Human-in-the-Loop (HITL) Trigger Definitions

Status: January 2026 - FSI-AgentGov v1.2 Related Controls: 2.12 (Supervision), 2.17 (Multi-Agent Orchestration), 2.6 (Model Risk)

Purpose

This specification defines when AI agent actions require human review or approval before execution. HITL triggers ensure appropriate supervision for high-impact decisions and maintain regulatory compliance for financial services organizations.

Industry Context

This specification aligns with the Agentic Oversight Framework published by Sardine for regulated financial services. The framework defines five components for autonomous AI governance:

Component	Description	FSI-AgentGov Coverage
Access	Constrained data and tool access	AAM sections 3-6, DLP policies
Agent Operation	Defined tasks, tool invocation, reasoning	AAM sections 2-5, Control 1.14
Decision and Presentation	Agent recommends, human approves	HITL triggers (this document)
Comprehensive Audit Trail	All data, actions, decisions logged	Control 1.7, AAM section 9
Board Governance	Integration within GRC framework	Control 2.12, escalation matrix

Reference: Sardine Agentic Oversight Framework Whitepaper

FSI-AgentGov Interpretation Layer

The mapping above reflects FSI-AgentGov's interpretation of Sardine's Agentic Oversight Framework applied to Microsoft 365 Copilot Studio agents. The original Sardine whitepaper addresses general agentic AI governance principles; the specific control mappings (AAM, HITL triggers, etc.) are FSI-AgentGov framework extensions designed for Microsoft Power Platform implementations.

AI Agent Autonomy Levels (FINRA 2026 Report)

The FINRA 2026 Annual Regulatory Oversight Report discusses autonomy levels that may inform appropriate HITL patterns:

Citation Verification Required

The autonomy level framework below (Assisted/Augmented/Automated/Autonomous) is an FSI-AgentGov interpretation applied to Microsoft 365 agent governance. Organizations should independently verify the specific FINRA publication and confirm the terminology aligns with their regulatory obligations.

Autonomy Level	Description	Recommended HITL Pattern	Zone
Assisted	AI suggests, human decides	Sampled Review	Zone 1-2
Augmented	AI recommends, human approves	Pre-Approval for material actions	Zone 2-3
Automated	AI executes pre-approved actions	Escalation-on-Threshold	Zone 3 restricted
Autonomous	AI self-directs	Not Recommended	Avoid for Zone 3

Autonomous Agents and Financial Services

FINRA expects human oversight of AI-assisted customer interactions. Fully autonomous agents without pre-approval or escalation mechanisms are not appropriate for Zone 3 customer-facing deployments.

Mapping Autonomy to HITL Triggers

# Autonomy-based HITL Configuration
autonomy_hitl_mapping:
  assisted:
    pattern: "sampled_review"
    sample_rate: 0.05  # 5%
    real_time_approval: false

  augmented:
    pattern: "pre_approval"
    approval_required_for:
      - "recommendations"
      - "account_changes"
      - "financial_transactions"
    sample_rate: 0.10  # 10% for non-approval items

  automated:
    pattern: "escalation_on_threshold"
    thresholds:
      confidence_minimum: 0.85
      transaction_maximum: 1000
      escalation_mandatory: true
    post_audit_rate: 1.0  # 100% post-execution review

  autonomous:
    pattern: "not_recommended"
    warning: "Zone 3 customer-facing agents should not operate autonomously"

HITL Pattern Definitions

Three primary patterns for human oversight in AI agent workflows:

Pattern	Description	Timing	Use Case
Pre-Approval	Human must approve before agent action executes	Before action	High-risk actions, Zone 3 agents, regulatory decisions
Sampled Review	Post-hoc spot checks on agent decisions	After action	Quality assurance, compliance monitoring, audit sampling
Escalation-on-Threshold	Automatic human handoff when thresholds exceeded	During action	Confidence scoring, risk limits, unusual activity

Pattern 1: Pre-Approval

Agent pauses and requests human approval before executing action.

Zone requirement: Mandatory for Zone 3 high-risk actions
Regulatory driver: FINRA 3110 supervision, Reg BI suitability
Implementation: Approval workflow, queue management, SLA tracking

Pattern 2: Sampled Review

Agent executes action; subset of actions flagged for post-hoc human review.

Zone requirement: Recommended for Zone 2, optional for Zone 1
Regulatory driver: FINRA 3110 reasonable supervision
Implementation: Random sampling, risk-based sampling, periodic audits

Pattern 3: Escalation-on-Threshold

Agent monitors confidence scores or risk indicators and escalates when thresholds are breached.

