Skip to content

Control Selection Guide

How to select the right controls for your AI system based on risk tier and use case characteristics.


Overview

Not every AI system needs every control. This guide helps you select appropriate controls based on:

  1. Risk tier — CRITICAL, HIGH, MEDIUM, LOW
  2. Use case type — Customer-facing, internal, agentic, batch
  3. Data sensitivity — PII, financial, regulated, public
  4. Decision impact — Consequential, advisory, informational

Step 1: Determine Risk Tier

Risk Tier Decision Tree

Risk Tier Decision Tree

Risk Tier Examples

System Tier Rationale
Credit decision support CRITICAL Directly affects lending decisions
Fraud detection AI CRITICAL False negatives allow fraud; false positives block customers
Trading signal generator CRITICAL Financial impact, regulatory scrutiny
Customer service chatbot HIGH Customer-facing, handles account queries
Document extraction (PII) HIGH Processes sensitive data at scale
Internal HR assistant HIGH Employment-related, PII
Meeting summariser MEDIUM Internal, limited sensitivity
Code assistant (internal) MEDIUM Internal productivity, no customer data
Marketing copy generator MEDIUM External-facing content, but reviewed
Sandbox experiments LOW No production data, no customer impact
Internal POC LOW Limited scope, controlled access

Step 2: Identify Use Case Characteristics

Use Case Type Matrix

Characteristic Control Implications
Customer-facing Output guardrails critical; Judge sampling higher; HITL escalation paths
Internal-only Can tolerate more latency; focus on data protection
Agentic Full AG.1-AG.4 controls; circuit breakers mandatory
Batch processing Can use heavier Judge evaluation; lower latency requirements
Real-time Guardrail latency budget critical; async Judge only
Decision support HITL mandatory for final decision; AI advisory only
Fully automated Higher scrutiny; regulatory constraints (GDPR Art 22)

Data Sensitivity Matrix

Data Type Additional Controls
PII AI.5.3 Privacy, output PII filtering, data minimisation
Financial AI.11 enhanced logging, AI.9 HITL for decisions
Health HIPAA/regulatory compliance, AI.5.1 classification
Credit SR 11-7 / SS1/23 model risk, AI.6.2 validation
Authentication Never in AI context; AI.7.1 input filtering
Regulated AI.13 vendor due diligence, AI.3.2 documentation

