Goal: Assess maturity of agent lifecycle, moderation, accountability, risk control, and audit across the full spectrum of autonomous agent operations.
“Every agent must have a traceable lifecycle from creation to deactivation, with clear ownership, constrained autonomy, and auditable decision trails.”
Audit Checklist
Full agent lifecycle: creation, authorization, memory, activation, constraints, deactivation
Ownership logic: who is responsible for agent decisions and how is it proven
Content publication and moderation, including agent-agent coordination detection
Audit trail of activity, log retention, replay audit capability
Agent risk scoring and automatic routing to human review on abuse patterns
Support capacity modeling, AI-to-human handoff efficiency, cost-per-resolution
Typical Weaknesses
Dispersed knowledge sources — no single source of truth for policies across platforms
No shared error classification — same error type categorized differently on Facebook vs. Instagram
Weak explainability — AI routing decisions lack transparency for escalation reviewers
Handoff friction — cross-platform case transfers lose context and priority
Capacity planning siloed — support teams optimized per-platform, not cross-Meta
Procedura Wysylkowa — Outreach Questions 12–16
Q12Please provide a process map from intake to resolution, distinguishing AI vs human actions.
Q13Please list appeal paths and exceptional handling with most common escalation causes.
Q14Please describe confidence thresholds, manual review triggers, and decision owners.
Q15Please identify 10 largest sources of friction, retry, and delays in support and moderation.
Q16Please describe decision logging, exception handling, and case history linkage.
🏦
Playbook 3 — Regulated Finance
Banks
Goal: Regulated-first environment. Map AI/agent deployment across advisory, support, operations, and research — with strict decision class governance.
“In banking, every agent decision must be classifiable: who decides, who approves, who audits, and who owns the outcome. No exceptions.”
Decision Classes
Class A — Human Only
Full human decision-making. No agent involvement permitted. Applies to credit decisions above threshold, regulatory filings, and customer complaints escalated to ombudsman.
Class B — Agent Recommends, Human Approves
Agent performs analysis and generates recommendation. Human reviews, approves, or rejects. Applies to fraud alerts, loan pre-qualification, and compliance checks.
Class C — Agent Acts, Human Audits Ex Post
Agent executes autonomously within defined parameters. Human audits decision after the fact. Applies to routine transaction monitoring, document classification, and FAQ responses.
Class D — Full Automation, Low Risk
Fully automated with no human review required. Applies to balance inquiries, statement generation, password resets, and standard notifications.
Type 1: Bank with Developed AI Support
Already deploying AI/LLM in customer support, advisory, or operations
Areas to Investigate
Current AI support coverage: which processes, which decision classes
Knowledge base integrity: are approved sources the only sources feeding AI decisions
Audit trail completeness: can every AI-assisted decision be replayed and explained
Exception handling: what happens when the AI hits a boundary or low-confidence scenario
Matched Agents
Agent
Role
Focus Area
Customer Support Auditor
Support quality
Resolution accuracy, response time compliance, escalation appropriateness
Knowledge Base Integrity Agent
Source validation
Approved source enforcement, knowledge freshness, conflicting information detection
Banking Ops Agent
Operational audit
Process throughput, SLA adherence, bottleneck identification
Typical Weaknesses
AI answers sourced from outdated knowledge bases without version control
No confidence threshold enforcement — AI gives answers even when uncertain
Audit trails incomplete — some decisions logged, others not
Customer escalation paths bypass AI governance entirely
Type 2: Bank with Employee Agents
Deploying agents as internal employee assistants for compliance, research, and workflow
Areas to Investigate
Employee agent scope: which tasks do agents handle, which are restricted
Policy compliance: are agents constrained by the same rules as employees
Human oversight cadence: how often and by whom are agent outputs reviewed
ROI measurement: is agent deployment justified by measurable efficiency gains
Matched Agents
Agent
Role
Focus Area
Policy & Governance Agent
Constraint enforcement
Employee-agent policy parity, access control, data handling rules