Human Skill Atrophy Monitoring Governance

2. Summary

Human Skill Atrophy Monitoring Governance requires that organisations detect and mitigate the degradation of human operator competencies that occurs when AI agents progressively absorb tasks previously performed by humans. The dimension mandates ongoing measurement of operator skill levels, detection of atrophy trends, and intervention before operators lose the ability to perform critical functions during agent failures, overrides, or fallback scenarios. When an AI agent handles 95% of credit decisions for two years, the human credit analysts who once made those decisions lose their ability to make them competently — and that loss is invisible until the agent fails and the humans cannot compensate.

3. Example

Scenario A — Pilot Skill Atrophy in Automated Flight: An airline adopts advanced autopilot systems that handle 97% of flight operations. Over three years, pilots' manual flying hours decline from an average of 120 hours per year to 8 hours per year. During an autopilot failure at 35,000 feet in turbulent conditions, the flight crew must hand-fly the aircraft. The captain's response time to establish manual control is 23 seconds — 3x longer than the pre-automation baseline of 7 seconds. The first officer inputs incorrect pitch corrections, demonstrating degraded instrument scanning patterns. The aircraft loses 4,000 feet of altitude before stable manual flight is established. Post-incident analysis reveals that the airline tracked automated vs. manual flying hours but had no threshold below which manual flying proficiency was considered degraded, no regular proficiency assessment, and no mandatory manual flying practice programme.

What went wrong: The airline allowed manual flying skills to atrophy without measurement or intervention. The automation operated reliably enough that the skill gap was never exposed — until it was. No periodic skill assessment detected the degradation. No minimum manual practice requirement prevented it. The safety case assumed pilots could always take over from automation, but the assumption was never validated after automation was introduced. Consequence: Serious incident investigation, mandatory remediation of pilot training programmes, and fleet-wide reassessment of automation dependency.

Scenario B — Credit Analyst Deskilling in Automated Underwriting: A consumer lending company deploys an AI agent to handle credit decisioning. Over 18 months, the agent handles 94% of applications autonomously, with human analysts reviewing only the 6% the agent escalates. Before automation, analysts processed 40 full applications per day, developing and maintaining deep expertise in financial statement analysis, fraud indicator detection, and risk assessment. After 18 months, analysts process 2-3 escalated applications per day, all pre-screened by the agent. An external audit introduces 20 synthetic applications with subtle fraud indicators into the analyst workflow. Before automation, analysts detected 85% of similar indicators. After 18 months of reduced practice, the detection rate drops to 41%. The company discovers that its human fallback capability has degraded by more than half without anyone measuring it.

What went wrong: No skill atrophy monitoring existed. The company assumed that analysts who were "still doing credit analysis" were maintaining their skills, failing to recognise that reviewing 2-3 pre-screened applications per day exercises a fundamentally different skill set than processing 40 full applications. No periodic competency assessment measured analyst capabilities. No minimum practice requirement ensured skill retention. Consequence: Degraded human fallback capability, audit finding for inadequate business continuity, and £3.2 million in undetected fraudulent loans discovered in the following quarter when the AI system was taken offline for model retraining and analysts had to process the full pipeline.

Scenario C — Diagnostic Skill Erosion in AI-Assisted Medicine: A dermatology practice adopts an AI diagnostic agent for skin lesion classification. Over two years, dermatologists rely increasingly on the agent's classifications, performing independent assessments only when the agent flags uncertainty. A junior dermatologist who joined the practice after the AI was deployed has never independently classified more than 50 lesions — compared to the 3,000-5,000 independent classifications that dermatologists historically performed during their first two years. During a system outage lasting 3 days, the junior dermatologist misclassifies 4 melanomas as benign naevi. Two patients experience delayed treatment. The practice's medical director recognises that the junior dermatologist's clinical skills never fully developed because the AI absorbed the volume of cases needed for experiential learning.

What went wrong: No monitoring distinguished between "working alongside the AI" and "developing independent diagnostic competence." The practice tracked throughput (patients seen) but not independent skill development (diagnoses made without AI assistance). No minimum independent practice requirement existed for clinicians working with AI support. The AI created an experiential learning deficit that went undetected. Consequence: Two delayed melanoma diagnoses, GMC referral for the practice, malpractice claims, and mandated remediation of training programmes.

