• integrated_82#Q4.10

RAIDT can be used in finance as a governance method for AI-assisted explanation work where reasons, reviewability and challenge are central. In the ageing-society paper, finance is translated through scenarios such as adverse action explanations for credit refusal, benefit-eligibility explanations, suspected fraud communications and complaint responses. The key move is to treat the "run as the unit of governance" so that each generated explanation is accompanied by a "run-level evidence pack" and a RAIDT "score profile". That record makes the five pillars (Responsibility, Auditability, Interpretability, Dependability, Traceability) inspectable for one concrete case rather than only at policy level. In practice, this means storing the prompt/template version, decision-record identifier, policy basis, any retrieval snapshot, output hash, reviewer judgement and escalation route.

For finance, the papers emphasise that interpretability is not merely stylistic clarity; it is procedural fairness and contestability. An adverse-action notice should therefore present plain-language reasons, reason codes, the policy basis, and clear next steps or appeal routes. RAIDT also treats "influence methods as governance interventions": for example, structured prompting is suited to fixed-template explanations, while RAG can ground outputs in the exact policy text used. The healthcare playbook adds the operational discipline for scoring and review: each run is assessed with fixed anchors 1=missing / 3=partial / 5=audit-ready, so a bank or lender can see whether an explanation was merely readable or genuinely reconstructable and reviewable. Crucially, the ageing-society paper warns that prompt-only improvements may improve clarity, but they do not by themselves secure auditability or traceability unless logging and version control are built into the process.

A lender uses GenAI to draft an adverse-action notice after declining an older customer's credit application. Under RAIDT, the system does not make the lending decision; it explains a recorded decision using a structured template. The notice states the reason codes, cites the policy basis, explains what information was considered, and gives a clear appeal route or next step. If policy text is retrieved, the retrieval snapshot is logged alongside the prompt version and decision-record ID.

The compliance team then reviews the run-level evidence pack and assigns a score profile across the five pillars (Responsibility, Auditability, Interpretability, Dependability, Traceability). If the draft is understandable but the policy snapshot is missing, Interpretability may be strong while Auditability remains only partial. That makes the governance gap visible before the notice is issued. The same record can later support complaint handling, internal audit or a customer challenge without relying on memory or informal reconstruction.

• 20-RAIDT_AgeingSoc_M_V50
• 21-RAIDT_Sector_Playbook_Healthcare_V2