Short answer
AI can help a reviewer organise approved multi-rater comments, group repeated observations and prepare questions for a development conversation. It should not decide which account is true, infer a person’s capability from a pattern of comments or convert contested feedback into an employment outcome. The useful output is a reviewable draft with sources, gaps and contradictions still visible.
AI does a different job from 360 feedback software
The buyer guide to 360 feedback software covers rater selection, invitations, access and follow-through. This page concerns a narrower moment: what happens after feedback has been collected and a reviewer is considering an AI-generated synthesis.
The distinction matters. A tool can automate collection without using generative AI. Conversely, an AI summary can be applied to comments collected elsewhere. Neither capability removes the need for a clear development purpose, a human reviewer and an employee conversation.
A reviewer-controlled synthesis sheet
Before uploading or processing feedback, define the permitted input and the review task.
| Field | Reviewer question |
|---|---|
| Purpose | Is this preparation for development, or is someone trying to use it for a different decision? |
| Material | Which comments, dates and relationship groups are approved for this use? |
| Prompt | Does the instruction ask for themes and counterexamples, rather than a verdict on the person? |
| Traceability | Can the reviewer find the source and context for each material theme? |
| Contradictions | Which accounts differ, and could they describe different work situations? |
| Conversation | What will the employee be able to add, question or correct? |
Use a prompt that asks the model to separate observation from interpretation. For example: “List recurring observations with their relationship group and date. List comments that do not fit the pattern. List questions a reviewer should discuss. Do not assign a rating or recommend a personnel decision.”
Keep traceable source references inside the authorised reviewer workspace. Decide separately what can be shared with the participant: relationship groups, dates and distinctive details can identify a rater even when names are removed. Check the agreed confidentiality rules and small-group handling before sharing a summary; do not promise anonymity from redaction alone.
Fictional example: preserve the disagreement
Fictional example: six raters comment on a project lead. Three peers describe clear written handovers. Two direct reports say priorities changed without enough explanation, while the sponsor praises the lead’s pace during a launch. An AI draft might call the lead “inconsistent in communication.” That phrase is too broad to act on.
The human reviewer keeps the distinction: handovers may be clear after decisions are made, while priority changes may not have been explained to the team at the time. The development conversation can ask about those two situations separately, with a concrete next practice and date to revisit it. The disagreement was evidence to explore, not noise to remove.
Controls to ask for and operate
NIST's AI Risk Management Framework provides a voluntary way to consider risks in AI design, use and evaluation. The CIPD's current AI and technology factsheet also frames technology selection and use as a people-profession responsibility. Put that into practice with a small operating checklist:
- Access: restrict processing to people authorised for the feedback purpose and to approved material.
- Evidence: retain the original comments and their relevant context for the reviewer; do not rely solely on a generated paragraph.
- Counterevidence: require the draft to list comments that conflict with its themes or lack enough context.
- Review: name a human reviewer who can amend or reject the draft before it is discussed.
- Conversation: let the subject respond to the themes and record agreed development actions separately from the AI output.
- Use boundary: do not use the synthesis as an automatic ranking, a determination of capability or a final people decision.
This is not a blanket claim about legal compliance, security or fairness. A buyer must check its own policy, contract terms and applicable requirements, especially when feedback is sensitive or contributes to a consequential process.
Questions for a vendor demonstration
Ask the provider to demonstrate the following with synthetic comments:
- Can the reviewer inspect the material behind a theme and identify its relationship group and period?
- Can the reviewer see disagreement or low-context comments, rather than only the most frequent wording?
- Can the reviewer correct, remove or add context before sharing a draft?
- What user roles, retention settings and processing terms apply to this particular deployment?
- Does the product describe the summary as decision support, or does it make a claim about an individual that the reviewer cannot verify?
These questions produce a more useful test than asking whether the model is “accurate” in the abstract.
Where Lontra fits
Lontra can provide a focused employee conversation and a manager brief to add work context to a development discussion. It is not a multi-rater administration system, an automated evaluator or a tool that resolves conflicting feedback. Managers receive briefs rather than raw conversations, and accountable people remain responsible for the discussion and any decision. Explore the product.
Frequently asked questions
Can AI summarise 360 feedback?
It can organise approved comments into a draft, but a trained reviewer should preserve context, check contradictions and decide what can fairly be discussed.
Should AI resolve conflicting feedback?
No. A contradiction can reflect different work relationships, time periods or expectations. It should be investigated by people, not averaged into a conclusion.