4. Requirement Statement

Scope: This dimension applies to all AI agent deployments where the agent performs or assists with tasks that humans must be capable of performing independently during agent failure, fallback, override, or degraded operation scenarios. The scope includes any situation where human competence is part of the business continuity plan, the safety case, or the regulatory compliance posture. It applies to both experienced operators whose skills may atrophy from disuse and new operators who may never fully develop skills because the AI absorbs the experiential learning opportunities. The scope extends beyond technical task performance to include situational awareness, judgement under uncertainty, and domain knowledge that require ongoing practice to maintain. If the organisation's safety case, business continuity plan, or regulatory compliance depends on humans being able to perform a function currently handled by AI — those humans' ability to perform that function must be measured.

4.1. A conforming system MUST identify and document all critical human competencies that the AI agent's operation displaces or reduces in practice frequency, creating a skill dependency register that maps each agent capability to the corresponding human competency required for fallback operation.

4.2. A conforming system MUST implement periodic competency assessments for each critical human competency at intervals no greater than 6 months, measuring operator performance against defined proficiency standards using practical assessments — not self-assessment or knowledge-only tests.

4.3. A conforming system MUST define minimum proficiency thresholds for each critical human competency and trigger a remediation intervention when any operator's assessed proficiency falls below the threshold.

4.4. A conforming system MUST implement minimum practice requirements — structured opportunities for operators to perform the displaced task independently of the AI agent — at a frequency and volume sufficient to maintain proficiency as validated by the competency assessments.

4.5. A conforming system MUST log all competency assessments, proficiency scores, threshold breaches, and remediation actions with timestamps and operator identifiers.

4.6. A conforming system SHOULD track leading indicators of skill atrophy — declining override quality, increasing error rates on escalated items, longer response times on manual tasks — as early-warning signals before formal competency assessments detect degradation.

4.7. A conforming system SHOULD differentiate atrophy monitoring between experienced operators (who had skills before automation and may lose them) and new operators (who may never develop skills due to insufficient experiential learning).

4.8. A conforming system SHOULD integrate skill atrophy data into business continuity planning, adjusting fallback capacity estimates based on measured (not assumed) human proficiency.

4.9. A conforming system MAY implement graduated automation levels that maintain minimum human involvement thresholds to preserve skills — for example, requiring that humans independently process at least 10% of the task volume without AI assistance.

5. Rationale

Skill atrophy from automation is a well-documented phenomenon across all industries where automation has displaced human task performance. Endsley's research on situation awareness (1995) established that operators who are removed from the active control loop lose not just procedural skills but situational awareness — the ability to perceive, comprehend, and predict system states. Bainbridge's seminal "Ironies of Automation" paper (1983) identified the fundamental paradox: automation is introduced because human performance is inadequate, but the humans who must take over when automation fails are the same humans whose skills have atrophied from disuse — making them less capable at the moment they are most needed.

The AI agent context intensifies the skill atrophy risk because AI agents can absorb a broader range of cognitive tasks than traditional automation. A robotic process automation system might handle data entry, but a human can still perform data entry if needed — the skill is procedural and retained relatively easily. An AI agent that handles credit analysis, medical diagnosis, or security threat assessment absorbs cognitive skills that require deep domain knowledge, pattern recognition, and judgement under uncertainty. These skills degrade rapidly without practice. Ericsson's research on deliberate practice (1993) demonstrates that expert-level cognitive performance requires ongoing practice; without it, performance declines to novice levels within 12-24 months for complex cognitive tasks.

AG-106 is the detective counterpart to the preventive controls in AG-104 and AG-105. Where AG-104 ensures operators are appropriately calibrated and AG-105 ensures they have sufficient cognitive bandwidth, AG-106 ensures they retain the underlying competency to exercise meaningful oversight. A perfectly calibrated operator with adequate bandwidth who has lost their domain expertise cannot provide effective oversight — they can recognise that the AI might be wrong (calibrated trust) and they have time to investigate (manageable workload), but they lack the skill to determine whether the AI is actually wrong.

The regulatory imperative is clear across domains. Financial regulators expect firms to demonstrate that human operators can perform critical functions during system failures (FCA operational resilience requirements, DORA Article 11). Healthcare regulators expect clinicians to maintain independent clinical competence regardless of AI assistance (GMC Good Medical Practice, CQC fundamental standards). Aviation regulators mandate minimum manual flying hours for pilots of automated aircraft (EASA requirements, FAA Advisory Circulars). AG-106 provides the governance framework that operationalises these regulatory expectations for AI agent deployments.

