AIGP domain - 27% of the exam

Understanding how to govern AI deployment and use

Understanding how to govern AI deployment and use is 27% of the AIGP exam. These are the objectives it covers, each with practice questions, with every answer explained.

The domain in numbers

94
Practice questions
6
Objectives
32
Easy
16
Medium
46
Hard

Where people go wrong

  • Swapping the NIST AI RMF functions

    The wrong answers hand an activity to the neighbouring function in the NIST AI Risk Management Framework. Tracking live performance for drift, and gathering trustworthiness evidence on a vendor model, belong to MEASURE, which defines metrics and assesses them. MAP frames context and identifies impacts, including downstream stakeholders. MANAGE responds to and recovers from a realised incident. GOVERN sets policies, roles and culture across the others without performing any one of those tasks.

    See it in a sample question
  • Bypassing retrieval when answers must cite changing content

    The wrong answers reach for fine-tuning, a larger model, a long context, a temperature change or a reworded prompt when answers must follow content that keeps changing and be traceable to a source. Fine-tuning writes facts into weights that cannot cite a clause and needs retraining after each amendment, and the other options supply no current source at all. Retrieval-augmented generation fetches approved passages at query time, so updating the knowledge source changes what the system can answer, and each response points to the text it relied on.

    See it in a sample question
  • Treating agentic risk as a general AI risk

    The wrong answers name a control or concern that applies to any AI system: token volume, retrieval relevance, transparency notices, better training data, latency dashboards or output-drift detection. An agentic architecture lets a model chain its own decisions into tool calls that change live systems, so errors compound without a person checking each step. The matching controls constrain and authorise the actions its tools may take, add human approval for high-impact steps, and capture multi-step action traces.

    See it in a sample question
  • Mistaking vendor documents for an impact assessment

    The wrong answers treat documents that already exist as the impact assessment: a SOC 2 report, a model card, a vendor-prepared assessment strong on accuracy, a single filed record, or one accuracy metric. An impact assessment evaluates how the system could affect the rights and interests of the people it decides about in this deployment, including the deployer's own population and how staff act on its outputs. It is then revisited through monitoring and review triggers when the system, its data or its context of use changes.

    See it in a sample question
  • Picking cloud or hybrid for a site-bound constraint

    The wrong answers choose cloud, hybrid, a hosted interface or batched upload where the site has a tight latency budget or a data-localisation rule. Each sends data off site and depends on a network link that is intermittent or too slow. An in-country cloud region still leaves the provider administering the infrastructure. Edge deployment on the line keeps inference local, and on-premise deployment gives the direct physical and administrative control that a regulator's residency expectation asks for.

    See it in a sample question

Objectives in this domain

What to study

The official documents this domain's practice questions cite most.

Sample questions from this domain

Free sampleUnderstanding how to govern AI deployment and usehard

A bank has deployed a credit-scoring model that passed validation at launch. Six months later the operations team notices approval rates have shifted, though no code or configuration has changed. Using the NIST AI Risk Management Framework, which function's activities are the team primarily carrying out when they set quantitative thresholds, instrument the live system, and track scoring performance over time to detect this degradation?

  • AMAP, because the team is establishing the deployment context and cataloguing the categories of risk that the credit-scoring system could plausibly create in operation.
  • BMEASURE, because the team is defining metrics and continuously assessing the deployed model's performance to identify drift and other emergent risks over time. Correct
  • CMANAGE, because the team is allocating resources to treat the identified drift and is deciding which mitigation or rollback response to apply first.
  • DGOVERN, because the team is enforcing the organisational policies and accountability structures that require credit models to be supervised after launch.
Continuous monitoring that quantifies and tracks a deployed model's performance to detect drift sits within the NIST AI RMF MEASURE function. The MEASURE function is defined by analysing, assessing, benchmarking and tracking AI risks and impacts using appropriate metrics; detecting post-deployment drift through instrumented performance metrics is a core MEASURE activity, distinct from framing risk (MAP) or treating it (MANAGE).

Why A is wrong: MAP is tempting because contextual framing precedes measurement, but MAP establishes context and identifies risks up front rather than continuously quantifying and tracking live performance, which is the activity described.

Why B is correct: MEASURE covers selecting metrics and using quantitative and qualitative methods to assess, analyse and track identified risks of a deployed system, which is exactly the instrumentation and monitoring of scoring performance described.

