Skip to main content
Decision Keep
← All posts
5 min readPerspectives & Ideas

XAI credit and fraud risk decisioning: explainability plus an evidence backbone

Explainable AI explains why a model scored a credit or fraud decision the way it did. That is not the same proof that the decision is intact and defensible. Here's how to pair XAI with a signed, hash-chained, offline-verifiable decision record.

About the author+

Jamil Luketic

Executive Director at Decision Keep

Former Data & Tech Leader at Oracle, Mastercard, Coles, Optus, and Reece.

Connect on LinkedIn
Illustration for XAI credit and fraud risk decisioning: explainability plus an evidence backbone
Perspectives & Ideas

A credit decline and a fraud flag look different at the point of decision, but they share the same regulatory problem: the person affected is entitled to know why, and the organisation must be able to prove the record hasn't been changed since the moment the model decided.

Explainable AI (XAI) - SHAP, LIME, counterfactuals - answers the why. A decision audit trail answers the did it change. Both matter, and neither substitutes for the other.

XAI tells you the reasoning. The trail proves the outcome.

When a credit decision is explained, XAI shows the contribution of each feature: tenure, repayment history, debt-to-income. That is essential for fairness and for answering a dispute. But XAI output itself is generated after the decision, by tooling that can be re-run or re-tuned - it does not bind the decision to a moment in time, and it does not prove that the inputs fed to the model were the ones recorded here, not ones swapped in later.

What regulators and courts actually require is a record that is:

  • Authentic - signed with the organisation's own key (Ed25519), not the vendor's.
  • Intact - hash-chained, so editing any entry breaks the chain.
  • Timed - anchored by an independent RFC 3161 timestamp.
  • Verifiable - an auditor checks a receipt against the published key with no account and no trust in the platform.

The European Commission's Trustworthy AI guidelines, NIST's AI RMF, and MAS's AI governance framework all converge here: explainability and traceability are distinct obligations, and the latter must be independent of the model that produced the explanation.

Where credit and fraud risk decisioning diverge

Aspect Credit risk decisioning Fraud risk decisioning
Decision weight Legal or similarly significant effect on a person Financial loss prevention; false positives block legitimate customers
Regulatory driver Equal credit opportunity, APP 1.7-1.9, GDPR, EU AI Act Art. 26 Transaction monitoring rules, PSD2 strong customer auth
Dispute surface "Why was I declined?" - logic + record integrity "Why was my card frozen?" - continuity + false-positive rate

In both, XAI gives the first line of defence - an answer to the customer. The tamper-evident trail is the second - an answer to the regulator or the court, which must hold up without the model team in the room.

How a defensible record contains all three

  • Model version + inputs (references, not raw PII) - so the XAI explanation can be tied to the exact decision.
  • Score, confidence and routing outcome - the actual outputs the model produced.
  • Your organisation's Ed25519 signature - proof of origin and integrity, over the canonical payload.
  • Hash chain - each decision links to the previous; tamper with one and the break is detectable.
  • Independent RFC 3161 timestamp - proves when the decision was sealed.
  • Verifiable erasure - so retention expiry is demonstrable without destroying the chain.
AI decides credit / fraud score + inputs XAI explains SHAP / LIME feature weights Forensic witness sign + chain + timestamp + verify
XAI explains the decision's logic; the Forensic Witness proves the decision itself is authentic, intact and time-anchored.

How Decision Keep fits

Decision Keep records each credit and fraud decision your models make - signed with your own key, hash-chained and time-anchored - so the XAI explanation you already produce has a tamper-evident backbone to stand on.

FAQ

Questions auditors, risk and legal actually ask

What is the difference between XAI and a decision audit trail?+
XAI (e.g. SHAP, LIME) explains the model's reasoning for a score or prediction. A decision audit trail proves what the system actually decided, by which model version, when, and that the record has not been altered since. Explainability shows the logic; the audit trail proves the outcome.
Do credit and fraud risk decisioning need both?+
Yes. Credit declines and fraud flags are legal-or-similarly-significant effects, so fairness and dispute handling need XAI. But the dispute, regulator or court will challenge the record itself - hence a tamper-evident, verifiable trail. Neither alone satisfies both 'was it fair' and 'has it been changed'.
Can I use my existing XAI toolkit with Decision Keep?+
Yes. Decision Keep records the decision your model made (inputs as references, score, model version, routing outcome) after the fact, so any XAI output for that same decision can be cited alongside a verifiable receipt.

Sources

References & further reading

Independent analysis and standards cited in this article.

Prove every AI decision

Decision Keep gives your organisation a tamper-evident, verifiable record of every automated decision. Book a demo to see it on your stack.

Book a demoSee compliance mapping

Keep reading

Documentation

Go deeper in the docs