AML Copilot ABPA 10-Day Starter

This is the first executable ABPA capstone package. It converts existing AML Copilot work into BA / PM / Architect decision assets.

0. Evidence Already Available

1. 10-Day Output Map

---

2. AI Opportunity Canvas v0.1

Scenario

Business Problem

Baseline Metrics to Collect

AI Fit

Candidate AI Roles

No-AI Boundaries

Initial Success Metrics

Recommendation

---

3. Stakeholder Evidence Map v0.1

Interview Questions by Stakeholder

AML Investigator

1. Which part of evidence gathering takes the most time?
2. Which data sources do you trust most and least?
3. What makes a SAR draft unusable?
4. What kind of AI suggestion would you override immediately?
5. What evidence must be visible before you trust a typology recommendation?
6. What shortcuts do analysts use today when backlog is high?
7. What is the most common reason a case gets returned for rework?
8. What should never be hidden behind an AI summary?

Compliance Officer

1. What minimum evidence is required before a SAR narrative is acceptable?
2. Which statement types are most dangerous if hallucinated?
3. What audit fields must exist for AI-assisted work?
4. What should trigger mandatory senior review?
5. How should AI-generated content be labeled?
6. What would make you block the pilot?
7. What sample size would you need for quality review?
8. Which regulatory expectations are non-negotiable?

Data / IT / Security

1. Which systems are source of truth for party, account, transaction, KYC, sanctions, and case status?
2. Which fields are PII or restricted?
3. What data can leave the bank boundary?
4. What identity and authorization model is required?
5. What logs can be retained, and for how long?
6. What availability and incident response expectations apply?
7. How will model/provider outages be handled?
8. What integration path is realistic for a pilot?

---

4. AS-IS Workflow v0.1

flowchart TD
  A[Transaction monitoring alert created] --> B[Alert assigned to AML investigator]
  B --> C[Open case and review alert reason]
  C --> D[Gather customer/KYC/account evidence]
  D --> E[Gather transaction and counterparty evidence]
  E --> F[Compare activity with typologies]
  F --> G{Suspicious enough?}
  G -->|No| H[Close as false positive with rationale]
  G -->|Yes| I[Draft SAR narrative]
  I --> J[Compliance / senior review]
  J --> K{Approved?}
  K -->|No| L[Return for rework]
  L --> D
  K -->|Yes| M[Submit / archive evidence]

Pain Metrics to Collect

TO-BE Hypothesis

flowchart TD
  A[Alert assigned] --> B[Copilot gathers evidence]
  B --> C[Copilot scores typology with cited evidence]
  C --> D{Confidence and risk gate}
  D -->|Low confidence/high risk| E[Mandatory human review]
  D -->|Normal advisory| F[Analyst reviews closure rationale]
  D -->|Suspicious advisory| G[Copilot drafts source-bound SAR]
  F --> H[Analyst disposition]
  G --> I[Analyst edits and submits for compliance review]
  E --> I
  I --> J[Audit trail export]

---

5. Requirements-to-Eval Matrix v0.1

---

6. AI Control Pack v0.1

---

7. Executive Decision Memo v0.1

Decision Requested

Fund a 30-day discovery and controlled prototype hardening phase for AML Investigation Copilot. Do not approve production rollout yet.

Why Now

• The repository already has a functioning AML domain model, synthetic labeled dataset, rule baseline, failure taxonomy, SAR draft generator, and eval scaffolding.
• The business workflow is high-value: AML investigations are evidence-heavy, language-heavy, regulated, and audit-sensitive.
• The current prototype is honest about its limits: rule-template SAR, synthetic data, no production PII, no automatic filing.

Options

First 30 Days

1. Validate production baseline metrics.
2. Confirm data source and access model.
3. Build stakeholder evidence map and workflow baseline.
4. Convert top requirements into evals.
5. Add human review and audit control pack.
6. Design pilot scope and stop rules.

Success Gate

Proceed to pilot only if:

• cited evidence validity is 100% on test set,
• compliance approves HITL operating model,
• historical or external labels are available for non-circular eval,
• analyst usability test shows task success without trust confusion,
• control owners accept RACI.

Stop Rule

Stop or revert to workflow-only if:

• the system cannot cite evidence reliably,
• labeled data cannot be obtained,
• reviewers cannot explain AI-assisted recommendations,
• critical hallucination appears in controlled tests,
• data/privacy review blocks model use.

---

8. Next Work Items