Why C is wrong: MANAGE is tempting since acting on drift eventually falls here, but it concerns prioritising and responding to risks already measured, not the measurement and tracking step that surfaces the drift.

Why D is wrong: GOVERN is tempting because a monitoring mandate originates in governance, but GOVERN sets the culture, policies and roles that cut across functions rather than performing the live quantitative tracking itself.

Free sampleUnderstanding how to govern AI deployment and usemedium

A compliance team deploys an assistant that answers staff questions about a regulatory rulebook that is amended several times a month, and auditors require each answer to point to the specific clause relied upon. The base model already writes fluently. Which technique best fits both the update frequency and the traceability requirement?

  • AFine-tuning the model on the rulebook after every amendment, so the current clauses are encoded directly into the model weights.
  • BRetrieval-augmented generation over an indexed copy of the rulebook, re-indexing on each amendment and injecting the matching clause text and its identifier into the prompt. Correct
  • CIncreasing the model's context window and pasting the entire rulebook into every system prompt so the assistant always has the full text available.
  • DAdding a lower temperature and a stricter system prompt so the model is instructed to answer cautiously and reference the relevant rule from memory.
Select retrieval-augmented generation when knowledge changes frequently and answers must be traceable to a specific cited source. Fine-tuning fixes knowledge in weights at training time and discards source identifiers, making frequent updates expensive and attribution unreliable. Retrieval-augmented generation decouples knowledge into a refreshable index and supplies the verbatim clause and its identifier in the prompt, so it both tracks frequent amendments at low cost and lets the assistant attribute each answer to its source.

Why A is wrong: Tempting because fine-tuning absorbs domain content, but retraining several times a month is costly and slow, and weights do not retain clause identifiers, so the model paraphrases rather than attributing to a specific source.

Why B is correct: Retrieval-augmented generation keeps knowledge in an external index that is cheap to refresh on each amendment, and because the retrieved clause and its identifier are placed in the prompt the assistant can ground and attribute each answer to a specific source.

Why C is wrong: Tempting because a long context exposes the whole text, but loading the full rulebook into every request is wasteful, and without a retrieval step the assistant has no reliable mechanism to cite the specific clause an answer relied upon.

Why D is wrong: Tempting because conservative decoding reduces invention, but the base model has no access to amendments after its training cutoff and cannot reliably reproduce a clause identifier from memory, so traceability is not achieved.

Free sampleUnderstanding how to govern AI deployment and usemedium

A governance team is explaining why an agentic AI architecture that can call tools and act over several steps needs a control category that a retrieval augmented generation assistant typically does not. Which governance concern is most distinctive to the agentic architecture?

  • AEnsuring the external knowledge store is kept current and well indexed, so that the AI system does not ground its answers in stale or missing source documents over time.
  • BConfirming that the base model's parameters were updated correctly during its training run, so that the specialised behaviour was encoded into the model as intended.
  • CVerifying that the user interface displays a clear disclaimer telling people they are interacting with an AI system rather than with a human agent.
  • DConstraining and authorising the actions the agent may take through its tools, since it can autonomously chain steps that affect external systems and compound errors with no human review. Correct
Recognise that agentic architectures require action-scoping and authorisation controls because autonomous tool use can affect external systems and compound errors. Because an agentic AI system plans and autonomously invokes tools that take real actions across several steps, a single early mistake can cascade through later actions and reach systems beyond the model before any person intervenes. The distinctive governance response is to constrain which tools and actions the agent may use and to require authorisation for consequential operations, a control category that a retrieval augmented generation assistant, which only returns grounded text, does not need to the same degree.

Why A is wrong: Tempting because freshness is a real concern, but keeping an external knowledge store current is the characteristic governance concern of retrieval augmented generation, not the distinctive issue raised by an agent that takes actions.

Why B is wrong: Tempting because validating training matters, but confirming that parameters were updated correctly is a fine-tuning concern; an agentic architecture's distinctive risk arises from autonomous action, not from how the model was trained.

Why C is wrong: Tempting because disclosure is a genuine transparency duty, but an interaction disclaimer applies broadly to many AI systems and does not address the action-taking autonomy that makes the agentic pattern distinctive.

Why D is correct: An agentic architecture autonomously selects and executes tool actions across multiple steps that can change external systems, so the distinctive control is bounding which actions are permitted and requiring authorisation, because errors propagate and effects reach beyond text without per-step human review.

