AI in Performance Management: Feedback, Goals and Review Drafts

By Kevin Kim · Updated Oct 11, 2026

AI in performance management means using software, usually a large language model, to help with the paperwork and prompts around performance: drafting and sharpening goals, collecting and summarizing feedback during the year, and preparing first drafts of review write-ups that a manager then edits, owns and delivers. It speeds up the writing and the pattern-spotting. It doesn't, and shouldn't, decide who is rated well, paid more or promoted.

This guide is for HR leaders, people managers and anyone being asked to approve an "AI-assisted review" feature. It covers what AI does well at each stage of the cycle, where it quietly goes wrong, and how to set it up so that managers stay accountable for every judgment about a person. It sits inside the AI in HR section. If you're asking a different question, how to evaluate people whose own work is heavily AI-assisted, that's covered in performance management when AI does the work.

Key Facts

• A meta-analysis of feedback interventions in Psychological Bulletin pooled 607 effect sizes from 23,663 observations. Feedback raised performance on average (d = .41), but over one third of the effects were negative (Kluger and DeNisi, 1996, APA). More feedback isn't automatically better feedback.
• Goal-setting research finds that specific, difficult goals lead to higher performance than urging people to do their best, with effect sizes in meta-analyses ranging from .42 to .80, and that goals plus summary feedback beat goals alone (Locke and Latham, 2002, American Psychologist).
• Gallup found 61% of employees who get both feedback and recognition from their manager at least weekly are engaged, versus 38% of those who get weekly feedback but less recognition. Only 27% of employees said they'd like feedback from their manager at least weekly (Gallup, October 2024).
• Researchers who tested LLMs on 600 student essays found that swapping gendered cues changed the feedback more in the male-to-female direction than the reverse, with more controlling feedback under female cues (Du, Borchers and Cukurova, arXiv). That was education data, not workplace data, but the mechanism is the one HR should test for.
• EU AI Act Annex III lists, as high-risk, AI systems intended to make decisions on promotion or termination, to allocate tasks based on individual behavior or traits, or to monitor and evaluate the performance and behavior of workers (Regulation (EU) 2024/1689, EUR-Lex). The Commission says high-risk rules for areas including employment apply from 2 December 2027 (European Commission).

What AI Actually Does in the Performance Cycle

In ACE Framework terms, performance tools ingest text (goal documents, check-in notes, peer comments) and some structured data (ratings, dates, targets), analyze and summarize it, and generate drafts. The patterns are generative research (pulling together a year of notes) and workflow copilot (suggesting the next thing a manager should write or do). None of that requires a model to predict anything about a person, and the safest deployments keep it that way.

Here's how it maps to the cycle.

The honest shorthand: AI is good at the first 80 percent of the writing and poor at the last 20 percent that carries the actual judgment. Keep it on the first part.

Goal Setting: Where AI Helps Most

Goal-setting research is one of the sturdier findings in applied psychology. Specific, difficult goals outperform "do your best", and the combination of a goal and regular feedback on progress beats a goal alone. The trouble in most companies isn't theory, it's that goals are written once a year, in a hurry, in vague language.

That's where an assistant earns its keep. Give it a rough objective and it can propose a version with a measure, a time frame and a first milestone. A customer success manager who writes "improve retention" can get back "reduce churn in the mid-market segment from the current rate to a target by the end of Q3, with a monthly review of at-risk accounts". A software team lead can turn "improve code quality" into a goal about review turnaround time and escaped defects.

Three rules make this work.

1. The employee proposes, the tool sharpens. If AI writes the goal from scratch, the person hasn't committed to it, and commitment is what makes goals work.
2. Check difficulty yourself. A model can't know what's a stretch for this person in this territory. Managers should adjust the target after talking it through.
3. Link goals to the work, not to the activity. Ask the assistant to flag goals that measure effort (calls made, tickets closed) when the point is an outcome. This matters even more when AI inflates output, which is a separate problem from the one in this guide.

Continuous Feedback: Useful, but Easy to Get Wrong

Gallup's October 2024 analysis ties frequent feedback paired with recognition to higher engagement, and it also shows a gap: only 27% of employees want weekly feedback from their manager, and only one in four strongly agree they get valuable feedback from the people they work with. That should make anyone cautious about a tool whose pitch is "more feedback, more often".

The Kluger and DeNisi meta-analysis is the sobering counterweight. Feedback helped on average, but more than a third of the effects were negative. Apply that to AI-assisted feedback and the practical advice is clear.

• Use AI to make feedback specific and task-focused. "The onboarding email sequence you shipped cut time-to-first-use, and the second email is the one to rewrite" beats "great job, keep it up" or "you need to be more strategic".
• Use it to summarize, not to score. Pulling five peer comments into three recurring themes is a good job for a model. Producing a number for someone's "collaboration" is not.
• Use nudges, not volume. A reminder to a manager who hasn't had a one-to-one in three weeks adds value. A stream of auto-generated praise doesn't.
• Keep recognition human. Gallup's data points to the combination of feedback and recognition from the manager. A machine-drafted thank-you that reads like one rarely counts. Pair this with the signals covered in AI for employee engagement.

Review Drafting: Faster Paperwork, Real Risks

Review season is where managers spend the most hours, so it's where AI looks most tempting. A manager pastes in notes and goals, and the model returns a structured write-up in seconds. Generative AI for HR teams covers the drafting techniques; here's what's specific to reviews.

What it does well. It organizes scattered notes by goal, fixes grammar, shortens rambling paragraphs and offers alternatives to loaded phrases. For a manager who writes twelve reviews a year and dislikes writing, it removes a real barrier.

