The AI system development life cycle
Seven stages from plan/design to decommissioning, with governance hooks at each. The life cycle is iterative, not linear and building AI is never a one-time process. Decommission when the use case is gone, value has dropped, or better tech replaces it.
The AI system development life cycle has the same broad phases as any tech project, plus a data obsession and relentless monitoring. Crucially, it is iterative, not linear → developers revisit earlier stages when data, business, tech, regulatory or economic conditions change, or when user feedback demands it.
- Plan & design
- Collect & prep data
- Develop model
- Test & evaluate
- Deploy
- Monitor & maintain
- Decommission
| Stage | Governance requirements |
|---|---|
| Problem definition & planning | Consider the user group · consider using an interpretable model |
| Data collection & preparation | Data must be representative of the problem · prevent bias in data labelling |
| Model development | Explainability by design · appropriate reporting and documentation |
| Testing & evaluation | Test for bias, maintain fairness principles · user testing and representation |
| Deployment | Enable user feedback · reporting function for incidents and errors |
| Monitoring & maintenance | Monitoring and reporting schedule · regular quality checks · action plan for taking the model offline or retraining |
| Decommissioning | Sensitive data archived or destroyed lawfully · document the process |
when → the use case is no longer needed, the system no longer delivers value, or it is replaced by better technology. It is not just a technical task → it requires a comprehensive evaluation of the system's impact and the implications of removal. Building an AI system is never a one-time process → continuous monitoring and refinement are mandatory.
Module 1's official takeaways: (1) definitions evolve - know the 7 common elements; (2) the ladder - ANI exists everywhere, broad AI is the intermediate step, AGI and ASI remain hypothetical; (3) the OECD framework's five dimensions classify AI and anchor risk assessment; (4) model types - classic vs generative, proprietary vs open source, small vs large, language vs multimodal; (5) the life cycle is seven iterative stages, keep data representative and unbiased, monitor forever, decommission carefully. Mnemonic bank → A COD SHiP (7 governance-critical characteristics) and PEDMT (5 OECD dimensions).