Architecture decision records for AI-enabled systems
A concise ADR structure for recording AI system boundaries, evaluation choices, fallbacks, and operational responsibilities.
Capture the decision in context
An architecture decision record should explain the workflow problem, users, constraints, considered options, and why the selected boundary is appropriate. For AI-enabled systems, include uncertainty and the consequences of a plausible wrong answer.
Add evidence and operations
Record the evaluation method, representative scenarios, acceptance threshold, data boundaries, human review, fallback behaviour, monitoring, and the person accountable for changing the decision.
- Decision and status
- Context and options
- Quality and risk evidence
- Security and data boundaries
- Fallback and operational owner
Keep ADRs alive
Link the record to tests, dashboards, runbooks, and follow-up decisions. Mark superseded records rather than rewriting history so future maintainers can understand why the system evolved.