Zone requirement: Required for Zone 2-3 customer interactions
Regulatory driver: SR 11-7 model risk management
Implementation: Confidence scoring, anomaly detection, risk scoring

HITL Trigger Framework

Trigger Categories

Category	Description	Regulatory Driver
Mandatory	Always require human review	Regulatory requirement
Configurable	Organization sets threshold	Risk-based decision
Conditional	Context-dependent triggers	Business rules
Exception	Unusual situations	Anomaly detection

Mandatory HITL Triggers

These triggers ALWAYS require human review before agent action proceeds:

Financial Threshold Triggers

Trigger	Condition	Review SLA	Approver Role
Large Transaction	Transaction value > $25,000	30 minutes	Supervisor
Account Modification	Account ownership change	1 hour	Operations Manager
Credit Decision	Any credit/lending recommendation	15 minutes	Credit Officer
Investment Recommendation	Suitability-impacting advice	15 minutes	Registered Representative
Wire Transfer	Any wire transfer initiation	15 minutes	Treasury

Suitability Triggers

Trigger	Condition	Review SLA	Approver Role
Suitability Determination	Agent makes suitability assessment	Immediate	Registered Representative
Risk Profile Change	Customer risk profile modification	30 minutes	Supervisor
Product Recommendation	Investment product suggestion	15 minutes	Compliance
Retirement Account Action	IRA/401k transaction	30 minutes	Supervisor

Regulatory Triggers

Trigger	Condition	Review SLA	Approver Role
SAR Indicator	Suspicious activity detected	Immediate	BSA/AML Officer
KYC Verification	Identity verification decision	1 hour	Compliance
Regulatory Disclosure	Required disclosure delivery	15 minutes	Compliance
Customer Complaint	Complaint identified	1 hour	Compliance

Configurable HITL Triggers

Organizations configure these thresholds based on risk appetite:

Confidence Score Triggers

# Confidence-based HITL Configuration
confidence_triggers:
  low_confidence_threshold: 0.7
  very_low_confidence_threshold: 0.5

  responses:
    below_very_low:
      action: "block_and_escalate"
      message: "I'm not confident enough to answer this. Let me connect you with a specialist."
      route_to: "human_queue"
      sla_minutes: 15

    below_low:
      action: "human_review"
      message: "Let me verify this information before proceeding."
      route_to: "verification_queue"
      sla_minutes: 30

    above_low:
      action: "proceed"
      message: null
      route_to: null

Complexity Triggers

Complexity Indicator	Threshold	Action
Query word count	> 100 words	Route for review
Multi-topic query	> 3 distinct topics	Human assist
Follow-up depth	> 5 clarifying questions	Escalate
Session duration	> 30 minutes	Supervisor check

Sensitive Topic Triggers

# Sensitive Topic HITL Configuration
sensitive_topics:
  always_escalate:
    - "legal_advice"
    - "tax_advice"
    - "estate_planning"
    - "regulatory_complaint"
    - "discrimination"
    - "fraud_allegation"

  require_confirmation:
    - "account_closure"
    - "beneficiary_change"
    - "address_change"
    - "fee_waiver"

  soft_escalate:
    - "competitor_comparison"
    - "rate_negotiation"
    - "product_complaint"

Conditional HITL Triggers

Context-dependent triggers based on business rules:

Customer Segment Triggers

Customer Segment	Additional Triggers	Rationale
High Net Worth	All recommendations	Enhanced service
Senior Customer	Product changes	Elder protection
New Customer	First 90 days of activity	Onboarding protection
Complaint History	Any service interaction	Relationship management

Agent Capability Triggers

Agent Type	Trigger Condition	Review Type
Advisory Agent	Any recommendation	Real-time approval
Transactional Agent	Value thresholds exceeded	Transaction review
Informational Agent	Sensitive topic detected	Content review
Orchestrating Agent	Depth limit reached	Chain review

Time-Based Triggers

Condition	Trigger	Rationale
After hours (6pm-8am)	All transactions	Fraud prevention
Weekend	Account modifications	Reduced staffing
Holiday	Large transactions	Enhanced review
End of quarter	Advisory interactions	Sales pressure mitigation

HITL Response Types

Response Actions

Action	Description	Use Case
Block	Prevent agent action entirely	High-risk situations
Pause	Hold until human approves	Approval workflows
Review	Allow action, flag for review	Audit sampling
Assist	Bring human into conversation	Complex queries
Escalate	Transfer to specialist	Subject matter expertise
Log	Proceed, enhanced logging	Monitoring only

