Decision engines — explainable, configurable, audit-trailed.

Both contribute to the same decision through one pipeline. Rules handle deterministic logic (hard declines, regulatory cut-offs, fraud signals); models handle probabilistic signals (default prediction, fraud scoring, affordability). The decision-record format is the same regardless of which path drove the outcome — every decision retains its rule path, its model contributions, the feature inputs at the time of the decision and an explanation a human reviewer can read. Rules and models are versioned separately but together — a decision is reconstructable from rule version + model version + inputs.

Three properties: (1) a reasoned-decision record per decision that’s human-readable; (2) per-decision model contributions (SHAP or equivalent), not just population-level model explanations; (3) an adverse-action notification flow that maps decision drivers to plain-language reasons. The reasoned-decision record is the centrepiece — it’s what you produce when the customer requests an explanation, what the regulator queries in a review, and what your compliance team uses when defending an outcome.

Yes, with parallel-run mode. The replacement engine runs alongside the legacy one for 30–90 days, both producing decisions on the same applications. Decisions are diffed continuously, root-cause analysis runs on every disagreement, and the rules / models are tuned until the disagreement rate is below the agreed threshold. Then the replacement engine takes traffic gradually — usually 5%, 25%, 50%, 100% — with rollback ready at any stage. Decision continuity is preserved because the legacy engine is still running for comparison the whole time.

Shadow-mode evaluation first — the challenger runs on the same applications as the champion but the customer never sees the challenger’s decision. Disagreements are logged and analysed. Once the challenger meets the agreed criteria (decision agreement, portfolio-level outcomes, compliance review), it takes a traffic slice — typically starting at 5–10%. Outcomes are tracked per slice and the challenger is promoted to champion when the comparison is conclusive. The discipline is in deciding the promotion criteria up-front, not after the data lands.

Credit scoring is a high-risk AI system under the EU AI Act from August 2026. The obligations include: a documented risk-management process, training-data governance, technical documentation, transparency to users, human oversight, accuracy / robustness / cybersecurity, and post-market monitoring. Most of this is buildable as platform capabilities once rather than per-model — risk-management dashboard, training-data lineage, model documentation generator, human-oversight workflow. The technical bar is genuinely meetable; the planning bar (deciding who in your firm is accountable for each obligation) is the harder part.