| Part | What it answers | Who reads it first |
|---|---|---|
| Part I: Foundations | What is categorically new about agentic AI; the AR-1…AR-16 risk taxonomy; the threat model for a financial institution | CRO, CAIRO, board risk committee |
| Part II: Governance & Accountability (GOV-01–14) | Who decides and who answers: board duties, AIMS, MRM extension, three lines, the autonomy ladder and its gates | Head of model risk, CRO, second line |
| Part III.A: Design-Time Controls (DES-01–14) | How to build agents so bad outcomes are structurally difficult: data governance, impact assessment, least privilege, supply chain, memory architecture | Engineering lead, security architecture |
| Part III.B: Pre-Deployment Evaluation (EVL-01–16) | What testing proves, what it cannot, and how to make evaluation results decision-grade, including validity, sandbagging, and the go/no-go gate | Head of validation, eval team |
| Part III.C: Runtime Controls (RUN-01–18) | What stands between the agent and the ledger while it acts: oversight modes, AI-control protocols, sandboxing, kill switches, hard limits, tool gating | Engineering/platform lead, first line |
| Part III.D: Monitoring & Post-Deployment (MON-01–12) | How to see what agents actually do, detect drift and incidents, and meet external reporting clocks | Operations, SOC, compliance |
| Part III.E/F: Multi-Agent & Third-Party (MAS-01–10, TPR-01–08) | What changes when agents interact and when the model belongs to a vendor: identity, delegation, collusion, concentration, upstream change | Platform lead, TPRM, model risk |
| Part IV: Assurance, Audit & Certification (ASR-01–10) | How to prove the controls exist and operate: internal audit, the 42001/42006 certification chain, vendor safety frameworks, transparency artifacts | Internal audit, second line |
| Part V: Regulatory Mapping | Which binding obligations attach, per regime, and which control families discharge them: one master table | Compliance, legal, CRO |
| Part VI: Fairness & Customer Outcomes (FCO-01–10) | How agentic flows acquire discrimination and how to prevent, detect, explain, and remediate it in covered decisions | Fair-lending compliance, model risk |
| Part VII: Implementation Playbook | Where to start and in what order: maturity model M1–M4, RACI, 90-day plan, year-one roadmap, board pack, the ten controls to build first | Program lead, CRO, everyone once |
| Appendices A–E | Master control index; crosswalk matrices; glossary; bibliography; incident severity & reporting matrix | Reference: all readers |
Flagship compendium · Section 3 of 24