Select Two (Multi-Select) Questions

Establish aligned risk scoring criteria across all business functions.

Aligning scoring criteria lets AI risks be rated on the same likelihood and impact scales as other enterprise risks, so they aggregate into existing heat maps and reporting rather than sitting in a separate silo. This directly satisfies the stem's requirement to integrate AI risk into the broader ERM framework.

Define a common risk taxonomy that includes AI-specific concepts like bias and drift.

A shared taxonomy embedding AI-specific concepts such as bias and drift lets ERM register, score and report AI risks alongside operational and strategic ones, satisfying the stem's requirement to integrate AI risk into the broader framework rather than managing it in isolation.

Ensure that AI risk owners report into the same governance channels as other risk owners.

Routing AI risk owners through the same governance channels as other risk owners embeds AI within existing ERM escalation and reporting structures, satisfying the stem's integration requirement. It prevents AI being managed as a siloed technical concern, ensuring consistent oversight, accountability and board-level visibility alongside enterprise risks.

Intended use cases

Model Cards record intended use cases, defining the tasks and contexts the model was designed and evaluated for, which guides appropriate deployment. This satisfies the stem's requirement for factors that must be documented in an AI Model Card.

Model limitations and known risks

Model Cards document limitations and known risks so downstream consumers understand failure modes, bias exposure and out-of-scope behaviour before deployment. This satisfies the stem's requirement for factors that must be recorded in an AI Model Card.

Model performance and accuracy

Model performance and accuracy must be covered because degraded, drifting or biased outputs create operational and safety risk. Monitoring these metrics ensures the programme detects failures, satisfying the stem's requirement that risk management addresses the model's ongoing behaviour rather than only deployment-time concerns.

Third-party/Vendor risk management

Third-party and vendor risk management must be covered because AI systems often rely on external models, APIs and data providers whose failures or compliance gaps become the organisation's exposure. Assessing vendors satisfies the stem's requirement that the risk programme extends across the full supply chain.

Legal and regulatory compliance

Legal and regulatory compliance must be covered because AI systems face data protection, sectoral and emerging AI legislation. Addressing this area ensures the risk programme identifies applicable obligations and controls, satisfying the stem's requirement that risk management spans governance concerns beyond purely technical model issues.

Data Sheets for Datasets explaining the provenance and makeup of training data.

Data Sheets for Datasets document a training dataset's provenance, composition, collection methods, and known limitations. This satisfies AI transparency by disclosing what data shaped the model, enabling stakeholders to assess suitability, bias exposure, and fitness for purpose rather than treating the model as an opaque artefact.

Model Cards detailing intended use, limitations, and performance.

Model Cards document a model's intended use, known limitations and performance metrics, directly satisfying AI Transparency requirements by giving stakeholders the information needed to understand how a model behaves and where it should not be applied.

Data privacy and residency requirements for training sets

Data privacy and residency requirements constrain which training data may lawfully be used and where it resides. For a generative chatbot, these legal and contractual limits directly shape acceptable risk exposure, so the committee must weigh them against organisational risk tolerance.

The likelihood of model drift impacting output accuracy

Model drift erodes accuracy as real-world data diverges from training distributions, so its likelihood determines how much output degradation the organisation will tolerate. This directly informs the chatbot's risk appetite, monitoring frequency and retraining triggers.

The impact of potential hallucination on brand reputation

Hallucination risk directly shapes brand exposure, so its potential reputational impact must be weighed against the committee's stated risk tolerance. This factor translates an abstract appetite into a concrete harm threshold, ensuring the chatbot's deployment stays within acceptable organisational limits.

Total number of AI-related incidents reported and remediated.

Counting reported and remediated AI incidents gives a direct, quantifiable measure of the risk programme's operational effectiveness, satisfying the stem's demand for metrics rather than qualitative assessments. Remediation volume evidences that identified risks are actually being treated, not merely logged.

Percentage of AI models that have undergone a formal risk assessment.

Tracking the proportion of models subjected to formal risk assessment gives a direct, quantifiable coverage measure of programme reach. It satisfies the stem's requirement for an effectiveness metric because it evidences governance activity across the AI inventory rather than counting intentions.

Rate limiting on API calls to prevent brute-force querying.

Rate limiting constrains the volume and frequency of queries an attacker can submit, directly blunting brute-force and model-extraction attempts that rely on mass probing. It satisfies the adversarial-risk constraint by capping automated abuse without blocking legitimate traffic.

Adversarial training to harden the model against perturbed inputs.

Adversarial training augments the training set with deliberately perturbed inputs labelled correctly, so the model learns decision boundaries robust to small malicious modifications. This directly hardens the model against evasion attempts, mitigating adversarial risk during the lifecycle.

Input sanitization and validation on all model endpoints.

Sanitising and validating inputs at every model endpoint strips malicious payloads before inference, defeating prompt injection, malformed data and adversarial-example delivery. This directly addresses adversarial risk by hardening the model's input surface against crafted manipulation.

