Risk Score Glossary
A complete reference for all six AI risk dimensions, their scoring rubrics, and how they contribute to the overall risk level.
Overview
The Exona risk assessment scores a company across six dimensions that together characterise the risk profile of an AI-driven product. Each dimension has a defined scale and rubric. The scores are combined into an overall_risk_level using a weighted model.
All scoring is AI-generated based on enrichment data and any questionnaire answers provided. Each score is accompanied by a rationale field citing the specific evidence used.
Overall risk level
The overall_risk_level is a composite rating derived from all six dimension scores.
| Level | Meaning |
|---|---|
Low | The product uses AI in a limited or peripheral way with minimal consequence if it fails. |
Medium | Moderate AI integration with some potential for harm, partially mitigated by controls. |
High | Significant AI autonomy in a consequential domain; governance gaps are likely relevant. |
Very High | Core AI autonomy in a high-stakes domain with limited oversight and broad potential for harm. |
Dimension 1: AI Intensity
What it measures: How central is AI to the product? A company where AI is the core engine is more exposed than one using AI as a peripheral feature.
| Score | Label | Description |
|---|---|---|
0 | No AI | No AI or machine learning components. |
1 | AI-Assisted | AI provides recommendations or analytics; humans make all decisions. |
2 | AI-Enhanced | AI is a significant product feature but not the primary value driver. |
3 | AI-Driven | AI is the primary mechanism: the product's value depends on AI output. |
4 | Autonomous Core | AI operates autonomously as the central decision-making engine; the product cannot function without it. |
Dimension 2: Autonomy
What it measures: To what degree does the AI act without human oversight? Fully autonomous AI decisions are harder to audit and more likely to cause undetected harm at scale.
| Score | Label | Description |
|---|---|---|
0 | Human-in-the-Loop | All AI outputs are reviewed and approved by a human before any action is taken. |
1 | Human-on-the-Loop | AI acts independently, but a human can review and override within a short window. |
2 | Supervised Autonomy | AI acts autonomously in most cases; humans are only involved in high-stakes edge cases. |
3 | Fully Autonomous | AI acts autonomously in all or nearly all cases with no human review. |
Dimension 3: Domain Risk
What it measures: How consequential is the domain in which the AI operates? A model making medical diagnoses carries higher inherent risk than one recommending playlists.
| Score | Label | Description |
|---|---|---|
0 | Low | Decisions have no direct consequence to individuals (e.g. content personalisation). |
1 | Moderate | Decisions affect user experience or business efficiency, but errors are easily corrected. |
2 | Elevated | Decisions have meaningful financial or reputational consequences. |
3 | High | Decisions have direct financial, legal, medical, or safety consequences for individuals. |
Dimension 4: Blast Radius
What it measures: How broad is the potential impact if the AI system fails or produces incorrect outputs? A model with 10 million users has a much larger blast radius than one with 100 enterprise clients.
| Score | Label | Description |
|---|---|---|
0 | Contained | Impact is limited to a small, defined group (e.g. internal users only). |
1 | Limited | Impact affects a bounded set of customers or partners. |
2 | Medium | Failures could affect tens of thousands of individuals or multiple downstream systems. |
3 | Large | Systemic failures could affect hundreds of thousands or millions of people, or create cascading infrastructure risk. |
Dimension 5: Data & Content Risk
What it measures: How sensitive is the data the AI system processes or produces? Sensitive data increases the severity of breaches, model inversion attacks, and compliance failures.
| Score | Label | Description |
|---|---|---|
0 | Public Data | Processes only publicly available or non-personal data. |
1 | Basic Personal Data | Handles standard PII (name, email, address) with standard protections. |
2 | Medium | Processes financial, behavioural, or biometric data requiring heightened controls. |
3 | Highly Sensitive | Processes health records, legal records, criminal data, or other categories with strict regulatory requirements. |
Dimension 6: Control & Governance
What it measures: How robust are the company's AI governance and oversight mechanisms? Strong controls reduce the likelihood that model failures go undetected or uncorrected.
| Score | Label | Description |
|---|---|---|
0 | Strong | Published model cards, third-party audits, bias testing, explainability tools, and clear escalation processes are all evident. |
1 | Adequate | Some documented controls exist (e.g. monitoring, limited human review for edge cases). |
2 | Partial | Controls are sparse, inconsistent, or apply only to a subset of decisions. |
3 | Weak or Absent | No meaningful oversight or audit trail is evident. The AI system appears to operate without checks. |
Control & Governance is an inverted scale: a score of 3 (Weak or Absent) is the most concerning outcome, while 0 (Strong) is the most reassuring. This mirrors how underwriters think: better governance means lower risk.
Regulatory Exposure
regulatory_exposure is a free-text field, not a scored dimension. It summarises the regulatory frameworks that apply to the company's AI use based on its geography, sector, and product type.
Common frameworks that appear:
| Framework | Region | Applies when |
|---|---|---|
| EU AI Act | European Union | AI systems classified as high-risk (Annex III): includes AI in insurance, HR, credit. |
| Consumer Duty | UK (FCA) | AI used in customer-facing financial services decisions. |
| GDPR / UK GDPR | EU / UK | Automated decision-making that produces legal or similarly significant effects. |
| NAIC Model Bulletin | USA | AI use in insurance underwriting and claims. |
| CCPA | California | AI processing of California residents' personal data. |
Reading a risk assessment in practice
Here is an example of how to interpret the scores for an insurtech claims platform:
What this tells an underwriter:
- The AI is not supplementing a human: it is the decision-maker.
- 85% of decisions are fully automated with no human in the loop.
- Errors in a claims context are not just inconvenient: they may cause financial harm to policyholders and trigger regulatory scrutiny.
- The governance gap (no review for low-value claims) is the key mitigant missing. Requiring the insured to implement human review above a lower threshold, or add bias monitoring, would materially reduce this risk.
Score aggregation
The overall_risk_level is not a simple average. Dimensions are weighted with particular emphasis on ai_intensity and autonomy because these are the most diagnostic indicators of systemic AI risk. A very high autonomy score in combination with high domain_risk will push the overall rating to High or Very High even if other dimensions are favourable.
The exact weighting model is calibrated against historical insured loss data and is updated periodically. The dimension scores are always individually reported so underwriters can apply their own judgement.