An accessibility audit for AI-assisted hiring

An accessibility audit should establish whether applicants can participate in the hiring process and be evaluated on the skills relevant to the job. For an AI-assisted step, the agency also needs to understand what the tool measures, where the result goes, and how a person can request a workable alternative when the process creates a barrier.

Begin with a real hiring workflow, not a general assurance about the software. Follow the applicant from the invitation through the assessment, review, and next decision. Include the instructions, time limits, communication channel, and the way the result is interpreted. A barrier can appear between tools even when each supplier describes its own interface as accessible.

Understand the responsibility behind the audit

U.S. Department of Justice guidance explains that employers must consider disability discrimination when using hiring technologies, including tools supplied by another company. It addresses screening out qualified people, assessing relevant job skills, and reasonable accommodations, subject to the applicable legal conditions. The guidance also recommends examining tools before use and during operation.

The EEOC's guidance on employment tests and selection procedures discusses job-related selection methods, discrimination risks, and the employer's responsibilities. Use these sources with the agency's qualified reviewer when determining the obligations that apply. A completed software checklist does not establish legal compliance for a particular hiring process.

This article offers a proposed operational audit. It is intended to help an agency collect clear evidence and assign follow-up work. It does not decide whether a particular assessment or employment decision is lawful.

Describe exactly what the tool does

For each AI-assisted step, record the input, output, and intended use. Does the tool organize information for a recruiter, rank candidates, score an assessment, or recommend who should proceed? A vague description such as helps recruitment can hide material differences in how applicants are affected.

Identify the person who sees the output and the action they can take. Record whether they receive the original evidence, whether they can identify uncertainty, and what happens when they disagree. If the team cannot explain how the output affects the next decision, resolve that gap before relying on it in a hiring process.

Ask the supplier for information that addresses the proposed use. Keep the response attached to the audit record with its date and version. A demonstration using one set of sample inputs may leave questions about other applicants, configurations, or conditions. Write those questions down and assign someone to obtain an answer.

Walk through the applicant's experience

Review the invitation and instructions as part of the process. They should explain what the applicant is being asked to do, how the step relates to the application, and where to ask for help. Confirm that the contact route is monitored by someone who can arrange a response rather than merely acknowledge receipt.

Use appropriate accessibility testing and qualified input to examine the experience. Record which environments and interaction methods were tested, the scenario being attempted, and the observed result. Avoid labeling the whole workflow accessible because one person completed a demonstration on one device.

A useful audit note separates a reproducible technical problem from an unanswered policy question. For example, an unavailable control requires a clear reproduction record, while a question about an alternative assessment requires the responsible hiring or legal reviewer. Both need owners, but they need different kinds of investigation.

Make the accommodation route operational

An applicant should not have to guess whom to contact when the process creates a barrier. The responsible team needs a way to receive the request, acknowledge it, determine the appropriate next step, and communicate that step to the applicant. Handle personal information through the agency's approved process and limit access to those who need it.

Test the handoff internally using a labeled sample scenario. Confirm who receives the request, who can decide the next action, and how the recruiting team learns that the process needs attention. The test should check the workflow rather than seek or record unnecessary medical details.

Record the agreed response in terms that the recruiter can follow. Avoid an ambiguous note that leaves the applicant waiting while different teams assume someone else owns the decision. Where the next action requires specialist review, identify that owner and the information they need.

Inspect how reviewers use the result

Interview the people who use the tool in daily work. Ask them to explain what a result means, what it does not establish, and how they handle a case that needs more context. Compare that explanation with the tool's intended use and the agency's documented process.

Look for practical evidence of review. Can the recruiter retrieve the relevant input? Can they record a disagreement and explain the next action? Does the workflow preserve an unresolved question instead of forcing an unsupported conclusion? These observations help the owner identify where training, access, or process design needs attention.

Human involvement should be described through the work the reviewer performs. The presence of a person somewhere in the workflow is not enough to demonstrate that a particular output has been assessed appropriately. Keep the review task and the responsible role explicit.

Use an audit record that leads to action

For each finding, record:

• The exact stage and configuration being reviewed.
• The scenario tested and the evidence observed.
• The potential effect on the applicant's ability to participate or be assessed.
• The owner responsible for investigating or correcting it.
• The interim process, if one has been approved.
• The evidence needed to close the finding and the person who will review it.

Recheck a finding after the change has been made. A statement that the supplier has fixed the issue is a reason to test the relevant scenario again. Record the version or configuration tested so the team can understand what the closure applies to.

Keep product claims specific

InferOwl can support recruiting records and human review of communication evidence through Owl Intel. Those capabilities should be described according to the workflow your agency uses. They do not establish that an external assessment is accessible, that every applicant can use it, or that an employment decision meets the applicable legal requirements.

Use the human-review guide to understand the product's review boundaries. When an assessment or another external service is involved, include its owner and the agency's responsible reviewer in the audit rather than assuming that a connection transfers responsibility to one tool.

Repeat the audit when the process changes

Revisit the affected scenario when the assessment, supplier, configuration, role requirements, or way reviewers use the output changes. Include feedback from applicants and recruiters in the review queue. A recurring support issue may reveal a process problem that was not apparent in the initial demonstration.

Finish each review with a decision about the specific workflow, the unresolved findings, and the next check. That record gives the agency a practical way to maintain the process and explain why it continued, changed, or paused a particular use of technology.