Why an AI score is not a shortlist explanation
An AI score compresses several inputs into a convenient signal. It may help a reviewer decide where to look first, but it does not say which approved requirement was met, which source supports it, or what uncertainty remains. “The score was 82” is therefore not an auditable reason to progress or reject a candidate.
The EU AI Act discusses transparency and human oversight for high-risk systems. It does not prescribe this record. This is Skilltage's operational guidance for keeping the human decision visible when software helps organise evidence.
The five-part evidence record
For each material shortlist decision, write five short fields:
- Criterion: the approved, role-related requirement being considered.
- Source evidence: the document or answer, with a concise excerpt or location.
- Uncertainty: missing, ambiguous, contradictory, or failed analysis evidence.
- Reviewer action: inspect, clarify, compare, pause, or proceed.
- Decision and owner: progress, pause, or do not progress, recorded by the named human reviewer.
Keep any score in a separate routing field. Never use it as a substitute for the evidence record.
Worked example
Criterion: “Can independently troubleshoot production equipment after onboarding.”
| Record field | Example |
|---|---|
| Criterion | Approved troubleshooting requirement in the role brief |
| Source evidence | CV: “diagnosed recurring line faults across two sites”; application answer gives one concrete incident |
| Uncertainty | Independence is described, but escalation limits are unclear |
| Reviewer action | Ask one consistent clarification question |
| Decision and owner | Pause pending answer — recorded by the hiring manager |
The score may have placed the candidate in the review queue. The record explains what the reviewer actually considered and why the next step is proportionate.
Checklist before closing the decision
- Was the criterion approved before candidate review?
- Can another reviewer find the source behind the summary?
- Have missing evidence and contradictory evidence stayed separate?
- Is the requested clarification the same for comparable candidates?
- Does the decision state who acted and when?
- Does the record avoid protected or proxy characteristics?
Skilltage guidance: use AI to surface and organise role-relevant evidence, then require a reviewer to inspect the evidence and select the candidate status. The product does not make final shortlist decisions.
Limitations and decision boundaries
An evidence record improves explainability; it does not make a criterion valid, a process lawful, or a decision fair by itself. Teams remain responsible for job-related criteria, accessibility, accommodations, candidate communication, retention, and final decisions. Do not infer capability from a high score or incapability from a low one. If source analysis fails, recover or inspect the source before acting on the gap.
Start with human-controlled AI screening boundaries, then use missing evidence as an explicit review state.
Next step
Take the next shortlist decision your team makes and complete the five fields before opening the score. If the record cannot be completed, fix the criterion or source workflow before relying on the ranking signal.