Free sampleUnderstanding how to govern AI deployment and usehard

An insurer is reviewing an AI impact assessment that its vendor prepared for a claims-triage AI system before the insurer deploys it. The assessment is thorough on the model's technical accuracy and the vendor's development controls, but it describes affected groups only as a generic population and does not consider the insurer's own claimant demographics or the way the insurer's adjusters will act on the system's recommendations. As the reviewer, what is the most important gap to address before relying on this assessment?

  • AThe assessment should be rejected outright because any impact assessment prepared by the vendor rather than the insurer is inherently invalid and cannot be relied upon for a deployment decision.
  • BThe assessment is sufficient as it stands because the vendor has documented the model's accuracy metrics and its internal development controls, which are the substance of an impact assessment.
  • CThe assessment must be extended to evaluate the system in the insurer's actual deployment context, including its specific claimant population and how adjusters will use the recommendations in practice. Correct
  • DThe reviewer should focus the remaining effort on renegotiating the vendor's liability cap so the contract covers a larger share of any future claims arising from the system.
A deployer reviewing a vendor-supplied impact assessment must extend it to its own context of use, affected populations and human decision behaviour. An AI impact assessment measures effects on affected individuals in the specific context where the system operates, so a generic vendor assessment that omits the deployer's real population and how humans act on outputs is incomplete and must be extended before the deployment decision relies on it.

Why A is wrong: Distrust of vendor-prepared documents is understandable, but a vendor assessment is a useful input that the deployer extends rather than something automatically invalid, so rejecting it outright discards relevant development-stage evidence the deployer would otherwise have to reproduce.

Why B is wrong: Strong accuracy and development documentation can look like a complete assessment, but an impact assessment is about effects on affected people in the context of use, not just technical performance, so treating it as sufficient overlooks the contextual harms the review exists to surface.

Why C is correct: This is correct because impact depends on the concrete context of use, and a generic vendor assessment cannot capture the insurer's real claimant demographics or its adjusters' decision behaviour, so the deployer must extend the assessment to its own setting before relying on it.

Why D is wrong: Adjusting the liability cap is a legitimate commercial control and feels protective, but it is a contractual risk-transfer step rather than part of completing the impact assessment, so it does nothing to identify the contextual harms the assessment currently omits.

Free sampleUnderstanding how to govern AI deployment and usemedium

A logistics operator wants an AI vision model to flag damaged parcels on a conveyor line inside a warehouse that has intermittent connectivity and a hard requirement that frame images never leave the site. Which deployment option best fits these constraints?

  • AA public cloud inference endpoint, because managed autoscaling removes the need to provision local hardware for the vision workload.
  • BA hybrid pattern that streams every captured frame to a central cloud model and caches only the verdicts locally for resilience.
  • CAn on-premise server in the head-office data centre that the warehouse connects to over the corporate wide-area network for each prediction.
  • DEdge deployment on a device at the conveyor line, because inference runs locally with low latency and image data stays on the premises even when connectivity drops. Correct
Recognise that edge deployment suits low-latency, offline-tolerant use cases where data must remain local to the point of capture. Edge deployment runs the model on hardware co-located with the data source, so predictions are produced without a round trip to a remote service. This both removes the dependency on a reliable network link and keeps raw inputs on the premises, which is why it fits scenarios combining latency sensitivity, intermittent connectivity, and a no-egress data requirement.

Why A is wrong: Tempting because cloud inference simplifies operations and scaling, but it depends on stable connectivity and would route frame images off site, breaching the stated data-localisation requirement.

Why B is wrong: Tempting because hybrid designs balance local and central processing, but streaming every frame still sends image data off site and remains exposed to the intermittent link, so it fails the core requirements.

Why C is wrong: Tempting because on-premise keeps data within the organisation, but routing each prediction over the wide-area network reintroduces a connectivity dependency and added latency that an at-line edge device avoids.

Why D is correct: Edge deployment places the model on hardware next to the data source, so inference continues during connectivity loss and the frames are processed locally rather than transmitted, satisfying both the latency and data-localisation constraints.

Other domains in this exam

See also the AIGP cert hub, the study guide, and the cheat sheet.

Examworthy is not affiliated with or endorsed by IAPP. Original, blueprint-aligned practice material only.