Performance disparity across different demographic cohorts.

Fairness requires measuring whether model outcomes differ across protected groups, so comparing error rates, selection rates and accuracy between demographic cohorts exposes disparate impact. This directly satisfies the stem's demand for a fairness factor, since parity across cohorts is the defining test of equitable model behaviour.

Defined metrics for success that explicitly account for equity.

Fairness cannot be assessed without explicitly defined success metrics that account for equity, such as equalised odds or demographic parity across groups. Choosing these measures determines what the review actually evaluates, making them a prerequisite factor rather than an afterthought.

Training data representation across protected groups.

Training data representation across protected groups determines whether the model learns patterns applicable to each population. Under-representation or skewed sampling produces disparate error rates, so auditing group composition within the training set is a core fairness evaluation factor.

Measurement bias (data collection)

Measurement bias arises when data collection instruments or sampling methods systematically distort values, so the training set misrepresents the population the model must serve. Recognising it as a distinct bias category satisfies the stem's requirement to identify three manageable bias types, since it can be mitigated through validated collection tools and representative sampling.

Sampling bias (data selection)

Sampling bias arises when training data is not representative of the population the model will serve, so the model's outputs skew towards over-represented groups. Managing it satisfies the stem's requirement to identify bias types, alongside other recognised categories such as exclusion, association and measurement bias.

Algorithmic bias (model logic)

Algorithmic bias arises from the model's own logic and training, such as skewed feature weighting or flawed objective functions, producing systematically unfair outputs. Managing it requires testing model behaviour across groups, since the defect lives in the algorithm rather than the data or its collection.

Applying fairness constraints during model training

Fairness constraints embedded in the training objective directly penalise biased outcomes during optimisation, so the model learns equitable decision boundaries rather than having bias patched afterwards. This satisfies the mitigation requirement by reducing disparate impact at its source, before deployment.

Training on more diverse and representative datasets

Training on more diverse and representative datasets reduces bias by exposing the model to broader demographic and contextual variation, preventing systematic underrepresentation. It targets the root data cause rather than masking symptoms through post-hoc adjustment.

The organization's tolerance for legal and reputational damage.

Tolerance for legal and reputational damage directly bounds how much AI risk an enterprise will accept, satisfying the governance constraint that risk appetite must reflect consequences the board deems intolerable. Unlike technical metrics, this factor translates AI failures into enforceable liability and stakeholder trust thresholds, anchoring the appetite statement in measurable exposure limits.

The complexity and autonomy level of the AI models being deployed.

Greater model complexity and autonomy directly widen the potential blast radius of failures, so they set how much risk an enterprise can absorb. Autonomous agents acting without human review demand a narrower appetite than simple classifiers, making this a defining input to the AI Risk Appetite.

The existing regulatory landscape and sector-specific compliance requirements.

Regulatory landscape and sector compliance requirements bound how much AI risk an enterprise may accept, since breaches trigger legal penalties and licence loss. They define the mandatory ceiling against which tolerance is set, making them essential to the AI risk appetite calculation.

Conducting fairness audits using specialized metrics (e.g., Disparate Impact).

Fairness audits apply specialised statistical metrics such as Disparate Impact to quantify outcome disparities across protected groups. Measuring bias on defined axes lets teams detect and remediate discriminatory behaviour before deployment, satisfying the mitigation requirement rather than merely documenting intent.

Sourcing training data from diverse, representative populations.

Diverse, representative sourcing ensures training data reflects the real deployment population across demographic and contextual dimensions. This directly reduces sampling bias, preventing the model from learning skewed correlations that would otherwise produce discriminatory predictions for under-represented groups.

Reviewing and approving the organization's AI Risk Appetite Statement.

Reviewing and approving the AI Risk Appetite Statement is a core Board responsibility, since setting risk tolerance is a governance function that cannot be delegated to management. This activity satisfies the stem's oversight constraint by ensuring the Board formally defines how much AI risk the organization is willing to accept before any AI initiative proceeds.

Designating executive-level accountability for AI risk management outcomes.

Assigning executive-level ownership ensures AI risk appears on the Board's agenda through a named accountable officer, satisfying the stem's requirement for adequate Board oversight within the Corporate Governance framework rather than delegating it to technical teams alone.

Establishing a reporting cadence for significant AI-related risk incidents and model performance trends.

A recurring reporting cadence gives the Board timely visibility of significant AI incidents and model performance trends, satisfying the oversight requirement rather than one-off updates. This continuous flow of risk information enables directors to challenge management, track emerging harms, and evidence discharge of their governance duties under the Corporate Governance framework.

Incomplete datasets with many missing values.

Incomplete datasets with many missing values directly degrade training data quality, since gaps bias model learning and reduce statistical representativeness. This satisfies the stem's training-phase constraint by addressing data completeness, one of the three recognised contributors to Data Quality Risk alongside accuracy and consistency issues during model development.

Systematic errors in the target label annotations.

