Skill Guide

Workplace hazard analysis for autonomous and semi-autonomous systems (HAZOP, FMEA adapted for AI)

The systematic process of applying structured hazard identification and risk assessment methodologies (HAZOP, FMEA) specifically to the design, development, and deployment of AI and autonomous/semi-autonomous systems to prevent operational failures and accidents.

This skill is critical for mitigating catastrophic safety, financial, and reputational risks inherent in complex AI-driven systems. It directly impacts business outcomes by enabling regulatory compliance, reducing costly recalls/litigation, and building stakeholder trust in AI products.

1 Careers

1 Categories

9.0 Avg Demand

15% Avg AI Risk

How to Learn Workplace hazard analysis for autonomous and semi-autonomous systems (HAZOP, FMEA adapted for AI)

1. Master the core concepts of traditional HAZOP and FMEA (e.g., guide words, severity/occurrence/detection ratings). 2. Study the unique failure modes of AI systems (data drift, adversarial attacks, model degradation, ambiguous output). 3. Learn the structure of AI/ML pipelines to understand where hazards can be injected.

1. Apply adapted frameworks to specific domains: e.g., FMEA for a computer vision module in a drone, or HAZOP for the decision logic of a warehouse robot. 2. Move beyond single-component analysis to system-of-systems hazards. 3. Avoid the common mistake of only analyzing the AI model in isolation, ignoring the data lifecycle, sensor fusion, and human-machine interface.

1. Architect organization-wide hazard analysis processes that integrate into MLOps and DevOps CI/CD pipelines. 2. Develop custom risk matrices and acceptance criteria for novel AI functionalities (e.g., generative AI outputs in safety-critical contexts). 3. Mentor cross-functional teams (engineers, ethicists, legal) on risk-informed design and champion a proactive safety culture.

Practice Projects

Beginner

Case Study/Exercise

HAZOP for a Warehouse Logistics Robot

Scenario

A company is deploying semi-autonomous mobile robots for pallet transport in a shared warehouse environment with human workers. You must identify initial hazards.

How to Execute

1. Define the system boundary: Robot navigation, obstacle detection, communication with central control. 2. Select a node (e.g., the obstacle detection system). 3. Apply HAZOP guide words (NO, MORE, LESS, AS WELL AS, PART OF, REVERSE) to the design intent (e.g., 'Detect obstacle' -> 'NO detection'). 4. Brainstorm potential causes, consequences, and existing safeguards.

Intermediate

Case Study/Exercise

FMEA for a Predictive Maintenance AI Model

Scenario

An AI model predicts machine failure for a critical industrial asset. The FMEA must cover the full ML lifecycle: data pipeline, model training, inference deployment, and maintenance action trigger.

How to Execute

1. Break down the system into subsystems/components: Data Ingestion, Feature Store, Model Training, Model Registry, Inference API, Alert Integration. 2. For each component, define potential failure modes (e.g., 'Stale data used for training'). 3. Assign Severity (S), Occurrence (O), and Detection (D) scores using an AI-adapted scale. 4. Calculate Risk Priority Number (RPN = S x O x D) and prioritize mitigation actions for high-RPN items (e.g., implement data drift monitoring).

Advanced

Project

Integrated Risk Assessment for an Autonomous Vehicle Perception Stack

Scenario

Lead a cross-functional team to assess the safety case for an AV perception system combining LiDAR, cameras, and radar, including the deep learning sensor fusion algorithm.

How to Execute

1. Establish a combined HAZOP-FMEA working group with systems safety, ML engineering, and V&V experts. 2. Use model-based systems engineering (MBSE) to create a functional architecture diagram. 3. Conduct a HAZOP on the high-level data flow and decision logic. 4. Perform a detailed FMEA on the fusion algorithm's failure modes (e.g., conflicting sensor inputs, latency-induced deadlock). 5. Synthesize findings into a unified risk register with traceable mitigation controls and verification activities.

Tools & Frameworks

Methodologies & Standards

IEC 61508 (Functional Safety)ISO/PAS 21448 (SOTIF - Safety of the Intended Functionality)Adapted HAZOP for software/AIAI-Specific FMEA

These are the foundational industry standards. SOTIF (SOTIF) is particularly vital for AI, addressing hazards from performance limitations and misuse, not just hardware faults. Use them as the canonical references for process and documentation.

Software & Platforms

Fault Tree Analysis (FTA) Software (e.g., Isograph, ReliaSoft)Simulation & Digital Twin Environments (e.g., CARLA, NVIDIA Isaac Sim)MLOps Platforms with Monitoring (e.g., MLflow, Evidently AI)Requirements Management Tools (e.g., IBM DOORS)

FTA software helps model complex causal chains of AI system failures. Simulation allows for controlled hazard injection testing. MLOps tools are essential for operationalizing the detection and mitigation controls identified in the analysis (e.g., monitoring for data drift).

Interview Questions

Answer Strategy

The candidate must demonstrate system thinking beyond a single model. The answer should break down the system into logical subsystems (sensor, perception, interface), identify interface-related failure modes (e.g., latency in bounding box output causing a planning overshoot), and emphasize the use of AI-adapted severity scales (e.g., 'collision' vs. 'minor deviation').

Answer Strategy

This tests influence, communication, and the ability to frame technical risk in business terms. The candidate should demonstrate how they quantified or articulated the risk (e.g., using a projected cost of failure from the risk matrix) and aligned it with long-term brand or compliance goals.