AI Governance and Control Training Course

Participants can identify the technical components that create AI governance and control requirements.

• Large language model components, prompts, context windows, and output generation
• Retrieval-augmented generation traced from source documents to generated responses
• Agent loops defined through goals, planning, tools, memory, and execution
• Tool calling risks created by APIs, databases, payment systems, and email
• Human-in-the-loop, human-on-the-loop, and human-in-command oversight models compared
• Architecture diagrams converted into control points, evidence sources, and owners

Participants can interpret major AI governance references and map them to internal policy requirements.

• ISO/IEC 42001 clauses mapped to management-system roles and documented information
• NIST AI RMF functions applied to govern, map, measure, and manage
• ISO/IEC 23894 risk guidance connected to existing enterprise risk processes
• EU AI Act obligations screened by role, use case, and risk category
• Data protection principles applied to training data, prompts, outputs, and retention
• Sector rules identified for financial decisions, public services, employment, and healthcare

Participants can create an AI inventory and classify use cases for proportionate governance treatment.

• AI inventory fields defined for owners, vendors, data, models, and decisions
• Use-case intake captures purpose, users, affected parties, and business criticality
• Risk classification considers autonomy, impact, scale, sensitivity, and reversibility
• Materiality thresholds separate experimentation, controlled deployment, and prohibited uses
• System boundaries record upstream data sources, downstream actions, and dependencies
• Inventory lifecycle updates capture changes, retraining, retirement, and transferred ownership

Participants can assess AI risks using structured impact, likelihood and control-effectiveness criteria.

• Risk scenarios written with cause, event, consequence, owner, and control
• Impact assessments cover customers, employees, citizens, markets, and organisational operations
• Likelihood ratings account for model uncertainty, exposure, misuse, and attack paths
• Inherent and residual risk distinguished through documented control effectiveness assessments
• Bias, discrimination, hallucination, privacy, security, and accountability risks assessed
• Risk acceptance decisions recorded with authority limits, conditions, and expiry dates

Participants can design an AI governance operating model with clear forums, roles and escalation routes.

• RACI matrices assign accountable executives, system owners, risk partners, and approvers
• AI steering committee charters define membership, authority, cadence, and decisions
• Approval gates set evidence requirements before pilot, production, and material change
• Exception processes document rationale, compensating controls, expiry, and senior approval
• Escalation routes distinguish operational issues, material incidents, and strategic decisions
• Policy hierarchies connect board principles, standards, procedures, and technical baselines

Participants can specify data controls for AI training, retrieval, prompting and output retention.

• Data lineage records collection, transformation, retrieval, usage, and deletion points
• Data classification controls prompts, attachments, retrieval sources, and generated outputs
• Purpose limitation tested against secondary use, model training, and vendor processing
• Data quality rules address completeness, timeliness, accuracy, representativeness, and provenance
• Retention schedules cover chat logs, agent memory, embeddings, and audit evidence
• Access controls apply least privilege to datasets, knowledge bases, and prompts

Participants can define proportionate model governance, validation and change-control requirements.

• Model documentation records intended use, limitations, assumptions, and known failure modes
• Performance measures selected for accuracy, relevance, refusal quality, and consistency
• Validation independence determined by materiality, autonomy, impact, and model complexity
• Test datasets controlled for provenance, sensitive data, representativeness, and versioning
• Model changes classified across prompts, configurations, versions, retrieval, and tools
• Release decisions require test evidence, approvals, rollback plans, and accountable owners

Participants can design human oversight that is timely, informed and matched to AI decision risk.

• Human approval gates placed before irreversible, high-impact, or external-facing actions
• Reviewer guidance defines authority, evidence, challenge questions, and rejection criteria
• Override controls capture rationale, approver identity, outcome, and recurring causes
• Automation bias reduced through uncertainty indicators, alternatives, and mandatory review prompts
• Decision rights separated between recommendation, execution, exception approval, and accountability
• Customer-facing disclosures specify AI involvement, review routes, and complaint handling

Participants can control agent autonomy through permissions, bounded actions and approval checkpoints.

• Agent authority profiles define permitted goals, actions, data, and transaction limits
• Least-privilege access applied to APIs, credentials, databases, and enterprise applications
• Tool allowlists restrict agents to approved functions and validated destinations
• Transaction limits cap payment values, record changes, communications, and deletions
• Segregation of duties prevents agents from initiating and approving sensitive actions
• Kill switches suspend agent execution during incidents, anomalies, or control failures

Participants can establish controls for prompts, knowledge sources and AI-generated outputs.

• Prompt templates define role, task, constraints, output format, and refusal rules
• System prompts protected through access control, versioning, approval, and change logs
• Prompt injection threats tested against malicious instructions and untrusted retrieved content
• Retrieval filters enforce source approval, metadata checks, recency, and access rights
• Output guardrails detect prohibited content, unsupported claims, and sensitive disclosures
• Citation requirements link generated advice to approved source material and evidence

Participants can identify AI-specific threats and integrate mitigations into security control plans.

