Domain 3 review: governance and responsible AI

Hannah Weiss, Cascade Ridge's privacy and security lead, had been part of every AI conversation since the autumn, usually as the person who asked what could go wrong. Noor invited her to the governance review. "This domain is where good business ideas meet the real world," Hannah said. Domain 3, AI Governance and Responsible AI Leadership, carries 24 of every 100 scored points and has three tasks: applying responsible AI principles to business decisions, establishing governance structures and compliance, and identifying enterprise AI risks and directing their mitigation.

Task 3.1: responsible AI in business decisions

The guide names six responsible AI principles as examples. The exam wants you to apply them to scenarios, not recite definitions, so the table pairs each with the kind of business question it raises.

Skill 3.1.2 is about tradeoffs when business goals conflict with these principles. A more accurate model may be less explainable; a more personalized service may need more personal data; a faster rollout may skip a fairness review. The exam's expected answer is almost never "ignore the principle" and rarely "abandon the goal". It is usually a balanced step: choose an explainable approach for high-stakes decisions, collect only the data needed, run a fairness test before rollout, or add human review where errors would harm people.

Skill 3.1.3 is about timing: responsible AI practices belong in project planning from the start, not as a review at the end. A scenario that describes a team planning to "add a fairness check after launch" is pointing at the wrong answer. Building requirements in at design time is cheaper and more effective than fixing problems once a system is in use.

Skill 3.1.4 covers human oversight and safeguards. The more consequential the decision and the more likely an error, the stronger the oversight needed. Safeguards the guide names include hallucination detection (checking whether generated content is supported by sources), guardrails (rules that block topics, filter harmful content, remove personal information or keep answers within scope; on AWS, Amazon Bedrock Guardrails provides these), and escalation criteria that define when the system must hand over to a person.

Task 3.2: governance structures and compliance

Skill 3.2.1 asks you to set up governance with cross-functional representation and clear accountability. A typical structure has an executive sponsor accountable to the board; an AI governance council or committee with business, technology, legal, privacy, security, risk and affected-function representatives; and a named business owner for each AI system who is accountable for its outcomes. Governance that sits entirely within IT, or entirely within legal, misses half the picture.

Skill 3.2.2 is about identifying and addressing regulatory compliance risks. The exam is international and will not test the details of one country's law, but it expects you to recognize when a scenario raises privacy, consumer protection, employment, financial, health or sector-specific rules, and to involve legal and compliance early.

For Cascade Ridge in the United States, there is no single federal privacy law for this sector. Obligations come from state privacy laws, the Federal Trade Commission's powers over unfair and deceptive practices, and sector rules such as those for card payments. Colorado has passed an AI law placing duties on developers and deployers of high-risk AI systems used in consequential decisions about people, such as employment, housing and lending; its requirements and dates have been revised, so check its current status.

Many organizations also look at the EU AI Act as a model of risk classification, even outside Europe, because it sorts AI uses into tiers: prohibited practices, high-risk systems with strict obligations, limited-risk systems with transparency duties (such as telling people they are talking to a chatbot), and minimal-risk systems. Skill 3.2.4 asks you to apply risk classification frameworks to prioritize governance effort across the AI life cycle. The business point is proportionality: a marketing image generator and a system that recommends treatment plans should not get the same level of review.

Skill 3.2.3 covers access controls and data security for AI systems: least-privilege access to models and data, separating environments, encrypting data, logging use, and making sure a generative AI assistant can only retrieve documents the asking user is allowed to see. On AWS, the managed services keep customer prompts and outputs within the customer's account and, for Amazon Bedrock, are not used to train the underlying models; checking such terms is part of any vendor decision.

Task 3.3: enterprise AI risks and mitigation

The final task asks you to identify enterprise AI risks and direct mitigation. Skill 3.3.1 is the recognition that AI systems in production need risk controls and monitoring, not just pre-launch testing. The remaining skills name particular risks.

Skill 3.3.2 makes a point that exam questions like to test: bias can arise at many stages, not just in training data. It can enter through which problem is chosen, how data is collected and labeled, how the model is designed, how outputs are used by people, and how the world changes after launch. That is why one fairness test before launch is not enough, and ongoing monitoring for bias drift is required.

Practice scenarios for Domain 3

Governance questions reward balance. Hannah's advice to Andre was to look for the option that lets the business proceed safely, rather than the one that sounds strictest or the one that sounds fastest.

Governance across the life cycle

Several governance questions ask when a control belongs. A simple life-cycle map helps. At the idea stage, classify the use by risk and decide the level of review. At design, set responsible AI requirements, data rules, access controls and the oversight model. Before launch, test for fairness, safety, robustness and accuracy against agreed thresholds, and confirm compliance. In operation, monitor performance, bias, harmful content and costs, keep logs, and review incidents. At retirement, remove access, handle data according to retention rules and record lessons learned. An option that places a control at the wrong stage, such as defining fairness requirements only after launch, is usually a distractor.

Questions on intellectual property tend to ask about practical steps rather than law: checking a vendor's terms on ownership of outputs and use of customer data, preferring providers that offer indemnity for outputs where available, reviewing generated content before public use, and confirming that any data used to adapt a model is licensed for that purpose.

One more pattern is worth knowing. When a scenario describes a harmful output that has already reached customers, the expected answer usually combines an immediate containment step, such as tightening a guardrail or adding human review, with an investigation of the cause and a report through the governance structure. Options that only apologize, or only retrain the model without containment, are incomplete.