6. Implementation Guidance

Skill atrophy monitoring requires three capabilities: identifying which skills are at risk, measuring whether they are degrading, and intervening before degradation reaches a critical level. The implementation must treat skill maintenance as an ongoing operational cost of automation, not an optional training activity.

Recommended patterns:

• Skill dependency register. For each AI agent capability, document the corresponding human competency required for fallback. Structure the register as: Agent Capability → Human Competency → Proficiency Standard → Assessment Method → Assessment Frequency → Minimum Practice Requirement. For example: AI credit decisioning → Human financial statement analysis → Correctly assess 80% of test applications within 15 minutes each → Quarterly practical assessment using 10 synthetic applications → Every 6 months → Independently process minimum 20 real applications per month without AI assistance.
• Practical competency assessments. Design assessments that test the actual skill, not knowledge about the skill. For credit analysts: present 10 synthetic applications with a controlled mix of approve, decline, and fraud indicators, and measure decision accuracy and time. For clinicians: present diagnostic cases without AI assistance and compare to reference diagnoses. For trading oversight: present market scenarios requiring manual position management. Assessment design should be validated by subject matter experts and calibrated against pre-automation performance baselines. A proficiency score should be computed as a percentage of baseline performance: an analyst who scored 85% accuracy pre-automation and now scores 52% has a proficiency score of 61% of baseline.
• Minimum practice requirements. Establish mandatory periods where operators perform the displaced task without AI assistance. The frequency and volume should be sufficient to maintain proficiency as validated by competency assessments — if quarterly assessments show degradation despite monthly practice sessions of 10 cases, increase to 20 cases. Common approaches include: dedicated manual processing days (e.g., every second Friday, the AI is placed in shadow mode and operators process independently), mandatory manual review quotas (e.g., each analyst must independently process at least 20 applications per month without AI pre-screening), and simulation exercises (e.g., quarterly tabletop exercises simulating a 72-hour agent outage).
• Leading indicator monitoring. Track operational metrics that signal skill degradation before formal assessments detect it: declining quality of override decisions (operators override the AI but their alternative decisions are worse), increasing error rates on the small volume of manually processed items, increasing time to complete manual tasks, and decreasing confidence in manual decisions (measured through post-decision confidence ratings or decision reversal rates). When leading indicators cross a threshold — for example, override decision quality drops below 70% accuracy — trigger an accelerated competency assessment rather than waiting for the scheduled cycle.
• New operator skill development tracking. For operators who join after AI deployment, track experiential learning milestones separately. Define the minimum number of independent decisions required for competency development (e.g., 500 independent credit decisions within the first 12 months). Ensure the AI deployment does not absorb so many cases that new operators cannot reach these milestones. If the AI handles 94% of cases and the remaining 6% is insufficient for skill development, create structured learning rotations where new operators process a curated set of cases independently.

Anti-patterns to avoid:

• Assuming presence equals competence. An analyst who has been "doing credit analysis for 10 years" but has spent the last 2 years reviewing 3 AI-escalated cases per day is not maintaining 10 years of expertise. Tenure is not a proxy for current proficiency when the task has fundamentally changed.
• Knowledge-only assessments. Multiple-choice tests on credit analysis principles do not measure the ability to analyse a financial statement under time pressure. Skill atrophy affects procedural and cognitive performance, not declarative knowledge — operators can describe how to do the task long after they have lost the ability to do it competently.
• Voluntary practice programmes. Making manual practice optional means the operators most susceptible to atrophy (those who trust the AI most and see practice as unnecessary) are the least likely to participate. Practice requirements must be mandatory and tracked.
• Measuring only task completion. An analyst who completes a manual assessment in the correct time but with 40% accuracy has not demonstrated proficiency. Assessments must measure quality, not just throughput.
• Ignoring the cost of skill maintenance. Mandatory manual practice reduces the operational efficiency that AI automation provides. Organisations must budget for this cost explicitly — typically 5-15% of operator time dedicated to skill maintenance activities — rather than treating it as overhead that can be eliminated during cost optimisation. The cost of skill maintenance is an insurance premium against the cost of agent failure when humans cannot compensate.

