Define the assessment before selecting the interface
Write down what the interview is meant to establish and how each question relates to the role. Decide what evidence would support a positive, negative or uncertain assessment. Avoid treating an accent, speaking pace or polished delivery as a substitute for the capability you actually need to assess.
This is a practical testing guide, not a claim that any interview model is accent-neutral or disability-safe. Those labels require evidence for the intended population and use. Melbourne employers should evaluate the actual interview pathway and provide accountable human review of consequential decisions.
Separate transcription from assessment
Test whether the transcript accurately represents the answer before you evaluate any summary or score built on it. Keep an authorised way to check the original recording where appropriate. Mark uncertain words visibly rather than silently replacing them with a plausible term.
Use approved test material that covers different accents, speech patterns, background conditions and domain vocabulary. Compare each transcript with a human reference and classify the mistakes: omitted qualifications, changed negation, incorrect names or harmless punctuation. One overall transcription score can hide errors that materially change an answer.
Give candidates a usable alternative
Give candidates a clear route to request an adjustment, and name the person who responds. Consider whether a live conversation, a written response or another accessible format would still assess the same capability. Start by locating the AI step in your process. The guide to AI recruitment software in Australia sets out recruitment stage by stage, with the point where a recruiter decides at each one. Explain the process before the interview, rather than making a candidate fail the tool first.
Check the interface with keyboard navigation, zoom, captions where relevant and assistive technology. W3C’s WCAG 2.2 materials provide criteria for evaluating web accessibility. A technical accessibility check and an appropriate interview adjustment are related but distinct tasks; do not assume one proves the other has been addressed.
Test errors that could change the decision
Take a hypothetical interview. The candidate says they have not yet completed a credential, but the transcript drops the word not, and the AI summary then describes the credential as completed. The review process should catch that transcription error before the statement reaches a shortlist or a client.
In another hypothetical case, a candidate pauses while using an assistive communication method. The assessment should go back to the agreed job evidence. If the system produces an unexplained fluency or confidence score, require evidence of its relevance and validity before anyone considers it in a decision.
Make the reviewer’s task explicit
Give the reviewer the role criteria, the source answer and any uncertainty flags. Ask them to verify factual summaries and to record the reason for a decision in their own assessment. Do not ask them merely to confirm a model score without access to the underlying material.
Sample accepted outputs as well as rejected ones, and record whether errors are corrected consistently and whether a candidate’s request for an alternative route changes how they’re treated. Take legal and accessibility advice for the circumstances; this operational guide is not a universal statement of every applicable duty.
Decide whether the tool is suitable
Agree the acceptance and stop conditions before the trial starts. Stop expansion if meaningful answers can’t be recovered, if uncertainty is hidden or if reviewers can’t explain the assessment using job-related evidence. Report the test population and conditions so nobody generalises the results beyond what was examined.
Repeat the relevant tests after any change to the transcription service, model, interview questions or interface. Keep a route open for candidates to raise errors and for staff to suspend a problematic workflow. A useful deployment decision should name its limits as clearly as its intended benefit.
Sources and reading
Primary sources checked 9 October 2026. The practical evaluation steps in this article are editorial recommendations, not measured customer outcomes.
