Model logs without context, no consumer-facing reason record, contested decisions handled ad hoc.
The scenario
Who: A consumer lender using AI-assisted credit decisioning.
The problem: Declined applicants appeal; the lender can’t reconstruct what features drove the decision two months later; OAIC asks for evidence of the human-review pathway.
Per-decision explanation with top reasons, reviewer queue, OAIC report pack ready on demand.
What it produces
From 10 December 2026, new APP 1.7-1.9 obligations require organisations using ADM in customer-facing decisions to disclose how those decisions are made in their privacy policy. The underlying records the disclosure summarises are the per-decision rationale Fydis captures: input features, model version, decision, and reasons. Contested decisions route to a human review queue. The OAIC report pack regenerates on demand from the same trace.
“Why was application #58931 declined?”
Declined
RetrieveModel v2.3 · 7 input features · all auditable
VerifyPlaid · Yoti · Equifax · all logged
EvidenceHuman review · 48h appeal window
EvidenceOAIC report pack · on demand
The analysis, step by step
- 01RetrieveInput features captured at decision time with provenance
- 02VerifyModel decision recorded with version hash
- 03VerifyTop reasons extracted (model-agnostic; SHAP or rule-based)
- 04EvidenceConsumer notice generated · plain-English reasons
- 05EvidenceContested decisions queued for human review · reviewer sign-off captured
- 06EvidenceOAIC report pack regenerated on demand from the trace
Frequently asked
Does this work with any model?
Yes. Fydis is model-agnostic, we capture the inputs, decision, and reasons regardless of whether the model is rule-based, gradient-boosted, or LLM-driven.
What about derived features?
On roadmap. We currently trace raw inputs and top-feature attributions; full pipeline lineage for derived features is in flight.
Run this on your data.
The live demo runs this exact analysis on sandboxed data. Book a 30-minute briefing and we’ll show you the same chain on your data, your approval policy, and your regulator clause map.