Inclusion starts before the model
Map the candidate journey from finding the vacancy to receiving an outcome. Include the application form, account creation, document upload, interview booking, assessment and communication. An accessible model output can’t repair a form a candidate can’t use, or a requirement that has no clear connection to the role.
Ask the hiring manager which tasks are essential and what evidence is acceptable. Separate required credentials from preferences and from wording inherited from the last job ad. If a criterion can’t be explained in job terms, resolve it before it becomes a search filter or a scoring instruction.
Make the application pathway usable
Test keyboard access, clear labels, understandable errors and whether a candidate can recover from an interrupted upload. Review the process on smaller screens and at increased zoom. Avoid asking candidates to re-enter the same information without a reason. Give them a contact route for when the digital process can’t be completed.
Use the W3C accessibility materials as a technical reference and involve people with relevant access needs in the evaluation. Record the scope of the test. A passing check on one page or device should not be described as proof that the entire recruitment process is accessible.
Keep alternative evidence paths open
For each assessment, decide whether the format is essential to the role or simply the one AI recruitment software in Australia offers by default. Define how a candidate can request an adjustment and how the request reaches someone able to act. Keep that conversation separate from assumptions about suitability.
A hypothetical role might require explaining a technical process accurately, while the proposed assessment requires a timed video response. Consider whether a different format can assess the same capability. Document the reasoning rather than treating the default platform setting as an occupational requirement. If the assessment is an AI-assisted interview, the accent-fair AI interviews guide covers testing interview access and human review separately.
Review AI-prepared material against sources
If AI prepares an application summary, require a human to verify it against the submitted evidence. Preserve uncertainty around missing dates, unfamiliar qualifications and career changes. Don’t turn missing information into a claim that the candidate lacks a capability.
Build a test set of varied application formats and employment histories, using authorised material. A smooth summary can still drop a detail or change what it meant, so examine the omissions and changed meaning. Ask reviewers to explain the reason for progression or non-progression using the agreed criteria.
Investigate differences without oversimplifying them
If outcomes differ across groups or application pathways, investigate the stage where the difference appears. Consider access barriers, source quality, criteria and reviewer behaviour. Obtain appropriate advice before collecting demographic or sensitive information for analysis, and restrict access to that material.
A result from a small trial may be uncertain. Report the sample, the missing data and the limits rather than labelling the process bias-free. Keep a record of what you did after a concern was raised and whether the change dealt with the problem you observed.
Give the programme an owner
Name an owner for each of the application pathway, assessment design, AI configuration and candidate support. Agree who can stop a workflow when a problem is found. A supplier should be able to explain how the employer can correct errors and change a configuration without losing control of records.
The Australian Human Rights Commission’s AI recruitment checklist is a useful review reference. Use it alongside advice for your organisation and evidence from the actual workflow. Inclusive recruitment is an ongoing operational responsibility, not a property established by an AI product label.
Sources and reading
Primary sources checked 9 October 2026. The practical evaluation steps in this article are editorial recommendations, not measured customer outcomes.
