A Hiring Bias Audit You Can Actually Run on Your Own Shortlist

Most conversations about bias in hiring stay abstract, which is why most companies conclude that they do not have a problem. An audit replaces the conversation with numbers. It takes a few hours and you can run it on data you already hold.

This matters more now than it did three years ago. Models trained on historical hiring data reproduce the patterns in that data, which means an automated screen can institutionalise past preferences and filter out capable candidates who lack conventional markers. But the point of the audit is broader than AI. Whether the filtering is done by a model or by a hiring manager, the audit tells you the same thing: who is dropping out of your process, and where.

What you are actually measuring

Not intent. You cannot audit intent and it does not matter. You are measuring pass-through rates by stage, split by attribute.

For each group, at each stage: what proportion moved forward?

If applicants from group A pass the first screen at 30% and applicants from group B pass at 12%, something in that screen is doing work you did not design it to do. Perhaps it is legitimate. Perhaps group B genuinely applies with less relevant experience. The audit does not tell you which. It tells you where to look.

Choosing your attributes

In the Nigerian context, the attributes worth checking are not identical to those used in American or European audits. The ones that carry real signal here:

• Gender. Straightforward to record and the most commonly consequential.
• University attended. Split into a small number of buckets: the handful of institutions your team habitually favours, other federal universities, state universities, private universities, polytechnics, and foreign institutions. This is frequently the largest source of unexamined filtering in Nigerian hiring.
• Age band. Nigerian job adverts still routinely carry age caps. If yours do, you have a written policy of filtering by age and you should know what it costs you.
• Region or state of origin, only if you already collect it. Do not start collecting it for this purpose.
• Employment status at application. Currently employed versus unemployed. Many processes quietly penalise gaps.
• Referral versus direct application. Referral pipelines reproduce the existing composition of your team. If referrals pass at three times the rate of direct applicants, your team will look increasingly like itself.

Do not audit on attributes you are not lawfully and appropriately collecting. Under the Nigeria Data Protection Act 2023 and the NDPC's implementation directive, collecting sensitive personal data needs a lawful basis and a stated purpose. Where you do not hold an attribute, leave it out rather than inferring it from names.

Running the audit

Step 1. Pick one role you have hired for at least three times, or three similar roles. Small samples produce noise. You need at least fifty applicants in total for the numbers to mean anything, and preferably two hundred.

Step 2. Lay out your stages. Typically: applied, passed screen, interviewed, final round, offered, hired.

Step 3. Build a simple table. Rows are your attribute groups. Columns are your stages. Cells are counts.

Step 4. Convert to pass-through rates. For each group, at each stage: how many moved to the next stage as a percentage of how many entered that stage. Stage-by-stage, not cumulative. Cumulative numbers hide where the drop happens.

Step 5. Look for the biggest gap, not every gap. There will be variation everywhere. You are looking for one or two stages where the difference is large and consistent across roles.

Step 6. Read the actual applications at that stage. This is the step people skip and it is the one that produces the answer. Take twenty applications from the group that passed and twenty from the group that did not. Read them side by side. The reason will usually be visible.

What the reasons usually turn out to be

In Nigerian hiring processes specifically, the recurring findings are:

Requirements that filter for access rather than ability. "Must have experience with [expensive enterprise tool]." The candidates who have it are those whose previous employer could afford it, which is a proxy for company size and sector rather than for skill.

Adverts that describe an ideal rather than a requirement. Long lists of desirables suppress applications from women and from candidates without insider knowledge of how job adverts work, both of whom are more likely to apply only when they meet every stated criterion.

Screening on school as a shortcut for screening on quality. Frequently unstated, frequently real, and easy to detect once you split the data.

Interview panels of one. A single interviewer's preferences become the company's preferences.

Unpaid multi-hour exercises. These filter for candidates who are not currently working long hours, which correlates with class more than with capability.

Auditing an automated screen specifically

If part of your process is automated, add two checks.

Check the inputs. List every field the system uses to score or rank. If school, address, or anything derived from a name is in there, remove it or verify it is not driving the outcome.

Run the shadow test. Take one role. Have the system produce its shortlist and have a human produce theirs independently, without seeing the other. Compare. If the system's shortlist has materially different composition on any attribute, you have found a problem before it cost you anything.

Ask any vendor directly whether they have tested their model on Nigerian applicant data specifically. A model tuned on American or European CVs will misread Nigerian career histories, which typically include more lateral moves, more contract work, and more informal-sector experience.

What to change first

Fix requirements before you fix anything else. It is the cheapest change and it has the largest effect.

• Cut your must-haves to three to five items you can answer yes or no to
• Move everything else to a clearly labelled desirable list, or delete it
• Remove age caps unless there is a genuine legal or operational requirement
• Remove the degree requirement for any role where you cannot articulate what the degree provides
• Replace tool-specific requirements with the underlying capability

Then fix the interview: same questions for every candidate, at least two interviewers, scored in writing during the conversation.

How often to repeat this

Annually is enough for most companies. Repeat sooner if you change your process, change your ATS, or turn on automated screening.

The value is not in any single audit. It is in having a number you look at again, because the first audit tells you where you are and only the second tells you whether anything you did worked.

Related reading

• 88% of companies now screen with AI. What it means for Nigerian employers
• Your applicants are using AI to write their CVs

CTA: BetternshipHR records stage-by-stage pipeline data as standard, so an audit like this is a report rather than a project. See how it works.

FAQ

Is a bias audit legally required in Nigeria? No. There is no Nigerian equivalent of the audit mandates emerging in some other jurisdictions. It is a commercial and ethical decision, not a compliance one.

What sample size do I need? At least fifty applicants across the roles you are auditing, and preferably two hundred. Below that, differences between groups are likely to be noise.

Should I collect attributes I do not currently record? Only where you have a lawful basis and a clear stated purpose under the NDPA, and only where the attribute is one you can act on. Do not infer attributes from candidate names.