ISACA Advanced in AI Risk

AI risk should live inside the risk framework the organization already runs. You connect AI to enterprise risk management, set appetite for it, and compare the main AI risk frameworks and standards.

• Enterprise risk management and where AI attaches
• Setting risk appetite for AI use
• The NIST AI Risk Management Framework
• ISO/IEC 42001 and ISO/IEC 23894 compared
• Mapping frameworks to existing controls
• Extending the three lines model to AI

You build A framework integration map showing where AI risk enters each step of the existing risk process.

Nobody manages a risk nobody owns. You assign accountability for AI systems, write the policies that govern them, set up escalation paths, and build the training that lets people recognize AI risk.

• Owners for models, data and outcomes
• AI policy suite and its hierarchy
• Review boards and escalation
• Training for developers, users and leaders
• Reward and challenge in risk culture
• Keeping an AI inventory current

You build An accountability model and policy outline for AI, with named roles and escalation rules.

AI is subject to law and to public expectation. You map regulation and ethical principles to use cases, assess impact on the people affected and record the obligations that become controls.

• Regulatory approaches to AI by jurisdiction
• Risk tiers and prohibited uses
• Fairness, transparency and explainability duties
• Human oversight requirements
• Impact assessments for affected people
• Turning obligations into control requirements

You build An obligations register for one AI use case, with each duty linked to a control.

Many AI risks are set long before training begins. You test whether a use case should be built at all, whether to buy or build it, and whether the data is suitable, lawful and representative.

• Use case intake and risk tiering
• Build, buy or adapt decisions
• Fitness of data for the intended purpose
• Representativeness and bias in training data
• Consent, licensing and privacy of data
• Data and model assets as things to protect

You build An intake and tiering form that sorts ten sample AI proposals by risk.

A model that works in a lab may fail in use. You set independent validation, test for robustness and fairness, define go-live criteria and decide what must be in place before a model reaches people.

• Independent model validation
• Performance, robustness and stress testing
• Fairness and explainability testing
• Adversarial and red team testing
• Go-live criteria and approvals
• Human oversight and rollback at deployment

You build A validation and release checklist for one AI system, with go and stop criteria.

Risk continues after release. You monitor for drift and misuse, control changes and retraining, and retire models and their data in a way that leaves no uncontrolled copies behind.

• Monitoring performance, drift and misuse
• Change control for models and prompts
• Retraining triggers and re-validation
• Version management and model lineage
• Retirement and data deletion
• Periodic review of live AI systems

You build A monitoring and review plan for one live model, with thresholds that trigger action.

Risk assessment turns worry into decisions. You build scenarios for AI failure and misuse, estimate likelihood and impact with available evidence, and record the results so they can be challenged.

• Building AI risk scenarios
• Cause, event and consequence analysis
• Estimating likelihood with little history
• Quantifying impact
• Expert judgment and its biases
• Recording scenarios in the risk register

You build Eight AI risk scenarios assessed and entered in a risk register with reasoning.

Each risk needs a decision. You choose treatments, design controls specific to AI such as guardrails, oversight and testing, and check whether those controls work.

• Treatment options for AI risk
• Guardrails and technical controls
• Human oversight as a control
• Control testing and effectiveness
• Residual risk and acceptance
• Exceptions and compensating controls

You build A treatment plan for three AI risks, with controls, owners and residual risk.

Leaders need a few honest signals. You design indicators for AI risk, build reports for risk committees and boards, and track issues to closure.

• Key risk indicators for AI
• Thresholds and escalation triggers
• Dashboards for risk committees
• Board reporting on AI risk
• Issue tracking and remediation follow-up
• Avoiding misleading measures

You build A one-page AI risk dashboard with indicators, thresholds and the action each triggers.

Most AI is built from parts someone else made. You assess model providers, open-source models, datasets and platforms, set contract terms and watch for concentration and provenance risk.

• Third-party models, datasets and platforms
• Due diligence questions for AI vendors
• Contract terms for data use and liability
• Open-source models and provenance
• Concentration and dependency risk
• Monitoring suppliers over time

You build A due diligence questionnaire and scoring method for an AI supplier.

AI systems fail in unfamiliar ways. You define AI incidents, plan detection and response, decide when to switch a system off, and keep the business running on fallback processes.

• Types of AI incident and their severity
• Detection and reporting channels
• Response playbooks for harmful output and data leakage
• Shutdown criteria and fallback processes
• Notifying regulators and affected people
• Learning from AI incidents

You build An AI incident response plan with severity levels, a shutdown rule and a fallback process.

One program draws all the parts together. You set the framework, inventory and assess a portfolio of AI use cases, treat their risks and report to leadership on what remains.

• Scoping the AI portfolio
• Assessing and tiering use cases
• Selecting treatments and controls
• Setting indicators and reporting
• Presenting the program to a risk committee

You build An AI risk program package: framework, inventory, assessments, treatments and reporting.