Skip to content
  • There are no suggestions because the search field is empty.

ISO 42001: 2023 - A.6.2.4 AI System Design

This article provides guidance on how to implement the ISO 42001: 2023 A.6.2.4 AI System Design

ISO 42001 Control Description

The organisation shall ensure that AI systems are designed in accordance with documented requirements and that design decisions are recorded, justified, and reviewed for alignment with the organisation's AI objectives, risk appetite, and applicable legal and ethical obligations.

Control Objective

To establish a disciplined and documented AI system design process that translates requirements into a coherent technical architecture, supports the identification and mitigation of design-stage risks, and provides a traceable record of design decisions that can be examined throughout the AI system lifecycle.

Purpose

The design phase of AI system development is a critical determinant of system capability, reliability, and risk. Decisions made during design — concerning model architecture, algorithmic approaches, feature engineering, data pipeline construction, and integration patterns — establish parameters that are often difficult or costly to revise once implementation is underway. Poor design choices can introduce risks that are not apparent until the system is deployed, including risks related to fairness, robustness, interpretability, and failure behaviour.

This control recognises that AI system design must be approached systematically and with deliberate attention to the full range of requirements and constraints that govern the system. Design activities shall be grounded in the requirements established under A.6.2.2, informed by the risk assessments conducted under A.6.1, and conducted in a manner that supports subsequent verification, validation, and audit activities.

Documented design rationale also supports knowledge continuity within the organisation, enabling future maintainers and auditors to understand why design decisions were made and to assess their continued appropriateness as the system's operational context evolves.

Guidance on Implementation

Design Scope and Architecture

The organisation shall develop and document an architectural design for the AI system that describes its principal components, the relationships and interfaces between them, and the manner in which data flows through the system. The architecture shall address the AI model or models at the core of the system, data pre-processing and post-processing components, integration interfaces with external systems, and operational infrastructure.

The architectural design shall demonstrate traceability to the requirements specification, showing how design choices address documented functional requirements, performance criteria, and constraints.

Algorithmic and Model Selection

The selection of algorithms, model types, and training approaches shall be justified with reference to the requirements of the intended use case, the characteristics of available data, and the risk profile established through assessment activities. The organisation shall document the rationale for algorithmic choices, including consideration of alternatives and the reasons for their rejection.

Where the intended use involves high-stakes decisions or significant impacts on individuals, the organisation shall give particular consideration to the interpretability and explainability properties of the chosen approaches, and shall document how these properties align with transparency and accountability requirements.

Fairness and Non-Discrimination by Design

Design activities shall explicitly address fairness considerations relevant to the intended use context. The organisation shall identify the fairness criteria applicable to the system, document how design choices are intended to achieve those criteria, and consider the potential for design decisions to introduce or perpetuate discriminatory outcomes.

Where trade-offs between fairness criteria and other performance objectives are identified, these shall be documented along with the rationale for the approach taken. Fairness requirements from the impact assessment conducted under A.6.1.1 shall be reflected in design decisions.

Robustness and Resilience

The design shall incorporate measures to support system robustness, including resilience to variations in input quality, handling of out-of-distribution inputs, and behaviour in the presence of adversarial conditions where relevant to the intended use. Failure modes shall be identified at the design stage, and the design shall include appropriate mechanisms for detecting and responding to system failures or degraded performance.

Where the AI system interacts with human operators or makes recommendations that inform human decisions, the design shall consider how the system supports human oversight and avoids inappropriate automation of consequential judgements.

Privacy and Security by Design

The design shall incorporate privacy and security considerations from the outset, rather than treating these as supplementary concerns. Privacy-enhancing design choices, including minimisation of personal data processing within the system architecture, shall be considered and documented. Security controls relevant to model integrity, data pipeline protection, and inference security shall be addressed in the design.

The design shall take into account the data protection impact assessments and security risk assessments that apply to the system.

Design Review

Design documentation shall be subject to formal review before implementation activities commence. Reviews shall assess conformance with requirements, alignment with risk assessment outcomes, and the adequacy of provisions for fairness, robustness, privacy, and security. Review outcomes, including any required design modifications, shall be documented.

Related Controls

  • A.6.2.2 – AI System Requirements and Specification: The design shall be demonstrably traceable to the documented requirements specification.
  • A.6.2.3 – Data for Development and Testing of AI Systems: Data characteristics and constraints shall inform data pipeline design and model selection decisions.
  • A.6.2.5 – AI System Implementation: The design documentation provides the authoritative reference for implementation activities.
  • A.6.2.6 – AI System Verification and Validation: Design documentation supports the planning of verification and validation activities and the development of test criteria.

  • A.6.1.2 – AI Risk Assessment: Design decisions shall be informed by, and shall address, risks identified through the risk assessment process.