What it does badly.

• Bland, interchangeable text. Models produce plausible praise and plausible criticism. Two employees with very different years can end up with near-identical paragraphs.
• Invented specifics. A model asked to "flesh out" a thin note may add achievements or incidents that never happened. This is the hallucination problem (AI hallucination), and in a review it becomes a false statement in someone's file.
• Biased language. LLM output reflects patterns in its training data. In the education study cited above, counterfactual gender swaps shifted the feedback, with more autonomy-supportive language under male cues and more controlling language under female cues. A separate paper on LLM-written recommendation letters found significant gender bias in language style and word choice (Wan et al., arXiv). Performance reviews are the same genre of text.
• Anchoring. A polished draft is hard to argue with. Managers tend to accept it and tweak, so the model's framing quietly becomes the verdict.

A simple discipline helps: the manager writes the judgment first (the rating rationale in a few plain sentences), and the AI helps with the surrounding structure and wording, not the other way around.

Real Examples by Function

Sales. A sales manager keeps weekly bullet notes on each rep. At mid-year the assistant groups the notes by goal (pipeline coverage, win rate, forecast accuracy) and drafts a summary. The manager removes two claims the notes don't support, adds context about a territory change and writes the rating rationale herself.

Engineering. A team lead asks the assistant to rewrite vague goals such as "own reliability" into measurable ones, then reviews them with each engineer. Engineers push back on one target, and the final number is theirs.

Customer support. A support manager uses AI to summarize themes across quality-assurance comments and customer feedback for each agent. The summary becomes the starting point for a coaching conversation, not an automatic score.

HR business partner. Before calibration, the partner runs a check on draft reviews for language that differs by group: vague praise for some people and specific praise for others, personality comments in place of behavior. This surfaces patterns for managers to fix. It doesn't change any rating on its own.

Fairness and Manager Accountability

Two questions decide whether this goes well: is it fair, and who is responsible when it isn't?

Fairness. Treat any AI-assisted review process as something to test, not trust. Compare draft length, specificity and tone across gender, age, location and tenure. Look at who gets "needs to be more assertive" and who gets "great team player". Where you're legally allowed to analyze protected characteristics, check ratings and promotion outcomes by group before and after you introduce the tool.

Accountability. The manager signs the review. That must stay true in policy and in practice. Write it down:

• AI may draft and summarize, and the manager is responsible for every statement.
• The employee should know AI helped prepare the document, and what data went in.
• Ratings, pay and promotion decisions are made by people, with a human able to explain the reasons in plain language. This is the human-in-the-loop principle.
• An employee can ask for a correction, and a person, not a tool, handles it.

Legal exposure. In the EU, GDPR Article 22 gives people the right not to be subject to a decision based solely on automated processing, including profiling, that significantly affects them, and its safeguards include the right to human intervention and to contest the decision (GDPR, EUR-Lex). The AI Act's Annex III points the same direction for systems that evaluate workers. Rules elsewhere differ and keep moving, so involve counsel for each jurisdiction. The practical test: if the tool influences a decision about someone's pay, promotion or job, treat it as high-risk and keep a person in charge of it.

Predictive scores. Resist using review data to predict who will quit or who has "potential" without strong safeguards. And connect any rating-linked pay decisions to the guidance in AI in compensation planning.

How to Implement AI in Performance Management Step by Step

1. Pick one stage. Start with goal drafting or note summarization. They're lower stakes than rating or calibration.
2. Define the red lines. AI doesn't produce ratings, rankings, promotion recommendations or termination rationales. Write that down before any pilot.
3. Fix the inputs. A model summarizing thin or one-sided notes produces thin or one-sided summaries. Agree what evidence managers should log through the year. HR data foundations for AI covers the groundwork.
4. Choose a tool with the right controls. Look for an enterprise setup where employee data isn't used to train public models, access is limited by role, and every draft shows what it was based on.
5. Train managers. Cover how to prompt well, how to spot invented detail, and how to rewrite the judgment in their own words. One hour on this beats a feature walkthrough.
6. Pilot with a feedback loop. Run one function for one cycle. Ask managers how much time it saved and employees whether reviews felt more or less accurate and personal.
7. Test for bias. Compare drafts and final outcomes across groups. Fix prompts and templates when you find gaps.
8. Tell employees. Explain what AI is used for, what it isn't, and how to raise a concern.
9. Measure what matters. Track time spent per review, quality scores from employees, rating distribution shifts and the number of corrections. Time saved alone is the wrong success measure.
10. Review yearly. Models, regulations and your own workforce change. Reconfirm the red lines and the testing each cycle.

Common Pitfalls

• Letting the draft become the verdict. Make managers write the rating rationale before they see any AI text.
• Measuring volume of feedback. Frequency without value doesn't help. Only one in four employees strongly agree the feedback they get is valuable.
• Auto-generated praise. Employees can tell, and it cheapens real recognition.
• Unreviewed specifics. Require managers to check every achievement or incident in a draft against their notes.
• Using the tool to rank. Moving from summarizing to scoring turns a writing aid into a high-risk system.
• Skipping the bias test. If you don't compare outputs across groups, you won't know.
• No disclosure. Hidden AI use erodes trust the moment someone finds out.

Where to Go Next

• For the wider AI-in-HR picture, start with AI in HR.
• To see how AI-assisted work changes what you should evaluate, read performance management when AI does the work.
• For agent-style tools that assemble review material automatically, see the AI performance review agent.
• To see how goals and feedback can feed a skills inventory, read AI skills inference.