Skip to content

Template: Escalation Decision Matrix (EDM)

Purpose: Define auditable escalation triggers, routing, SLAs, and approval authority for agent behaviors—especially where agents may touch sensitive data or take actions.
Applies to: Zone 2 (recommended) and Zone 3 (required).
Regulatory driver: FINRA emphasizes that GenAI can implicate supervision and recordkeeping and highlights the need for governance, testing, and ongoing monitoring (including prompt/output logs, model version tracking, and human-in-the-loop review).
Related controls (examples): 2.12 Supervision & Oversight, 3.4 Incident Reporting/RCA, 1.7 Audit logging, 3.3 Compliance reporting.

1) Matrix metadata

  • Matrix name: (e.g., “Zone 3 Autonomous Agents – Escalation Matrix”)
  • Business unit / function:
  • Applies to agent(s): (IDs/names)
  • Zone: Zone 2 / Zone 3
  • Governance level: Recommended / Regulated
  • Owner (Compliance):
  • Owner (Engineering/Platform):
  • Approver (Risk/Legal):
  • Effective date:
  • Next review date: (recommended: quarterly for Zone 3)

2) Severity model (choose one and standardize)

Option A: 3-level severity (simple)

  • S1 (Critical): Immediate customer/regulatory harm likely, or security breach indicators.
  • S2 (High): Material policy violation, sensitive data exposure risk, repeated failures.
  • S3 (Medium): Non-material issue, degraded performance, isolated minor policy warnings.

Option B: 4-level severity (more granular)

  • P0 / P1 / P2 / P3 (incident-response style)

Pick one model for the entire program to ensure reporting consistency.

3) Escalation triggers table (core of the matrix)

Trigger category Trigger condition (objective) Severity Auto action Escalate to (role) SLA Required evidence
Unauthorized action Agent attempts prohibited action from AAM S1 Block + alert Compliance On-Call + SecOps 15 min Decision log + audit event IDs
Scope drift New connector/data scope detected (not approved) S1 Block + alert SecOps + Platform Owner 30 min Drift signal + baseline
Sensitive data event Confidential/Restricted label accessed unexpectedly S1 Alert; optional block Data Protection Officer 30 min DLP match + resource IDs
Low confidence confidence_band = Low for “recommend/execute” S2 Require human review Assigned Reviewer 4 hrs Decision log + rationale
Hallucination risk Conflicting sources or “unsupported claim” flags S2 Require review SME Reviewer 8 hrs Source list + output category
High volume Rate anomaly exceeds threshold S2 Throttle Platform Owner 4 hrs Telemetry + threshold
Repeated minor issues 3+ S3 within 7 days S2 Review required Compliance Lead 1 biz day Summary report
Policy exception request User requests override S2 Block until approved Approver role 1 biz day Override request record
Minor QA issue Non-critical formatting/UX defect S3 Log only Product Owner 5 biz days Issue ticket

Notes - “Trigger condition” must be measurable (avoid subjective language). - “Auto action” must be explicit (Block, Throttle, Require Review, Log Only). - "Required evidence" must reference the specific logs/artifacts (Decision Log Schema + Purview audit references where applicable).

4) Routing map (who gets notified)

Routing roles (fill in with org-specific groups)

  • Compliance on-call: (name / DL / Teams channel)
  • SecOps on-call: (name / DL / Teams channel)
  • Data Protection: (name / DL)
  • Business owner: (name / DL)
  • Model/AI owner: (name / DL)
  • Legal: (name / DL)
  • Incident commander: (name / DL)

Notification channels

  • Teams channel:
  • Email distribution list:
  • Ticketing system queue:
  • Pager/on-call tool:

5) Override policy (how exceptions are handled)

FINRA notes that supervisory systems must be tailored and that firms should consider integrity/reliability/accuracy where AI is used, with ongoing monitoring and human-in-the-loop review.
Overrides must therefore be rare, explicit, and auditable.

  • What can be overridden? (e.g., throttling, non-critical blocks)
  • What cannot be overridden? (e.g., prohibited actions, sensitive data boundary violations)
  • Approval authority: (role-based)
  • Override duration: (time-boxed)
  • Logging: every override generates a decision log record + ticket.

6) Evidence bundle checklist (for audits/exams)

For each S1/S2 escalation: - [ ] Decision log record ID(s) - [ ] Audit log event reference(s) - [ ] Scope baseline reference (if drift) - [ ] AAM clause reference (if unauthorized action) - [ ] Ticket ID and timestamps (SLA evidence) - [ ] Root cause analysis summary (if incident) - [ ] Corrective action + prevention (CAPA) entry

7) Operational metrics (monthly reporting)

  • Count of S1/S2/S3 events by agent and category
  • SLA adherence (met vs missed)
  • Override rate
  • Repeat event rate (“recurrence”)
  • Mean time to detect (MTTD) and mean time to respond (MTTR)

These feed the Real-time Compliance Dashboard and compliance reporting controls in your Reporting pillar.

8) Approvals and versioning

  • Version:
  • Approved by (Compliance):
  • Approved by (Risk/Legal):
  • Approved by (Platform owner):
  • Approval date:
  • Change record (ticket/PR):