Run an AI bias report

What this achieves

A bias report examines the decisions AI has influenced over a period and reports whether outcomes were distributed evenly across groups of people. It exists because AI here touches recommendations, readiness, and promotion, and those reach individuals.

Required role: Administrator.

Steps

Go to Administration → AI Governance → AI Bias.
Choose the monitoring area.
Select Generate report.
Set the Monitoring type, the Period start, and the Period end.
Select Generate.
Read the fairness score, then the flags, then the breakdowns.
Select Mark as reviewed and record your Notes.

The five monitoring areas

Option	Description
Recommendations	Suggestions the AI made to people.
Readiness scores	Readiness scoring against target roles.
Mobility matching	Matching people to internal opportunities.
Feedback cycles	Feedback cycle participation and outcomes.
Promotion readiness	Promotion readiness assessment.

Each area is analysed separately and holds its own report. A clean result in one says nothing about the others, and the areas that matter most — promotion and readiness — are the ones people are least likely to check.

Choose the period deliberately

The report covers the dates you give it, and the summary states how many decisions fell inside them.

Too short a period gives you a handful of decisions, where an uneven split is chance rather than a pattern. Too long a period hides a recent change inside an average.

Example: a quarterly report across HC Corp’s promotion-readiness decisions gives enough volume to mean something. A one-week window covering four decisions does not, whatever percentage it produces.

Reading the report

Option	Description
Generated on	When the report was produced.
Decisions	How many decisions it examined.
Fairness	An overall fairness score for the period.

Read the decision count before the fairness score. A fairness percentage over a small number of decisions is arithmetic, not evidence.

Bias flags

Flags are the specific findings, each with a severity of High, Medium, or Low, the deviation it found, and the expected against the actual figure.

The expected-against-actual pair is the part to act on. It says what the distribution would have looked like if outcomes had followed the population, and what it actually looked like — which is a statement you can investigate, unlike a score.

Where no flags are raised, the report says so for that period rather than leaving the section blank.

Breakdowns

Outcomes are broken down by gender, department, grade, tenure, and age group.

Read these carefully, because a skew is not automatically bias. A department breakdown showing Engineering receiving more learning recommendations may be reporting that Engineering has more skill requirements, not that anyone is being treated unfairly. The flag tells you where to look; the explanation is yours to find.

Small groups identify people

Breakdowns over small groups stop being statistics. HC Corp’s People department has 6 people and Finance has 9, so a departmental breakdown showing “one” is describing a person.

Treat small-group output as personal information, keep it with the people who would legitimately handle it as such, and do not put it in a circulated document.

Trend

The fairness score over time is more informative than any single report. One report tells you the position; the trend tells you whether something you changed worked.

Mark it as reviewed

Mark as reviewed records who read the report, when, and what they concluded, in the Notes field.

Write what you actually assessed and what you plan to do — for example, that a department skew was investigated and traced to differing requirements, or that a remediation was scheduled. Those notes are the record that the finding was examined, and an unreviewed report is indistinguishable from an unread one.

Review on a schedule, not after a complaint

Set a cadence — quarterly is the usual choice — and hold it. The value of bias monitoring is that it finds a pattern before somebody else does, and a report generated in response to a complaint has already lost that advantage.

What happens next

A flag is a prompt to investigate, never a conclusion about the AI or about anyone using it. Where a flag turns out to reflect something real, an action plan generated from that finding tracks the remediation and keeps it attached to the report that raised it.