Performance Review Bias: An Evidence Audit With Counterexamples

Short answer

Performance review bias is a risk that evidence, standards or judgment are applied unevenly. Audit the decision rather than guessing a manager's intent: trace each material statement to a source, check coverage across the review period, look for counterexamples, compare similar cases and examine whether protected or less-visible groups receive different treatment. A structured process can reveal and correct problems, but it cannot certify a review as bias-free.

Audit the decision path, not a list of labels

Terms such as recency, halo, visibility, similarity and attribution bias can prompt useful questions. They are not diagnoses. A reviewer needs to show where the evidence, criterion or process changed the decision.

Use this audit map:

The result of an audit should be a corrected record, a process change or an unresolved question. “Bias training completed” is an activity, not proof that decisions are now consistent.

Build a review-audit sheet

Sample decisions with enough variation to test the process. Include different managers, roles, sites, work arrangements and proposed outcomes, subject to lawful and appropriate data use.

For each material criterion, record:

1. Decision: proposed rating or conclusion and any related action.
2. Standard: exact criterion communicated to the employee and when.
3. Role fit: why that criterion applies to this person's work and authority.
4. Supporting source: dated evidence and what it actually shows.
5. Coverage: which part of the review period and work the source represents.
6. Counterexample: relevant evidence inconsistent with the proposed conclusion.
7. Comparable case: how the same criterion was applied in a similar situation.
8. Context: changed scope, dependencies, tools, leave, workload or adjustments.
9. Employee response: factual corrections, disagreement and additional evidence.
10. Audit outcome: retain, narrow, revise, remove or investigate further.
11. Owner: person accountable for the decision and correction.

Use access controls appropriate to the material. An audit does not require every participant to see every employee's record.

Ask whether the standard was fair before testing evidence

A perfectly documented decision can still apply the wrong criterion. Check whether the standard was relevant to the role, within reasonable control, communicated in advance and supported by appropriate access and adjustments.

Acas guidance on performance reviews says objectives should be fair and reflect usual tasks and workload. The US Equal Employment Opportunity Commission advises employers to communicate performance standards, apply them consistently and include factual details in evaluations. Its guidance for employers conducting evaluations also warns against holding employees to higher standards or rating them negatively because of protected characteristics.

Those sources apply within their jurisdictions and contexts. They do not turn this worksheet into legal advice or establish that one audit method satisfies every obligation. Involve the organisation's employment and legal owners when a decision may affect pay, promotion, discipline or dismissal.

Test the evidence behind each sentence

Take one material sentence from a review and split it into parts:

Replace a broad label with a bounded statement the employee can understand and challenge. If the source shows one late handover, write that. Do not generalise to reliability across the year without further evidence.

Use counterexamples deliberately

Counterexamples are not courtesy notes placed after the conclusion. They test whether the conclusion survives relevant contrary evidence.

Ask reviewers to find:

• a different period under the same criterion;
• an example from a less-visible assignment;
• a similar case involving another employee;
• evidence from the employee or recipient of the work;
• a condition outside the person's control;
• an instance where a different style achieved the required result;
• a fact that would change the proposed decision if verified.

Record why a counterexample changes the decision or why it does not. “Not persuasive” is incomplete without a reason linked to the standard.

Fictional audit: two service-team reviews

This example is fictional. Two service coordinators, Jordan and Kim, are reviewed on the same criterion: hand over unresolved priority cases by the agreed cut-off, with owner, status and next action.

Jordan's draft review says “meets the standard” based on three recent handovers. Kim's says “inconsistent” because one high-profile handover lacked a next action. The audit sheet reveals:

• Jordan completed 10 of 12 sampled handovers; two lacked an owner.
• Kim completed 11 of 12; one lacked a next action.
• Jordan's examples came from the final month, while Kim's high-profile miss came from the first month.
• Both received the revised template halfway through the period.
• Kim's manager cited the miss in two separate criteria, while Jordan's omissions appeared in neither.

The audit does not prove why the managers treated them differently. It does show inconsistent evidence coverage and a possible duplicated penalty to examine. One incident can legitimately inform more than one criterion when each connection is justified; it should not silently receive extra weight merely because it is repeated. The reviewers return both drafts to the managers, apply the same sampling window and decide each criterion once. They document the revised rationales and check whether the template change altered what could reasonably be expected in the first half.

No algorithm decides either outcome. The audit corrects the comparison process and preserves any remaining disagreement.

Examine patterns without treating disparity as a diagnosis

Aggregate review outcomes can reveal where to investigate. Compare ratings, changes after calibration, development access and resulting pay or promotion decisions across relevant groups when lawful and appropriate. Keep denominators, small-group protections and job differences visible.

A difference does not by itself prove discrimination, and an overall average can hide inconsistent decisions. Move from the pattern back to the criteria, evidence and process. The EEOC's guidance on internal performance-evaluation complaints recommends examining the challenged statements and actions and, where appropriate, reviewing other evaluations for evidence of discrimination. The UK Employment Statutory Code of Practice discusses transparent, objective and justifiable criteria, consistent procedures and reasonable adjustments.

Define the challenge route in advance. Preserve the original, correction, decision owner and date rather than overwriting the record without an audit trail.

Audit any AI-assisted step too

AI does not remove reviewer bias and can introduce new errors through missing context, training data, prompt design or confident summaries. If a system organises evidence or proposes a theme:

• document the permitted input and intended use;
• keep the source available to the reviewer;
• test omissions, misattribution and unsupported inference;
• compare performance under conditions similar to actual use;
• involve affected users in testing where appropriate;
• record human corrections and stop conditions;
• prohibit the output from deciding an individual's rating, pay or promotion.

The NIST AI Risk Management Framework Core calls for documented scope, knowledge limits, human oversight and evaluation in deployment-relevant conditions. Those practices strengthen governance. They do not validate a performance-review model or guarantee a fair outcome.

Turn findings into process changes

Classify each audit finding so the remedy matches the problem:

For building the underlying criteria, use the objective performance review worksheet. For comparing application across managers, use the performance calibration protocol.

Where Lontra can support preparation

Lontra can help prepare an employee conversation and a manager brief around a specific work question. Managers do not receive raw employee responses. HR and managers still need to review the actual performance evidence, find counterexamples, apply the organisation's standards and own every individual decision. Lontra does not score performance, rank employees or validate a bias audit.

The Lontra product overview shows the current conversation and brief workflow. Use conversation context to ask a better question, then keep the formal performance, pay and promotion records in the accountable processes designed for them.