Systematic annotation errors teach the model the labellers' consistent mistakes rather than the true mapping, producing biased predictions that validation against the same flawed labels cannot reveal, which is a core data quality risk at training time.

Selection bias in the training set creation process.

Selection bias arises when the sampling method over- or under-represents certain groups, so the training set fails to reflect the deployment population. This skews learned patterns, producing systematically unfair or inaccurate outputs — a data quality risk originating in training-set creation rather than labelling or collection.

The distribution of prediction outputs.

Monitoring the distribution of prediction outputs detects drift because shifts in input data or concept change the spread and frequency of predicted classes or values. Tracking this axis reveals degradation even when ground-truth labels are unavailable in the deployment phase.

The confidence score of the model predictions.

Monitoring prediction confidence scores exposes degradation in the model's decision certainty, satisfying the deployment-phase requirement to detect drift. A sustained fall in confidence indicates the input distribution has shifted away from training data, making this a direct, measurable signal of model drift.

The distribution of incoming input features.

Comparing live input feature distributions against the training baseline detects data drift directly. Shifts in feature ranges, categories, or frequencies mean the model is scoring inputs unlike those it learned from, which is the earliest drift indicator.

Identification of potential failure modes and their impact.

Identifying potential failure modes and their impact directly satisfies the risk assessment requirement by mapping what could go wrong and the severity of each outcome. This structured analysis lets stakeholders prioritise mitigations before deployment, covering both likelihood and consequence dimensions essential to a defensible AI risk report.

The residual risk level after applying controls.

Recording residual risk after controls shows the exposure remaining once mitigation is applied, which the risk assessment report must state so approvers can judge whether the project proceeds within tolerance. Without it, the report only lists inherent risk and gives no basis for an informed go/no-go decision.

Documentation of mitigating controls for identified risks.

Documenting mitigating controls directly satisfies the report's purpose of showing how each identified risk is reduced to an acceptable level. A risk assessment that lists threats without corresponding controls leaves residual risk unquantified, so this item is essential for governance sign-off and audit evidence.

Model validation results

Model validation results evidence that the model meets defined performance, fairness and robustness thresholds before deployment. Auditing these records confirms independent testing occurred, satisfying the lifecycle compliance requirement for AI risk integration within enterprise risk management.

Incident response logs

Incident response logs must be audited because they evidence how AI-related failures, drift or misuse were detected, triaged and remediated. Within ERM, these records demonstrate that model incidents feed back into risk registers and control testing, satisfying compliance and continuous-improvement obligations.

Data lineage and provenance

Data lineage and provenance must be audited because AI compliance depends on tracing training, validation and inference data back to its sources, confirming lawful collection and detecting drift or contamination. Without documented lineage, model outputs and regulatory attestations cannot be substantiated during ERM review.

Status of compliance certifications for high-risk models.

Displaying compliance certification status for high-risk models shows senior management whether regulatory and internal obligations are currently met, satisfying the stem's reporting requirement. It highlights expired or pending attestations needing escalation, complementing aggregate risk scoring with governance assurance.

Aggregate risk score per business unit.

Aggregating risk scores per business unit gives senior management a comparable, portfolio-level view of AI exposure, satisfying the stem's requirement for board-appropriate visualisation. It surfaces which units carry the greatest residual risk, enabling prioritised investment and accountability without overwhelming executives with model-level detail.

Clearly defined RACI matrix for AI initiatives

A clearly defined RACI matrix assigns who is responsible, accountable, consulted and informed for each AI initiative, removing ambiguity over ownership. This satisfies the coordination requirement by ensuring every risk decision has a named owner across functions.

Standardized AI risk taxonomy

A standardized AI risk taxonomy supplies common categories and definitions so legal, security, data science and business teams classify and compare risks consistently. Without it, cross-functional coordination breaks down because each function describes the same risk differently.

A centralized AI risk register

A centralized AI risk register provides the shared, authoritative record where risks identified by different functions are logged, tracked and owned. It satisfies the coordination requirement by giving all teams one source of truth, preventing duplicated or orphaned risks across silos.

Recording full data lineage for every training run.

Full lineage records each dataset, transformation and version feeding every training run, so corrupted or drifted inputs can be traced to their source and excluded. This directly satisfies the stem's data quality risk constraint by making provenance auditable rather than assumed.

Automated outlier detection to identify noisy or corrupted records.

Automated outlier detection flags anomalous records before they reach model training, directly satisfying the stem's requirement to minimise data quality risk. By identifying noisy or corrupted entries at the pipeline stage, it prevents skewed gradients and biased outputs, ensuring the Training pipeline consumes only validated, representative data rather than propagating defects downstream.

Implementing strict schema validation to catch data type errors.

Strict schema validation enforces declared column types and structures at ingestion, rejecting malformed records before they reach model training. This directly satisfies the stem's data quality risk constraint by preventing type mismatches and corrupt values from silently degrading the Training pipeline's output.