AI in APS recruitment: the APSC principles and what to demand of a vendor
APS agencies were expected to implement the APSC’s principles for AI in recruitment from 1 June 2026. They sit on top of the merit principle rather than replacing it — which means an agency has to show both that AI was used responsibly and that the assessment remained genuinely comparative.
- The APSC has issued principles for agency use of AI in recruitment, with agency implementation expected from 1 June 2026.
- Merit under the Public Service Act still governs: the assessment must be competitive, relevant and based on work-related qualities.
- Agencies need a documented statement of where AI is used, a named delegate for each decision, candidate disclosure and an accessible alternative path.
- A vendor that cannot evidence an individual assessment cannot support a merit-based delegate decision.
- Public-sector procurement questions are a good proxy for what private buyers will be asked for next.
Why the APS moved first
The Australian Public Service hires at scale, publishes its selection criteria, and is subject to a statutory merit requirement that private employers are not. That combination makes AI-assisted screening both unusually attractive — bulk rounds, thousands of applicants, published criteria that map cleanly to a rubric — and unusually exposed, because a merit decision has to be explainable to an unsuccessful candidate and to the Merit Protection Commissioner.
The APSC has responded with principles for agency use of AI in recruitment, which agencies were expected to have implemented from 1 June 2026. They do not displace the merit principle in the Public Service Act; they describe how to use AI without breaching it.
Merit is the constraint that shapes everything else
Merit-based selection requires an assessment that is competitive, based on work-related qualities, and relevant to the duties. Three consequences follow for any AI tool an agency puts in the funnel.
- 01The rubric has to trace to the published work-level standards and the selection criteria for the role. A generic vendor competency model that cannot be mapped to those documents is not assessing merit; it is assessing something else and calling it merit.
- 02The assessment must remain comparative. A tool that scores each candidate against a fixed threshold in isolation does not, on its own, produce the comparative assessment merit requires — the delegate has to be able to compare candidates on the same evidence.
- 03The delegate decides. AI output is material before the delegate, not a substitute for the delegate’s judgement, and the record must show the delegate engaged with it.
Principles mapped to the artefacts to require
Principles are useful to agencies and useless to procurement unless they are translated into things a vendor must actually hand over. This is that translation.
| Principle theme | What the agency must be able to show | Artefact to require from the vendor |
|---|---|---|
| Transparency to candidates | That applicants were told AI would be used, before it was | Candidate disclosure text, the point in the flow it appears, and a timestamped acknowledgement record |
| Accountability | A named human accountable for each selection decision | Per-decision record naming the reviewer, their rationale and any override of the system ranking |
| Merit and relevance | That criteria trace to the work-level standards and role duties | Rubric configuration exported in full, versioned, with the mapping to your selection criteria |
| Fairness | That the tool does not disadvantage a protected group | Per-requisition adverse-impact analysis at the four-fifths threshold, plus any independent bias audit |
| Privacy | What personal information the model uses and what decisions it supports | Written input field list and APP 1.7 limb classification |
| Accessibility | That candidates who cannot use the AI channel are not excluded | Documented alternative assessment path and reasonable-adjustment process |
| Explainability | Why a specific candidate scored as they did | Evidence-linked scoring: verbatim, timestamped transcript evidence behind every rating, or an explicit abstention |
| Records | That the decision can be reconstructed later | Immutable per-decision record with a retention period meeting agency and Fair Work requirements |
The alternative-channel requirement people underestimate
Every AI screening rollout in the public sector runs into the same question late and expensively: what happens to the candidate who cannot complete an AI interview? Low bandwidth in a regional town, a speech difference, a screen-reader dependency, a disability that makes timed video intolerable, or simply a refusal to be assessed by a machine.
Under the Disability Discrimination Act the agency owes reasonable adjustments regardless of what the tool supports, so an alternative path is not optional. Design it up front and it costs little: a text-based interview channel, an untimed mode, an extended-time mode, and a documented route to a human-conducted equivalent.
- Offer the interview in text as well as voice and video, so the modality itself is not a barrier.
- Make timing adjustable per candidate rather than per requisition.
- Publish the adjustment request route in the invitation, not buried in a help centre.
- Score the alternative path against the same rubric so the comparative assessment survives.
- Record the adjustment and its outcome in the decision file — it is evidence of compliance, not a blemish.
Questions to put in the approach to market
- 01Produce a complete per-decision record for one candidate, redacted, as part of your response.
- 02State in writing every field of personal information your scoring model receives, and confirm whether video frames or audio features are used at any stage.
- 03Describe how a criterion that would be unlawful under the Fair Work Act or the discrimination Acts is prevented from being configured — not detected afterwards, prevented.
- 04Describe the alternative assessment path for a candidate who cannot or will not use the AI channel, and how it is scored comparably.
- 05State your data residency options and your retention configuration ceiling.
- 06Provide your adverse-impact methodology, the threshold used, and whether analysis runs before shortlists ship.
- 07Provide the classification of decisions your product makes solely automatically versus substantially and directly supports, for our APP 1.7 disclosure.
- 08Provide model, prompt and rubric versioning evidence, and describe how a mid-round change is recorded.