Step 3: Select Controls by Tier

Control Selection Matrix

Control CRITICAL HIGH MEDIUM LOW
AI.1 Governance
AI.1.1 Policy framework ✅ Required ✅ Required ✅ Required ⚠️ Recommended
AI.1.2 Governance structure ✅ Required ✅ Required ⚠️ Recommended ○ Optional
AI.1.3 Accountability ✅ Required ✅ Required ✅ Required ⚠️ Recommended
AI.2 Risk Management
AI.2.1 Risk classification ✅ Required ✅ Required ✅ Required ✅ Required
AI.2.2 Risk assessment Full assessment Full assessment Streamlined Self-assessment
AI.2.3 Ongoing monitoring Continuous Continuous Periodic Spot checks
AI.3 Inventory & Documentation
AI.3.1 System inventory ✅ Required ✅ Required ✅ Required ✅ Required
AI.3.2 System documentation Comprehensive Comprehensive Standard Basic
AI.3.3 Data lineage Full Full Key flows Basic
AI.3.4 Explainability Full audit trail Key factors General approach Basic
AI.4 Development Security
AI.4.1 Secure development ✅ Required ✅ Required ✅ Required ⚠️ Recommended
AI.4.2 Testing Statistical (≥50 runs) Statistical (≥20 runs) Statistical (≥10 runs) Basic (≥5 runs)
AI.4.3 Pre-deployment review Independent + committee Security team Streamlined Self-assessment
AI.5 Data Governance
AI.5.1 Data classification ✅ Required ✅ Required ✅ Required ⚠️ Recommended
AI.5.2 Data quality Continuous validation Continuous validation Periodic validation Basic checks
AI.5.3 Privacy protection Full PIA, minimisation PIA, minimisation Standard handling Basic
AI.5.4 RAG content integrity Full validation Full validation Validation Basic monitoring
AI.6 Model Security
AI.6.1 Model protection ✅ Required ✅ Required ✅ Required ⚠️ Recommended
AI.6.2 Model validation Independent (SR 11-7) Internal validation Functional testing Basic testing
AI.6.3 Model monitoring Continuous + trends Continuous Periodic Basic
AI.6.4 Capability assessment On every change On every change On major changes Initial only
AI.6.5 Baseline comparison Daily Weekly Fortnightly Monthly
AI.7 Guardrails
AI.7.1 Input guardrails ✅ Required ✅ Required ✅ Required ✅ Required
AI.7.2 Output guardrails ✅ Required ✅ Required ✅ Required ⚠️ Recommended
AI.7.3 Guardrail maintenance Monthly adversarial test Quarterly Biannually Annually
AI.7.4 Context isolation Dedicated instances Strict isolation Session isolation Standard
AI.8 LLM-as-Judge
AI.8.1 Judge evaluation ✅ Required ✅ Required ✅ Required ⚠️ Recommended
AI.8.2 Sampling strategy 100% 20-50% 5-10% 1-5%
AI.8.3 Finding management 1-hour SLA (critical) 24-hour SLA 1-week SLA Monthly batch
AI.8.4 Judge governance ✅ Required ✅ Required ⚠️ Recommended ○ Optional
AI.8.5 Confidence calibration ✅ Required ✅ Required ⚠️ Recommended ○ Optional
AI.9 Human Oversight
AI.9.1 HITL All decisions All escalations + sampling Periodic + escalation Spot checks
AI.9.2 Escalation procedures ✅ Required ✅ Required ✅ Required ⚠️ Recommended
AI.9.3 Human override ✅ Required ✅ Required ✅ Required ⚠️ Recommended
AI.9.4 Accountability ✅ Required ✅ Required ✅ Required ✅ Required
AI.9.5 HITL effectiveness Weekly canaries Monthly canaries Quarterly canaries Biannual
AI.10 Agentic Controls
AI.10.1-10.6 Full AG.1-AG.4 Full AG.1-AG.4 AG.2 + AG.3 Basic AG.2
AI.11 Logging & Monitoring
AI.11.1 Logging Full, tamper-evident, 7yr Full, 3yr Metadata + sampled, 1yr Basic, 90 days
AI.11.2 Real-time monitoring ✅ Required ✅ Required ✅ Required ⚠️ Recommended
AI.11.3 Alerting Immediate Within 1 hour Daily Weekly
AI.12 Incident Response
AI.12.1 AI-specific playbooks ✅ Required ✅ Required ⚠️ Recommended ○ Optional
AI.12.2 Investigation capability ✅ Required ✅ Required ✅ Required ⚠️ Recommended
AI.12.3 Remediation ✅ Required ✅ Required ✅ Required ⚠️ Recommended
AI.12.4 Notification Regulatory + customers Regulatory Internal Basic
AI.13 Supplier Management
AI.13.1 Vendor assessment Full + training data Full Standard Basic
AI.13.2 Vendor agreements Full AI terms Full AI terms Standard Basic
AI.13.3 Model provenance Full documentation Full documentation Standard Basic
AI.13.4 Training data risk Full assessment Full assessment Streamlined Basic
AI.14 Security Awareness
AI.14.1 Training All audiences + HITL bias All audiences Key personnel Basic
AI.15 Continuity
AI.15.1 Continuity planning ✅ Required ✅ Required ⚠️ Recommended ○ Optional
AI.15.2 System resilience Multi-region, failover Failover Graceful degradation Basic
AI.16 Intellectual Property
AI.16.1 Model IP protection ✅ Required ✅ Required ⚠️ Recommended ○ Optional
AI.16.2 Third-party IP ✅ Required ✅ Required ✅ Required ⚠️ Recommended

Step 4: Apply Use Case Modifiers

Agentic AI Modifier

If your system is agentic (takes actions, not just generates content), add:

Control All Agentic Systems
AG.1.1 Plan disclosure ✅ Required
AG.1.2 Plan guardrails ✅ Required
AG.1.3 Plan approval Based on tier
AG.2.1 Action guardrails ✅ Required
AG.2.2 Circuit breakers ✅ Required
AG.2.3 Scope enforcement ✅ Required
AG.2.4 Tool controls ✅ Required
AG.2.5 Tool protocol security ✅ Required
AG.3.1 Trajectory logging ✅ Required
AG.3.2 Trajectory evaluation Based on tier
AG.3.3 Agentic HITL Based on tier
AG.4.1 Agent inventory ✅ Required
AG.4.2 Orchestration controls If multi-agent
AG.4.3 Trace correlation ✅ Required

Customer-Facing Modifier

If your system is customer-facing, add:

Enhancement Rationale
Increase Judge sampling by 1 tier Higher reputational risk
Output guardrails mandatory Customer protection
Escalation SLAs tightened Customer impact
HITL queue prioritisation Customer experience
Incident notification to customers Regulatory requirement

