Use AI to help express developmental feedback only after the manager has recorded what happened, what they can substantiate and what they want to discuss. Review the rewrite for changes in meaning, then personally own the message and the follow-up. A fluent paragraph cannot supply observations the manager never made or replace the employee’s opportunity to explain the situation.
For HR and learning teams, the practical decision is how to support managers with writing while keeping evaluation and conversation accountable. This guide proposes a bounded process for ordinary developmental feedback. It does not advise on an individual employee, determine ratings or cover disciplinary, compensation or hiring decisions. Those require the organization’s established processes and appropriate expertise.
The examples below are fictional. We have not tested a feedback product, run a manager-training experiment or measured employee outcomes for this article.
Research on source labels is not a test of better writing
In an August 2026 Frontiers in Psychology study, Marót and colleagues randomly assigned employed participants recruited through convenience sampling in Hungary to four descriptions of feedback production: human-written, AI-generated, AI-generated with human refinement, or human-written with AI refinement. Everyone read the same feedback excerpt in a hypothetical workplace scenario.
The human-written label elicited more favorable reactions than the fully AI-generated label across several measures. Hybrid conditions showed a more nuanced pattern; the two hybrid conditions did not significantly differ. This does not establish that the workflows are equivalent or that one drafting sequence improves real performance.
The authors emphasize immediate anticipated reactions, many single-item measures and limited generalizability beyond the sample. This was not a comparison of actual managers producing different-quality feedback. It also does not justify hiding AI involvement. Our recommendation is to make the process truthful and accountable, rather than optimize the label attached to unchanged text.
Prepare the substance before opening the assistant
A useful starting note identifies the purpose of the discussion and the evidence available. What event did the manager observe? What expectation had been communicated? What consequence can they actually describe? What is still uncertain? Keep secondhand reports distinguishable from direct observations instead of merging them into a confident narrative.
The Center for Creative Leadership’s Situation-Behavior-Impact guidance recommends describing the situation, observable behavior and impact, then asking about intent rather than assuming it. That is practitioner guidance for feedback conversations, not evidence that an AI prompt can assess an employee accurately. We use it to inform the preparation step, not to claim a validated AI workflow.
Choose an approved tool and data-handling arrangement before using real personnel information. Supply only the context needed for the permitted task. Removing a name does not necessarily make an account anonymous if the event identifies the person. Do not paste private HR records into a convenient assistant simply because the task is “only editing.”
If the notes contain only “poor attitude” or “not leadership material,” pause the drafting task. Ask the manager to identify the observations and relevant expectations, or acknowledge that they lack sufficient information. Asking AI to add specificity would create the appearance of evidence without providing it.
Keep observation, interpretation and proposal distinct
Consider a fictional manager preparing a conversation about an internal project handover. The manager’s invented notes say:
- In the Tuesday handover, the checklist did not name the person responsible for the remaining verification step.
- The receiving team asked the manager who should complete that step before continuing.
- The manager has not yet asked why the name was missing.
- The manager wants to discuss how responsibility should be recorded in future handovers.
These notes can support a specific discussion. They do not establish a pattern across months, a personality trait or a reason for the omission. The following table shows how a rewrite could change the meaning even while sounding professional. The suggested checks are our editorial proposal.
| Possible rewrite | Change introduced | Manager’s check |
|---|---|---|
| “You repeatedly leave handovers incomplete.” | One event becomes a recurring pattern | Require additional supported examples or remove the generalization. |
| “Your lack of commitment caused the delay.” | An observation becomes a motive and causal claim | Do not infer commitment or unrecorded consequences. |
| “The receiving team asked who owned the verification step.” | Retains the supplied event | Confirm that the description matches the record. |
| “We have agreed you will own every verification step.” | A discussion goal becomes an agreement | Keep it a proposal until the conversation establishes an agreement. |
The last change is particularly easy to overlook during a tone review. It does not merely make the wording stronger: it invents consent and a continuing responsibility. The reviewer should examine verbs such as “agreed,” “caused” and “repeatedly,” not just whether the message sounds kind.
Positive feedback deserves the same care. Turning one helpful contribution into “you always support everyone” can add an unsupported pattern. Praise should remain specific enough that the recipient understands what contribution was recognized.
Give AI a writing role with visible limits
Provide the approved notes and ask for a shorter or clearer version without new claims. A bounded request could say: “Rewrite these notes for a developmental conversation. Preserve uncertainty. Do not add events, frequency, motives, ratings, impacts or commitments. List any wording that requires information not supplied. Keep proposed next steps distinct from agreed ones.”
This is a suggested instruction, not a tested prompt or a guarantee against errors. The manager must still compare the result with the notes. If the assistant flags missing context, return to the evidence or leave the gap visible; do not reward it for producing a complete-looking paragraph regardless.
A parallel comes from another writing domain. In her September 2026 account of NN/g’s editorial process, Raluca Budiu describes AI assistance with clarity, formatting and critique while human editors retain responsibility. She also notes that a rewrite can lose nuance. This is an editorial practice account, not a study of HR outcomes; the useful analogy is that assistance with expression does not transfer responsibility for the content.
Retain the original note and final version in the approved process, with access and retention appropriate to personnel information. A large organization may use an established HR system; a small team should still make the final record recoverable without scattering sensitive drafts across personal chats and documents.
Review meaning before reviewing tone
Read the output once against the source notes, sentence by sentence. Can the manager explain the basis of every factual statement? Does the scope still match the observed event? Are uncertainty and secondhand information preserved? Did the assistant insert a diagnosis, a trait or an inference about motivation?
Then inspect the action language. A manager’s suggestion is not an employee’s commitment. An intention to discuss support is not a promise that resources have been approved. Remove or qualify statements that exceed what the manager can actually offer.
Only after those checks, review clarity and tone. Shortening is useful when it preserves the meaning; softened language is unhelpful when it makes the concern impossible to understand. Avoid mechanically demanding the same emotional style in every case. The manager should be able to say the final message naturally and answer reasonable questions about it.
Our evidence-based review guide explains why checking support differs from accepting polished prose. In this setting, review also includes authority: who observed the event, who made the judgment and who can agree to the next step?
Explain the process honestly and allow correction
Describe AI’s role in terms that match what happened and follow the organization’s disclosure policy. If the manager supplied the observations and used AI to shorten the wording, that is different from asking a model to generate an evaluation from a collection of records. Do not call the latter “just spell-checking” or claim an expert review that did not occur.
An illustrative explanation for the fictional editing-only workflow is: “I wrote the observations and used an approved AI tool to help make the wording clearer. I checked the final message and am responsible for it.” This statement is appropriate only when every part is true. A disclosure statement cannot repair unsupported substance.
Deliver the feedback through a process that permits dialogue. In the handover example, the manager can describe the missing responsibility and ask what happened. The employee may identify an earlier assignment or another piece of context the manager missed. Correct the record where warranted rather than treating the AI-polished draft as settled evidence.
Evaluate the support process, not just the polished paragraph
For manager training, start with fictional scenarios and compare draft rewrites with their source notes. Ask reviewers to identify invented frequency, motives, impacts and agreements. Assess whether the final wording remains specific and whether the manager can explain and revise it. These are proposed training checks, not an exercise we have run.
If the organization pilots AI support with approved real work, account for preparation, review, corrections and follow-up as well as drafting time. Ask whether recipients could understand the concern and contribute missing context; do not treat message length or positive reactions alone as evidence of better development.
Keep consequential or contested matters with the appropriate human process. For ordinary developmental feedback, the useful boundary is concrete: the manager supplies and verifies the substance, AI may assist expression, and the conversation remains open to evidence the draft did not contain.