Fundamental Principles

Principles are the part of an architecture that survives the next reorganisation, vendor and model generation. Chapter 3 mapped what the law requires; this chapter sets out how to decide where the law leaves the choice open, which is most of the time. It proposes eight principles for AI Architecture, relating each to the principles of Architecture as Code (Chapter 2) from which it descends and to the AI Act (66, Nordqvist, 2025), for use as review criteria, as the backbone of an AI policy, and as the source of the checks that Chapters 8 and 9 build into governance gates.

A principle earns its place only if it changes decisions. Richards and Ford (2020) argue that architectural characteristics have to be explicit and measurable before trade-offs can be reasoned about (71, Richards & Ford, 2020), so each principle below is stated as an obligation on the design with at least one verification a machine can perform, in the spirit of the fitness functions of Architecture as Code (66, Nordqvist, 2025).

The principles also share a scaling rule. Regulation (EU) 2024/1689, the AI Act, applies obligations in proportion to risk, from prohibited practices in Article 5, through high-risk AI systems in Article 6 and Annex III, to transparency duties in Article 50 and minimal-risk systems with no specific obligation (1, European Union, 2024). The depth of evidence each principle demands follows the same gradient: a spam filter and a benefits eligibility system both need declared intent, yet only one needs it reviewed by a fundamental rights impact assessment.

Primary topics explored

1. What a principle has to do
2. Principle 1: Declared intent
3. Principle 2: Data and model as first-class architectural elements
4. Principle 3: Reproducibility and lineage
5. Principle 4: Bounded autonomy
6. Principle 5: Evaluation before and after deployment