Short answer
AI can make performance-review preparation less administrative by organising approved evidence, identifying missing context and drafting questions. It should not decide what a person achieved, explain why performance changed, or determine a consequential outcome. A useful workflow makes each claim traceable, gives the employee a chance to correct context and leaves the final judgment with accountable people.
Start with the decision, not the draft
An AI-written paragraph can sound more certain than the records behind it. Begin by writing the review decision in plain language: is this a development conversation, a periodic assessment under an existing policy, or preparation for a separate decision? The answer determines what evidence is relevant and who needs to review it.
The CIPD describes performance reviews as one part of wider performance management, alongside regular discussion, objectives and feedback. Its current guidance also notes a shift toward regular conversations while retaining a role for structured reviews. CIPD's performance reviews factsheet is useful context for that distinction.
A five-part preparation worksheet
Use a shared working document, not an unreviewed generated report.
| Field | What to record | Human check |
|---|---|---|
| Review purpose | Development, periodic review or another defined process | Does the employee understand the purpose? |
| Criteria | Published role expectations or objectives | Are they still applicable to this period? |
| Evidence log | Dated work examples, deliverables and agreed feedback | Can each item be located and challenged? |
| Missing context | Changed scope, dependencies or inaccessible work | Who can clarify it? |
| Conversation plan | Questions, possible support and a review date | What must be discussed before any conclusion? |
AI may help sort a long evidence log into the stated criteria, identify repeated wording or suggest questions. It must label a statement as a draft when the record does not establish it. “Delivery was late because priorities were unclear” is an interpretation, for example, unless the conversation and records support it.
Claim-checking before the meeting
Ask the manager to check each proposed sentence against this short test:
- Source: What approved record, dated example or accountable observer supports this statement?
- Scope: Does the example describe the full period and role, or one moment only?
- Alternative: What work condition, dependency or change in scope could also explain it?
- Employee context: What will the employee be invited to add or correct?
- Decision boundary: Is this a discussion prompt, or is someone treating it as a conclusion?
This protects against both stale manager memory and confident language generated from partial inputs. It also makes a disagreement useful. If the employee says a missed milestone followed a changed dependency, record that context and decide what further checking is needed instead of asking AI to settle the account.
Fictional example: a draft is not an assessment
Fictional example: a manager uploads three project notes and asks for a review summary. The draft says that Alex “struggled with stakeholder management.” In the evidence log, two notes instead show that the project sponsor changed twice and Alex documented unanswered decisions. The manager changes the draft to a question: “What support would have made the decision path clearer, and what would you handle differently next time?” The conversation can then consider both Alex’s contribution and the operating conditions.
That is a better use of assistance: it prepares a discussion without converting incomplete records into a personal verdict.
Where AI needs stricter controls
Employment-related tools need more than a useful prompt. NIST's AI Risk Management Framework is a voluntary framework for incorporating trustworthiness considerations in design, use and evaluation. In the US, the EEOC has also warned that AI and other software used in employment decisions can create disability-discrimination risks, including when a tool screens out workers or makes disability-related inquiries. Read the EEOC and DOJ notice.
For a performance-review workflow, practical controls include:
- restrict inputs to material the organisation is permitted to use for that purpose;
- keep the source and date beside every material claim;
- let the employee see and respond to the evidence used in their discussion;
- test draft outputs for unsupported assumptions and inconsistent treatment of comparable cases;
- keep a named human owner for any final review, development, reward or employment decision.
Those are operating controls, not a claim that a tool is compliant or fair in every setting. Local policy and specialist advice may be needed where reviews affect pay, progression or employment.
Where Lontra fits
Lontra can support a focused employee conversation and a manager brief before a review. It does not assign a performance rating or make a people decision. Use its context alongside the organisation’s own approved records, then let the manager and employee discuss what is accurate and what happens next. Explore the product.
Frequently asked questions
Can AI write a performance review?
It can help organise approved material or draft neutral wording, but a manager and employee must check the evidence, discuss it and make the consequential judgment.
What should never be delegated to AI in a review?
Do not delegate the assessment of a person, a pay, promotion or employment decision, or the resolution of disputed facts to an AI output.