Industry Considerations

Financial Services. FCA operational resilience requirements (PS21/3) mandate that firms can continue to deliver important business services during severe but plausible scenarios, including technology failures. If the important business service depends on AI agents, the fallback to human processing must be viable — which requires that humans retain the skills to perform it. Skill atrophy monitoring directly supports the firm's operational resilience self-assessment. Competency assessments should align with existing T&C (Training and Competence) schemes under SYSC 5.1.

Healthcare. GMC Good Medical Practice requires doctors to maintain and develop their knowledge and skills. Where AI absorbs clinical tasks, clinicians must demonstrate continued competence in those tasks through validated assessments. The Royal Colleges' CPD (Continuing Professional Development) frameworks should incorporate AI-displacement skill maintenance. CQC inspections increasingly examine whether AI-assisted care maintains clinician competence to operate independently.

Aviation. EASA and FAA requirements for pilot proficiency in automated aircraft provide the most mature model for skill atrophy governance. EASA SIB 2013-05 recommends that operators actively promote and monitor manual flying skills. The airline industry's approach — mandatory simulator sessions, minimum manual flying hours, and recurrent proficiency checks — is directly applicable to other domains deploying AI agents.

Critical Infrastructure. IEC 61511 and ISA 84 require that safety instrumented system operators maintain competence in manual operations. AI agents in process control must not displace the operator skills needed for emergency manual intervention. Training and competency requirements should reference ANSI/ISA-84.00.01, which mandates documented competency requirements and periodic assessment.

Maturity Model

Basic Implementation — The organisation has created a skill dependency register mapping AI agent capabilities to required human competencies. Competency assessments are conducted at least every 6 months using practical assessments. Minimum proficiency thresholds are defined. Operators below threshold are flagged for remediation. Assessment results are logged. This level meets the minimum mandatory requirements but relies on periodic assessments as the primary detection mechanism, which may detect atrophy only after significant degradation has occurred.

Intermediate Implementation — All basic capabilities plus: leading indicators of skill atrophy are monitored continuously (override quality, manual task error rates, response times). Minimum practice requirements are mandatory and tracked — operators must complete a defined volume of independent task performance per month. New operator skill development is tracked separately with experiential learning milestones. Skill atrophy data is integrated into business continuity planning, with fallback capacity estimates based on measured proficiency rather than assumptions. Competency assessment design is validated by subject matter experts and calibrated against pre-automation baselines.

Advanced Implementation — All intermediate capabilities plus: predictive models identify operators trending toward atrophy before thresholds are breached. Graduated automation levels maintain minimum human involvement to preserve skills (e.g., 10% of task volume processed manually). Skill atrophy trends are reported to governance committees quarterly. Independent assessors validate competency assessment design and scoring annually. The organisation can demonstrate to regulators that human fallback capability is measured, maintained, and viable at current proficiency levels — with quantitative evidence.

7. Evidence Requirements

Required artefacts:

• Skill dependency register. The documented mapping of AI agent capabilities to required human competencies, including proficiency standards, assessment methods, assessment frequencies, and minimum practice requirements. Format: structured data (spreadsheet, database export, or equivalent).
• Competency assessment records. Per-operator assessment results with dates, assessment type, proficiency scores, threshold comparisons, and assessor identity. Must include the assessment materials used (e.g., synthetic test cases) and scoring methodology. Minimum 12 months retention.
• Practice requirement compliance records. Records demonstrating that operators completed their minimum practice requirements, including dates, volume of independent task performance, and performance quality metrics. Minimum 12 months retention.
• Remediation records. For operators below proficiency thresholds: the remediation plan, activities undertaken, and reassessment results demonstrating proficiency recovery. Minimum 12 months retention.

Retention requirements:

• Competency assessment records and practice compliance records: minimum 7 years for regulated financial services; minimum 5 years for healthcare; minimum 3 years otherwise.

Access requirements:

• Producible to regulators or auditors within 48 hours of request. Evidence must exist as retained artefacts, not be reconstructable after the fact.

8. Test Specification

Test 8.1: Skill Dependency Register Completeness

• Stimulus: Review the skill dependency register against the full inventory of deployed AI agent capabilities.
• Expected behaviour: Every AI agent capability that displaces a human function required for fallback has a corresponding entry in the skill dependency register.
• Pass criteria: No AI agent capability required for fallback operation is missing from the register. Each entry includes: competency description, proficiency standard, assessment method, assessment frequency, and minimum practice requirement.
• Fail criteria: Any fallback-critical AI agent capability lacks a corresponding skill dependency register entry, or any entry is missing mandatory fields.

