The AI-FMEA system uses three engineering scales—Severity, Occurrence, and Detection—to evaluate how conversational AI systems can fail, how often those failures may appear, and how easily the system can detect them before they reach the user. These scales allow developers to prioritize risks, identify weak points in the system, and implement corrective actions based on measurable criteria.
This reference page provides clear, practical definitions of each scale and explains how they apply to AI behavior patterns. For deeper regulatory or academic analysis, a full technical specification is available below.
Severity (S)
Severity measures the potential impact on the user or society if the failure reaches them.
In conversational AI, Severity reflects outcomes such as:
- Emotional manipulation or dependency
- Harmful guidance or misinformation
- Trust degradation or exploitation
- Legal, ethical, or reputational consequences
Scores generally range from Minor (1–3) to Critical (9–10).
Higher scores indicate outcomes that may cause irreversible harm, escalate rapidly, or bypass user awareness.
Severity answers the question: “How bad is the consequence if this failure happens?”
Occurrence (O)
Occurrence estimates how likely it is for a failure mode to appear in real user interactions.
For conversational AI, Occurrence considers:
- Model tendencies and training patterns
- Reinforcement loops
- User behavior that triggers vulnerabilities
- Systemic biases or learned habits
- Recurrence of certain harmful patterns across sessions
Scores range from Rare (1–3) to Frequent/Systemic (8–10).
Occurrence answers the question: “How often will this failure happen if we do nothing?”
Detection (D)
Detection measures how likely the system is to recognize the failure before it reaches the user.
In AI systems, Detection is influenced by:
- Internal safety classifiers
- Guardrails and filters
- Monitoring systems
- Human oversight
- Transparency of internal reasoning
- Ability to interrupt or self-correct failure patterns
Scores range from High Detectability (1–3) to No Detectability (9–10).
Detection answers the question: “How likely is the system to catch this failure before the user is affected?”
The Heat Map
The Heat Map is a color-coded visualization of the Risk Priority Number (RPN), which is calculated by:
RPN = Severity × Occurrence × Detection
Higher RPN values indicate urgent or systemic risks requiring immediate mitigation.
The Heat Map helps teams quickly identify:
- Critical risks (red)
- Moderate risks (yellow/orange)
- Lower-priority risks (green)
It functions as a triage tool, guiding developers to focus on the highest-impact issues first.
Download the AI-FMEA Rating Scales & Heat Map (FMEA-05)
Use this reference sheet when completing the Master Template or reviewing FMEA results.