• Threat models identify assets, actors, attack paths, impacts, and mitigations
• OWASP LLM risks examined for prompt injection, data leakage, and tool misuse
• Jailbreak testing probes policy bypass, unsafe outputs, and instruction conflicts
• Secrets management protects API keys, service accounts, tokens, and model credentials
• Supply-chain controls assess model sources, libraries, plugins, and hosted services
• Security logging correlates user requests, agent actions, tool calls, and outcomes

Participants can assess AI suppliers and establish contractual, technical and assurance controls.

• Supplier due diligence examines data use, model hosting, security, and subcontractors
• Contract clauses define ownership, confidentiality, incident notice, audit, and exit rights
• Vendor model changes assessed for performance shifts, new risks, and approvals
• Service-level measures cover availability, response quality, support, and incident handling
• Assurance evidence evaluated through certifications, audit reports, test results, and attestations
• Exit planning addresses data return, deletion, replacement, and business continuity

Participants can plan and perform evidence-based testing of AI governance and operational controls.

• Control objectives translated into test procedures, evidence requests, and pass criteria
• Design effectiveness assessed against stated risks, requirements, and control owners
• Operating effectiveness tested through samples, logs, approvals, and observed workflows
• Scenario testing examines harmful outputs, unauthorised actions, and failed escalations
• Control deficiencies rated by impact, likelihood, root cause, and remediation priority
• Assurance reports distinguish findings, observations, management actions, and deadlines

Participants can design monitoring and reporting that enables timely AI governance decisions.

• Key risk indicators measure harmful outputs, overrides, incidents, and control exceptions
• Key performance indicators track accuracy, latency, adoption, cost, and service outcomes
• Agent activity dashboards show tool calls, approvals, failures, and transaction values
• Thresholds trigger alerts, investigation, escalation, suspension, and board notification
• Drift monitoring identifies changing inputs, outputs, users, patterns, and performance
• Management information distinguishes operational detail, executive decisions, and board assurance

Participants can prepare response arrangements for AI failures, harmful outputs and unsafe agent actions.

• Incident taxonomy separates model failure, security event, privacy breach, and misconduct
• Triage criteria assess affected parties, severity, containment urgency, and notification duties
• Containment actions include disabling tools, revoking access, reverting versions, and isolating data
• Forensic evidence preserves prompts, outputs, logs, configurations, and approval records
• Root-cause analysis uses five whys, fault trees, and control failure mapping
• Recovery plans define retesting, reapproval, communications, compensation, and lessons learned

Participants can assemble defensible evidence for internal audit, regulators and external assurance reviews.

• Evidence registers link obligations, controls, artefacts, owners, dates, and storage locations
• Audit trails capture model versions, prompts, data sources, approvals, and agent actions
• Control narratives explain objectives, processes, systems, evidence, and exceptions
• Regulatory enquiries answered through factual records, accountable sign-off, and consistent terminology
• Issue management tracks findings, actions, due dates, validation, and closure authority
• Records retention balances legal obligations, investigation needs, privacy, and storage costs

Participants can prepare decision-useful AI reporting and support effective board-level challenge.

• Board papers summarise AI inventory, material risks, incidents, and decisions required
• Risk appetite breaches presented with exposure, controls, options, and accountable recommendations
• Executive challenge questions test purpose, proportionality, evidence, ownership, and residual risk
• Scenario analysis examines business disruption, customer harm, enforcement, and reputational consequences
• Assurance maps show first-line controls, second-line oversight, and third-line review
• Meeting minutes record challenge, decisions, conditions, dissent, and follow-up actions

Participants can sequence AI governance implementation priorities, resources and adoption actions.

• Maturity assessments compare current practices against target governance capabilities
• Gap analysis prioritises policy, inventory, control, skills, data, and assurance actions
• Ninety-day plans assign milestones, owners, dependencies, budget, and decision points
• Change impact assessments identify affected roles, processes, systems, and suppliers
• Training plans differentiate board awareness, owner responsibilities, user conduct, and specialist skills
• Communications plans explain approved use, prohibited actions, reporting routes, and consequences

Participants can build, review and present a complete AI governance and control pack for a defined agentic AI use case.

• Select a material agentic AI use case with defined business objectives
• Build the AI inventory record, system boundary, and stakeholder map
• Complete risk assessment, impact analysis, classification, and residual-risk decision
• Design governance roles, approval gates, control matrix, and evidence register
• Specify agent permissions, tool restrictions, monitoring metrics, and incident actions
• Present board-ready recommendations, receive challenge, and refine implementation priorities