Regulated Workflow Matrix

This page maps common regulated AI workflows to the technical controls engineers usually implement with Briefcase AI.

Workflow Matrix

Workflow	Typical Inputs / Versioned Artifacts	Decision Trace Fields	Replay / Debug Loop	Human Review Trigger
Onboarding / KYC	Identity docs, policy/rule set commit, scoring config	decision_id, model/version, confidence, artifact commit refs	Re-run onboarding decisions after policy/model changes	Low confidence, conflicting signals, unresolved references
Sanctions Screening	Watchlist snapshot commit, matching thresholds, screening config	match score, threshold, list commit, route outcome	Replay match decisions against historical list/config state	Borderline match score or policy override
Fraud Decisioning	Transaction features, fraud model version, risk policy commit	risk score, threshold, action, feature snapshot	Compare historical decisions after model update or drift alert	Score in review band, anomaly flags
AML Alerting	Rule configuration version, monitoring parameters, entity context	alert reason codes, rule version, escalation path	Reconstruct why alerts were generated/suppressed at a point in time	High-risk or unresolved alert rationale
Credit / Underwriting	Application features, policy version, model config	decision class, confidence, feature set version, reason codes	Replay denials/approvals after rule or model revision	Low confidence, policy-exception request
Real-Time Payment Routing	Routing model version, risk rules, operational configuration	route decision, confidence, latency bucket, policy refs	Re-run routing decisions for disputes or post-incident analysis	Confidence below threshold or high-impact transaction

Implementation Pattern

For each workflow, a common implementation sequence is:

1. Resolve and attach versioned artifact references.
2. Capture decision input/output metadata with confidence.
3. Persist immutable trace records.
4. Expose retrieval and replay endpoints for operations/audit.
5. Route low-confidence outcomes to structured human review.

Engineering Checklist

• Stable function_name values for replay comparability.
• Commit-linked artifact references on every decision.
• Confidence and decision-route tags for triage.
• Replay checks in deployment validation workflows.
• Policy-driven routing thresholds managed as config.

Next Steps

• lakeFS Integration
• Version Control Systems
• Multi-Agent Correlation
• Prompt-Knowledge Validation
• System Design