Test 8.2: Competency Assessment Validity

• Stimulus: Administer the competency assessment to 10 operators with known pre-automation proficiency levels (established from historical records). Compare assessment scores to known proficiency.
• Expected behaviour: Assessment scores correlate with known proficiency levels — operators with historically higher proficiency score higher on the assessment, and operators with known atrophy indicators score lower.
• Pass criteria: Correlation between assessment scores and known proficiency is statistically significant (p < 0.05). The assessment distinguishes between proficient and degraded operators.
• Fail criteria: No significant correlation between assessment scores and known proficiency, indicating the assessment does not validly measure the target competency.

Test 8.3: Minimum Practice Requirement Enforcement

• Stimulus: Configure a minimum practice requirement of 20 independent cases per month. Simulate an operator who completes only 8 cases in a given month. Verify system response.
• Expected behaviour: The system detects the shortfall and generates a compliance alert to the operator and their supervisor.
• Pass criteria: The shortfall is detected by month-end. A compliance alert is generated with the operator's ID, the required volume (20), and the actual volume (8).
• Fail criteria: The shortfall is not detected, or no alert is generated.

Test 8.4: Proficiency Threshold Breach Response

• Stimulus: Record a competency assessment where an operator scores 55% against a proficiency threshold of 70%. Verify system response.
• Expected behaviour: The system flags the operator as below threshold, triggers a remediation requirement, and restricts the operator from unsupervised fallback duty until reassessment demonstrates recovery.
• Pass criteria: The breach is recorded, a remediation plan is required, and the operator's fallback eligibility status is updated to reflect the proficiency gap.
• Fail criteria: The breach is not flagged, no remediation is required, or the operator's fallback eligibility is not affected.

Test 8.5: Leading Indicator Detection

• Stimulus: Simulate an operator whose override decision quality degrades from 82% accuracy to 58% accuracy over 60 days.
• Expected behaviour: The system detects the degradation trend and triggers an accelerated competency assessment before the next scheduled assessment date.
• Pass criteria: The degradation trend is detected within 30 days of crossing the leading indicator threshold. An accelerated assessment is triggered.
• Fail criteria: The degradation is not detected until the next scheduled assessment, or no accelerated assessment is triggered.

Test 8.6: New Operator Skill Development Tracking

• Stimulus: Onboard a new operator. Verify the system tracks their independent decision count toward the experiential learning milestone (e.g., 500 independent decisions within 12 months).
• Expected behaviour: The system maintains a running count of the operator's independent decisions, displays progress toward the milestone, and alerts if the projected pace is insufficient to reach the milestone within the required timeframe.
• Pass criteria: The count is accurate (matches actual independent decisions), progress is displayed, and an insufficiency alert fires if the operator is on pace to reach only 300 of 500 required decisions by month 12.
• Fail criteria: Independent decisions are not counted, progress is not visible, or pace alerts do not function.

Test 8.7: Assessment Logging Completeness

• Stimulus: Conduct 20 competency assessments across 10 operators over 3 months. Query the assessment log for completeness.
• Expected behaviour: All 20 assessments have complete records including date, operator ID, assessment type, proficiency score, threshold comparison, and assessor identity.
• Pass criteria: 100% of assessments have complete records with all mandatory fields populated.
• Fail criteria: Any assessment lacks a complete record, or any mandatory field is null.

Conformance Scoring

• Score 0: No skill atrophy monitoring exists — the organisation assumes human operators retain competence without measurement.
• Score 1: A skill dependency register exists but competency assessments are not conducted — the organisation has identified the risk but does not measure it.
• Score 2: Competency assessments are conducted at defined intervals, proficiency thresholds are enforced, and minimum practice requirements are mandatory — the organisation actively detects and mitigates skill atrophy.
• Score 3: All Score 2 capabilities plus leading indicator monitoring, new operator skill development tracking, business continuity integration, and independent validation of assessment effectiveness — verified through external review or audit.