Response Flow Configuration

# HITL Response Flow
hitl_response_flow:
  trigger_activated:
    - log_trigger_details
    - capture_conversation_context
    - determine_response_action

  block_action:
    - notify_user_politely
    - route_to_human_queue
    - set_priority_based_on_trigger
    - start_sla_timer
    - alert_if_sla_breach

  pause_action:
    - notify_user_of_wait
    - present_approval_request
    - capture_approver_decision
    - resume_or_terminate_based_on_decision
    - log_approval_with_justification

  assist_action:
    - add_human_to_conversation
    - provide_context_to_human
    - human_can_guide_or_takeover
    - log_assistance_interaction

SLA Definitions

Review SLA Tiers

Tier	Target Time	Max Time	Escalation
Immediate	5 minutes	15 minutes	Auto-escalate to manager
Urgent	15 minutes	30 minutes	Alert supervisor
Standard	30 minutes	1 hour	Queue management
Normal	1 hour	4 hours	Standard review

SLA Breach Handling

# SLA Breach Configuration
sla_breach_handling:
  warning_at: 80%  # of SLA time
  breach_at: 100%  # of SLA time

  warning_actions:
    - notify_assigned_reviewer
    - add_to_supervisor_dashboard

  breach_actions:
    - escalate_to_backup_reviewer
    - notify_supervisor
    - log_sla_breach
    - update_compliance_metrics

  repeated_breach:
    threshold: 3  # breaches per day
    actions:
      - alert_operations_manager
      - review_staffing_levels
      - consider_temporary_agent_restriction

Implementation Guidance

Queue Management

# HITL Queue Configuration
hitl_queues:
  immediate_review:
    routing: "round_robin"
    backup_after: "5_minutes"
    max_queue_depth: 10
    overflow_action: "escalate_all"

  standard_review:
    routing: "skill_based"
    backup_after: "30_minutes"
    max_queue_depth: 50
    overflow_action: "extend_sla"

  audit_sample:
    routing: "random_assignment"
    sample_rate: "10%"
    priority: "low"
    max_queue_depth: 100

Approver Availability

# Approver Availability Configuration
approver_availability:
  business_hours:
    start: "08:00"
    end: "18:00"
    timezone: "America/New_York"

  after_hours:
    on_call_rotation: true
    escalation_path:
      - "on_call_supervisor"
      - "operations_manager"
      - "security_operations"

  fallback:
    action: "block_and_notify"
    message: "This request requires human approval. Please try again during business hours or contact support."

Metrics and Reporting

Metric	Calculation	Target
HITL Trigger Rate	Triggers / Total Interactions	<10%
Approval Rate	Approved / Total Reviewed	>90%
Average Review Time	Mean time to decision	<SLA
SLA Compliance	Reviews within SLA / Total	>95%
False Positive Rate	Unnecessary triggers / Total	<5%

Zone-Specific Configuration

Zone 1 (Personal Productivity)

zone_1_hitl:
  enabled_triggers:
    - "sensitive_topic_detection"  # Minimal
  disabled_triggers:
    - "confidence_score"  # Too noisy for personal use
    - "financial_threshold"  # Not applicable
  rationale: "Minimal friction for low-risk personal agents"

Zone 2 (Team Collaboration)

zone_2_hitl:
  enabled_triggers:
    - "confidence_score"
    - "sensitive_topic"
    - "customer_segment"
  thresholds:
    confidence: 0.6  # Moderate
    financial: 10000  # Team appropriate
  rationale: "Balanced controls for shared agents"

Zone 3 (Enterprise Managed)

zone_3_hitl:
  enabled_triggers:
    - "all_mandatory"
    - "all_configurable"
    - "conditional_by_segment"
  thresholds:
    confidence: 0.7  # Strict
    financial: 5000  # Conservative
  additional:
    - "audit_sample_all_interactions"
    - "supervisor_spot_check_enabled"
  rationale: "Comprehensive oversight for customer-facing agents"

Integration Points

Control	Integration
2.12 Supervision	Supervisory review procedures
2.17 Orchestration	Multi-agent HITL checkpoints
3.4 Incident Reporting	HITL rejections as incidents
2.6 Model Risk	MRM oversight requirements

FSI Agent Governance Framework v1.2.51 - January 2026

Specification	Relationship
Confidence and Routing	Defines the confidence scoring model and routing thresholds that determine when HITL triggers activate
Zone 1 Minimum Explainability	Specifies how HITL decisions and confidence scores are communicated to end users