As AI systems become increasingly central to safety-critical domains—automotive, robotics, and industrial autonomy—the question is no longer whether to apply safety standards, but how those standards translate into real engineering practice.

In a recent discussion with industry experts, we explored how guidance from ISO/PAS 8800 and ISO/IEC TS 22440 is actively influencing modern safety workflows, system architectures, and regulatory strategies.

The takeaway is clear:

These standards are not theoretical—they are actively reshaping how safety-critical AI systems are built, validated, and communicated.

1. From Standards to Structured Workflows

At Saphira, we don’t treat standards like ISO/PAS 8800 or ISO/IEC TS 22440 as standalone compliance checklists. Instead, we embed their principles directly into how safety workflows are structured and executed.

This includes:

Human-in-the-loop approval gates for all safety-critical artefacts
Full traceability across HARA, FMEA, and safety concepts
Structured inputs and outputs, replacing ad hoc documentation
Explicit handling of operational assumptions (ODD/OED)

This approach reflects core ideas from:

ISO/IEC TS 22440 (Clause 8, Annex A, Annex C) — tool classification, error detection, and confidence in use
ISO/PAS 8800 (Clause 10, Annexes F & G) — robustness, architectural measures, and handling of AI uncertainty

Rather than asking “Is AI allowed?”, the practical question becomes:

How do we constrain AI so that it operates within reviewable, auditable safety processes?

2. A Shift Toward Explicit Assumptions and OOD Awareness

One of the most significant impacts of ISO/PAS 8800—especially Clause 10 and Annexes F & G—is the emphasis on:

Explicit system assumptions
Operational design domain (ODD/OED) clarity
Out-of-distribution (OOD) behavior

Historically, many of these elements were implicit. Today, they are becoming first-class engineering artefacts.

In practice, this means:

Safety workflows now require clear documentation of system limits
Teams systematically explore edge cases and OOD scenarios
Perception systems are evaluated not just on accuracy, but on failure behavior outside training distribution

This is especially critical for:

Autonomous vehicles
Industrial robotics
Perception-heavy systems

3. OOD Detection as a Safety and Audit Mechanism

A particularly important concept is out-of-distribution (OOD) detection.

In simple terms:

In-distribution → scenarios the model was trained on
Out-of-distribution → novel or unexpected inputs

OOD detection provides a mechanism to:

Identify when the system is operating outside its known capabilities
Trigger fallback or safe-state behavior
Log events for audit and analysis

In real deployments, this has led to:

Structured logging of interventions and edge cases
Improved auditability of AI system behavior
Use of OOD signals as a qualitative safety metric

In some cases, teams have reported significant reductions in incident rates (e.g., up to ~90%) through:

Data collection during development
Simulation-based testing and mitigation
Iterative improvement cycles

These workflows align directly with the intent of ISO/PAS 8800.

4. Bridging AI Behavior to Safety Artefacts

A recurring challenge in AI safety is translating model behavior into:

Hazards
Safety goals
Requirements
Safety case evidence

Standards like ISO/PAS 8800 and TS 22440 are driving a shift toward:

Systematic linkage between AI failure modes and hazard analysis
Explicit reasoning about how uncertainty propagates into risk
Stronger connections between:
- HARA
- FMEA
- Functional safety concepts

In practice, this results in:

More consistent safety artefacts
Earlier identification of gaps or missing mitigations
Better alignment between engineering teams and assessors

5. Architectural Measures for Robustness

Annexes F and G of ISO/PAS 8800 emphasize architectural and development techniques for improving robustness.

These are already widely adopted in industry, including:

Architectural Measures

Model ensembles
Redundant perception pipelines
Reject classes / uncertainty thresholds

Development Techniques

Data augmentation
Adversarial or fault-aware training
Transfer learning

Explainability Methods

Saliency maps
Structural coverage metrics
Model introspection techniques

While some of these are standard in ML engineering, their formalization in safety standards is driving:

More consistent adoption
Better auditability
Stronger safety case arguments

6. Impact on Safety Cases and Regulatory Communication

Perhaps the most important shift is how these standards influence communication with regulators and assessors.

Key changes include:

Safety cases now incorporate explicit assumptions and limitations
OOD behavior is used to justify:
- System boundaries
- Residual risk
- Continuous improvement strategies
Metrics and evidence are becoming more structured and auditable

As highlighted in the discussion:

Regulatory and legal processes increasingly require not just internal claims—but publicly defensible evidence of how standards are applied.

This is driving demand for:

Clear documentation
Measurable safety indicators
Public-facing technical material

7. The Role of Tooling in Enabling This Shift

Finally, standards like ISO/IEC TS 22440 play a critical role in defining how tools—especially AI-assisted tools—can be safely used.

In practice, this means:

Treating AI as an assistive tool (AUL-B1/B2)
Implementing strong error detection mechanisms (AI-TD)
Ensuring outputs are:
- Structured
- Reviewable
- Traceable

This allows organizations to:

Scale safety workflows
Reduce documentation burden
Maintain compliance and auditability

🚀 Key Takeaway

Across automotive and industrial AI systems, we are seeing a clear shift:

From implicit assumptions → explicit, documented constraints
From manual enumeration → systematic exploration of edge cases
From fragmented artefacts → traceable, structured safety workflows

Standards like ISO/PAS 8800 and ISO/IEC TS 22440 are not just influencing compliance—they are actively reshaping engineering practice.

From AI Safety Standards to Practice