know what they want to say but struggle to say it cleanly. AI can turn rough notes into clear, specific feedback, then flag language that leans vague or loaded. Phrases like “not a culture fit” or “lacks executive presence” get caught and questioned, which nudges the manager toward observable behavior instead of impressions.

A simple prompt can do a lot of this work:

Here are my rough notes on an employee’s last quarter. Rewrite them as specific, behavior-based feedback. Flag any sentence that is vague, judgmental, or could carry bias, and suggest a more concrete version. Do not invent achievements I did not mention.

The point is not to outsource the message. It is to sharpen it before it reaches a person.

Summarizing 360 input

A round of 360 feedback can produce dozens of comments across many reviewers. Reading them fairly is hard, and tired managers skim. AI can group the comments into themes, separate signal from one-off remarks, and show where reviewers agree or contradict each other. The manager reads a structured summary, then goes back to the raw comments for anything that matters. That is faster and, in practice, fairer than scanning a wall of text at the end of a long day.

Prepping for tough conversations

Difficult conversations go badly when managers walk in unprepared. AI can act as a rehearsal partner: draft an opening, anticipate how the employee might respond, and suggest how to keep the conversation factual rather than personal. It can also help a manager check their own framing before they speak. None of this scripts the meeting. It lowers the odds of a clumsy start.

Spotting patterns

Across a team or a year, useful signals hide in plain sight. One person’s feedback quietly drifting downward. Goals that keep slipping for structural reasons rather than individual ones. Review language that reads differently for different groups. AI can surface these patterns for a human to investigate. The pattern is a question, not a verdict.

If you want more grounding on these working patterns, our broader notes on AI for HR and our collection of ChatGPT prompts for HR cover the practical mechanics in more depth.

What AI should not decide on its own

There is a short list of decisions that AI should never make alone, no matter how confident the output sounds.

• Ratings. A model can summarize evidence, but the rating is a judgment that carries consequences. It belongs to a manager who can be held accountable for it.
• Pay. Compensation decisions touch fairness, equity and law. They need human deliberation, context the model cannot see, and a clear record of who decided and why.
• Promotion calls. Readiness for a bigger role depends on potential, trust and team context. These are exactly the things AI estimates poorly and humans must own.

There is also a practical reason beyond principle. In several regions, automated decisions that significantly affect someone’s work or pay carry legal obligations around transparency and human review. Treating AI as a decision-maker for ratings, pay or promotions invites risk that a drafting tool simply does not.

Keeping a human in the loop

“Human in the loop” is easy to say and easy to fake. A real version has a few properties:

• A named person reviews every AI-assisted output before it reaches an employee. Review is the default, not the exception.
• The reviewer can see what the AI was given, so they can judge whether the summary is fair to the source.
• Managers are told plainly that AI can be wrong, can flatten nuance, and can carry the biases of its training. Skepticism is part of the job, not a sign of resistance.
• Sensitive personal data is handled with care, and you know where employee information goes when it enters a tool.

The goal is to make the human review meaningful rather than a rubber stamp. If a manager approves AI output without reading it, you have automated the decision in everything but name.

A practical way to introduce it

Most failed rollouts start too big. A calmer path looks like this.

• Start with one low-stakes use case. Drafting feedback or summarizing 360 input are good first steps because a human still writes the final version. Leave ratings, pay and promotions untouched.
• Pilot with a few willing managers. Pick people who will give honest feedback, not just enthusiasts. Run it for a review cycle and ask what helped and what felt off.
• Write down what AI is and is not used for. A one-page note is enough. Clarity here prevents quiet scope creep into decisions it should not touch.
• Be open with employees. Tell people that AI helps managers prepare feedback and that a human writes and owns every assessment. Transparency builds trust; discovery later destroys it.
• Review the outputs, not just the tool. After the pilot, read a sample of the actual feedback. Did it get clearer? Fairer? Or just longer? Adjust based on what you see.

Expand only once the first use case is genuinely working and trusted. The pace should be set by quality, not by the calendar.

The honest summary

AI performance management works best as a thinking aid, not a verdict machine. It drafts, summarizes, surfaces and rehearses. People decide. If you keep that division clear, you get better-prepared managers, fairer feedback and more time for the conversations that matter. If you blur it, you get a confident tool making decisions no one wants to defend later. The line is simple, and it is worth holding.