9. Regulatory Mapping

Regulation	Provision	Relationship Type
EU AI Act	Article 14 (Human Oversight)	Direct requirement
EU AI Act	Article 9 (Risk Management System)	Supports compliance
FCA PS21/3	Operational Resilience	Supports compliance
FCA SYSC	5.1 (Training & Competence)	Direct requirement
EASA SIB 2013-05	Manual Flying Skills	Supports compliance (aviation)
GMC Good Medical Practice	Domain 1 (Knowledge, Skills, Performance)	Direct requirement (healthcare)
NIST AI RMF	GOVERN 1.4, MANAGE 2.4	Supports compliance
ISO 42001	Clause 7.2 (Competence)	Supports compliance
DORA	Article 11 (Response and Recovery)	Supports compliance

EU AI Act — Article 14 (Human Oversight)

Article 14(1) requires human oversight measures that are "appropriate to the risks, the level of autonomy and the context of use of the high-risk AI system." Meaningful human oversight presupposes that the humans exercising oversight are competent to do so. AG-106 ensures that the competence assumption underlying Article 14 remains valid over time — if AI agents absorb the tasks that developed and maintained human competence, Article 14 compliance erodes silently. The European Commission's guidance on human oversight will likely address competence maintenance as AI deployment matures; AG-106 positions organisations ahead of this regulatory evolution.

FCA PS21/3 — Operational Resilience

The FCA's operational resilience framework requires firms to identify important business services, set impact tolerances, and demonstrate they can remain within those tolerances during severe but plausible disruption scenarios. If an important business service relies on AI agents, the firm must demonstrate a viable fallback — typically human processing. AG-106 provides the evidence that human fallback capability is measured and maintained, not assumed. A firm that cannot demonstrate measured human proficiency in fallback-critical functions has a gap in its operational resilience self-assessment.

FCA SYSC 5.1 — Training and Competence

SYSC 5.1.1R requires firms to employ personnel with the skills, knowledge, and expertise necessary for the discharge of responsibilities allocated to them. For personnel whose responsibilities include fallback operation when AI agents are unavailable, competence must be actively maintained and assessed — not assumed from historical capability. AG-106 extends T&C obligations to cover the specific risk of AI-driven skill atrophy.

DORA — Article 11 (Response and Recovery)

DORA Article 11 requires financial entities to establish ICT business continuity policies and disaster recovery plans. Where business continuity depends on human fallback from AI agent operations, Article 11 implicitly requires that human operators are capable of performing the fallback function. AG-106 provides the governance framework to validate that this capability exists at measured proficiency levels.

10. Failure Severity

Field	Value
Severity Rating	High
Blast Radius	Organisation-wide — affects the viability of human fallback for all AI-assisted operations and the credibility of business continuity plans

Consequence chain: Unmonitored skill atrophy creates a hidden dependency: the organisation depends on AI agent availability because human operators can no longer perform the displaced functions competently, but this dependency is undocumented and unrecognised until an agent failure forces the fallback. The immediate failure mode is a capability gap during an agent outage — humans cannot compensate at the quality, speed, or volume required. In the credit analysis example, a 2-week model retraining period required human processing that produced £3.2 million in undetected fraud. In healthcare, skill atrophy in AI-assisted diagnosis can produce patient harm during system outages. The secondary consequence is regulatory: organisations that claim human fallback capability in business continuity plans but cannot demonstrate measured human proficiency face regulatory action for inadequate resilience planning. The long-term consequence is organisational brittleness — the organisation becomes entirely dependent on AI agent availability for functions it believes it can perform manually, creating a single point of failure that compounds with every month of unmeasured atrophy.

Cross-references: AG-019 (Human Escalation & Override Triggers) requires human competence to evaluate escalations; AG-106 ensures that competence is maintained. AG-038 (Human Control Responsiveness) requires timely human response; AG-106 ensures the responding humans are competent. AG-104 (Trust Calibration Governance) manages trust alignment; AG-106 ensures the underlying skills that informed trust requires are preserved. AG-105 (Oversight Workload and Alarm Fatigue Governance) manages cognitive load; AG-106 manages the cognitive capability available to bear that load. AG-022 (Behavioural Drift Detection) detects agent risk changes; AG-106 ensures humans can evaluate whether those changes are problematic. AG-107 (Override Usability and Actionability Governance) ensures override mechanisms are usable; AG-106 ensures the operators using them are competent. AG-108 (Operator Role Segregation Governance) assigns roles; AG-106 ensures operators in each role maintain the competencies that role requires.