Cube A Cloud

Data Sensitivity

How sensitive is the data processed by this system?

1Public data only — no personal, financial, or confidential data.2Internal non-sensitive data.3Confidential business data.4Personal data (names, contact details, behavioural data).5Special category data: health, biometric, financial, legally privileged.

Decision Reversibility

How easily can decisions based on this system's output be undone?

1All decisions are easily reversible within minutes.2Most decisions reversible within hours.3Decisions reversible but at significant cost or delay.4Some decisions are difficult to reverse (e.g. communications sent).5Decisions are irreversible: financial transactions, legal filings, physical actions.

Audience Vulnerability

Who is affected by this system's outputs?

1Internal staff making routine operational decisions.2Internal staff making consequential decisions.3External business customers or partners.4General public.5Vulnerable individuals: children, patients, financially stressed, legally exposed.

Refusal Coverage

How well-defined are this system's refusal conditions?

1No refusal conditions defined.2Refusal conditions discussed but not documented.3Documented for some use cases.4Documented, tested, and assigned to a named owner.5Documented, tested, automated where possible, and reviewed on every model change.

HITL Maturity

How robust is human oversight for this system?

1No human review of outputs.2Ad hoc human review with no defined process.3Human review defined for high-stakes outputs.4Human review enforced and logged for all consequential outputs.5Review logged, reviewer qualifications verified, SLA-tracked.

Model Transparency

How well understood is this system's decision-making logic?

1Black-box model; no explainability mechanisms.2Limited transparency; general model type known.3Partial transparency; feature importance documented.4Explainability mechanisms deployed for high-stakes outputs.5Full audit trail: inputs, reasoning, outputs, and reviewer decisions logged.

Regulatory Exposure

Is this system subject to AI-specific regulation?

1No regulatory requirements applicable.2General data protection requirements apply.3Sector-specific guidance applies (not mandatory compliance).4Subject to mandatory AI regulation (e.g. EU AI Act high-risk category).5Subject to mandatory regulation and active regulatory scrutiny.

Third-Party Dependency

How dependent is this system on external providers?

1Fully self-hosted; no external dependencies.2Minor external APIs; easily replaceable.3Key inference provided by a third party.4Core functionality depends on a single external provider.5Mission-critical dependency on a third party with no fallback.

Drift and Degradation Risk

How likely is this system to degrade or produce harmful outputs over time?

1Static rule-based system; no drift possible.2Low drift risk; monitored monthly.3Moderate drift risk; monitoring in place.4High drift risk; daily monitoring required.5Very high drift risk; real-time monitoring required and output distributions shift significantly with data changes.

Scale and Reach

How many decisions does this system influence per day?

1Fewer than 10 decisions per day.210–100 decisions per day.3100–1,000 decisions per day.41,000–10,000 decisions per day.5More than 10,000 decisions per day.

Adversarial Exposure

How exposed is this system to adversarial or malicious inputs?

1No external input; internal data only.2Limited external input; low adversarial surface.3Moderate external input; prompt injection possible.4High external input; adversarial attack likely.5Directly accessible to adversarial actors with financial or reputational motivation.

Incident History

What is this system's track record?

1No incidents in the past 12 months.2Minor incidents only; resolved quickly.3One or more significant incidents in the past 12 months.4Repeated significant incidents; root cause not fully addressed.5Systemic failure pattern; incidents escalated externally.

AI Risk Worksheet.