Regulated Data Modifier

If your system handles regulated data (credit, health, financial):

Enhancement Rationale
Full AI.6.2 validation (SR 11-7 style) Regulatory requirement
AI.3.4 Explainability at CRITICAL level Right to explanation
AI.9.1 HITL mandatory for decisions GDPR Art 22, fair lending
AI.11.1 Logging at CRITICAL level Audit trail
AI.12.4 Regulatory notification Reporting requirements

Batch Processing Modifier

If your system operates in batch mode (not real-time):

Adjustment Rationale
Guardrail latency budget relaxed No real-time requirement
Judge can evaluate 100% even at lower tiers Processing time available
HITL can review before results released Batch allows pre-release review
Rollback easier Can reprocess batch

Step 5: Document Control Selection

For each AI system, document:

Item Content
System name Unique identifier
Risk tier CRITICAL / HIGH / MEDIUM / LOW
Tier rationale Why this tier was selected
Use case type Customer-facing, internal, agentic, batch
Data sensitivity PII, financial, regulated, public
Decision impact Consequential, advisory, informational
Modifiers applied Which modifiers and why
Controls selected Full list with evidence requirements
Controls deferred Any controls not implemented and rationale
Review schedule When to reassess tier and controls

Common Patterns

Pattern 1: Customer Service Chatbot

Typical profile: - Tier: HIGH - Type: Customer-facing, real-time - Data: Account data (PII, financial) - Decision: Advisory (human makes final decision)

Control selection:

Control Area Selection
Guardrails Full input + output, <50ms budget
Judge 20-50% sampling, 24hr SLA
HITL Escalation path, not every interaction
Logging Full content, 3-year retention
Agentic Not applicable (reactive, not agentic)

Pattern 2: Credit Decision Support

Typical profile: - Tier: CRITICAL - Type: Internal, decision support - Data: Credit data, PII - Decision: Consequential (affects lending)

Control selection:

Control Area Selection
Guardrails Full input + output + grounding verification
Judge 100% sampling, 2hr SLA for critical findings
HITL Human decides all — AI is advisory only
Logging Full, tamper-evident, 7-year retention
Validation Independent validation per SR 11-7
Explainability Full audit trail, decision rationale

Pattern 3: Internal Document Assistant

Typical profile: - Tier: MEDIUM - Type: Internal, real-time - Data: Internal documents (some confidential) - Decision: Informational

Control selection:

Control Area Selection
Guardrails Standard input + output
Judge 5-10% sampling, weekly SLA
HITL Periodic review, standard escalation
Logging Metadata + sampled content, 1-year retention
Agentic Not applicable

Pattern 4: Agentic Research Assistant

Typical profile: - Tier: HIGH (or CRITICAL if external actions) - Type: Internal, agentic - Data: Various (depends on tools) - Decision: Advisory with autonomous actions

Control selection:

Control Area Selection
Guardrails Full input + output + action guardrails
Judge 20-50% trajectory evaluation
HITL Plan approval for high-risk actions
Circuit breakers Step limits, time limits, cost limits
Scope enforcement Tool allowlist, data scope, outcome boundaries
Logging Full trajectory, 3-year retention

Pattern 5: Sandbox/POC

Typical profile: - Tier: LOW - Type: Internal, experimental - Data: Synthetic or public only - Decision: None (experimentation)

Control selection:

Control Area Selection
Guardrails Basic input validation
Judge Spot checks only
HITL Not required
Logging Basic metadata, 90-day retention
Isolation Separate from production

Review and Reassessment

Triggers for Reassessment

Trigger Action
Use case scope expands Reassess tier
New data types added Reassess data sensitivity
Customer-facing deployment Likely tier increase
Model upgrade Capability assessment
Regulatory change Control alignment review
Incident occurs Post-incident control review
Annual review Full reassessment

Review Schedule

Tier Full Reassessment Control Verification
CRITICAL Quarterly Monthly
HIGH Biannually Quarterly
MEDIUM Annually Biannually
LOW Annually Annually

Checklist: Control Selection Process

Step Complete
1. System identified and named
2. Risk tier determined using decision tree
3. Tier rationale documented
4. Use case type identified
5. Data sensitivity assessed
6. Decision impact classified
7. Base controls selected from matrix
8. Modifiers applied (agentic, customer-facing, regulated, batch)
9. Control selection documented
10. Deferred controls justified
11. Review schedule established
12. Sign-off obtained

AI Runtime Behaviour Security, 2026 (Jonathan Gill).