Turning responsible AI policy into controls: AIP-C01 study notes and practice questions
What to know about Turning responsible AI policy into controls for the AIP-C01 exam, from the AI Safety, Security, and Governance domain (20% of the exam guide), then 1 free question to test it. Part of the AIP-C01 study guide.
AIP-C01 gives a written policy, such as “no investment advice,” “a person must decide loan denials,” or “users must know they are talking to AI,” and asks which technical control enforces it. The skill being tested is translation: each rule needs a mechanism, an owner and a test.
From clause to control
Write each policy rule as a testable statement with an owner and a risk level. Then map it to a control:
A prohibited subject becomes a denied topic in a guardrail.
A data-handling rule becomes a sensitive information filter and a retention rule.
A rule that a person decides becomes an approval step in the workflow.
A rule about disclosure becomes a UI element and a logged acknowledgment.
A rule with no control is a statement of intent, not governance.
Guardrails from policy
Configure the guardrail from the policy document, not from memory. Store the guardrail definition as code, version it, and apply the same version across applications. A policy change becomes a reviewed change to the guardrail version, rolled out through the pipeline.
Pre-release compliance checks
Add a compliance stage to the release pipeline. It runs the red team suite, grounding tests, PII tests and fairness checks, confirms that a model card exists and is current, and blocks the release on failure. An automated check can be a Lambda function or a CodePipeline stage that calls the evaluation jobs.
Human oversight
For high-risk decisions, the model proposes and a person decides. Implement this with a Step Functions task-token callback that waits for approval, and record who approved and when. Oversight must be real: the reviewer needs the evidence, the authority to reject, and enough time. A reviewer who approves everything meets the form of the policy but not its intent.
Disclosure and opt-out
Tell users they are interacting with AI, and offer a route to a human agent. Keep the disclosure visible in the interface, and log that it was shown when the policy requires that record.
How to choose
Scenario wording
Choose
“Policy forbids investment advice”
Denied topic in the guardrail
“Release only if policy tests pass”
Automated compliance stage in the pipeline
“Loan denials must be decided by a person”
Human approval through a Step Functions callback
“Users must know the system is AI”
Interface disclosure with a human handoff option
“The policy changes every quarter”
Guardrail version managed as code and rolled out through the pipeline
Common traps
Writing a policy with no enforcement. Each clause needs a control and a test.
Relying on the model’s own refusals rather than a guardrail that enforces the rule.
Approving once and never re-reviewing after a policy or model change.
Treating human review as a rubber stamp, without evidence or authority to reject.
Editing a live guardrail in place instead of publishing a tested version.
An energy utility's employee benefits team is launching an assistant that answers staff questions about leave, pay and health plans. The utility has a written responsible AI policy that sets the rules the assistant must follow.
The policy says the assistant must not give medical advice, must not reveal employee identifiers, and must give leave and pay answers that match the approved leave policy document. The leave rules include numeric limits and waiting periods that the assistant must not misstate.
The team wants each rule enforced in the runtime configuration, and it wants a record that ties each control back to the policy clause it implements.
Which approach meets these requirements?
Choose one answer.
AA. Paste the full policy into the system prompt, ask the model to follow it, and review a random sample of conversations each quarter to confirm that the assistant complied with each clause.
BB. Map each policy clause to a guardrail control: a denied topic for medical advice, a sensitive information filter for identifiers, and Automated Reasoning checks for leave answers.
CC. Fine-tune the model on the policy document and on sample sets of approved answers, so the model learns the rules internally, and then test the tuned model on a fixed list of staff questions.
DD. Set every content filter to HIGH strength, and require help desk staff to review each assistant answer about leave or pay before it is sent to an employee.
Show the answer and explanation
Answer: B
A. Wrong: A system prompt is a request the model may not honor, and a quarterly sample cannot show that each clause was followed in every conversation. The option is not enforced at runtime and leaves no clause-level record.
B. Right: Each clause becomes a runtime control in the guardrail, so the rule is enforced on every reply. The Automated Reasoning policy checks numeric leave and pay answers against formal rules from the approved document, and the mapping gives the clause-level record that the team needs. Learn more: Automated Reasoning checks in Amazon Bedrock Guardrails (opens in a new tab)
C. Wrong: Fine-tuning can shape behavior, but it does not enforce each rule on each reply, and a policy change means another training run. A fixed question list also gives no clause-level record of how each rule is enforced. Learn more: Customize your model (opens in a new tab)
Mapping each policy clause to a guardrail control enforces the rule on every reply, and an Automated Reasoning policy checks leave and pay answers against formal rules from the approved document. Keeping that mapping gives the record tying controls to clauses. Prompts, fine-tuning and human review do not enforce each clause at runtime.
Want more on Turning responsible AI policy into controls? The 5 full-length practice sets have 6 more Turning responsible AI policy into controls questions, 375 in all, none of them in the free exam. One payment of $14.99, no subscription. See the practice sets →