Model lifecycle
Phases from design and data curation to decommissioning.
How the laws define this
Regulators and policy frameworks agree that a model lifecycle represents a structured sequence of phases throughout an AI model's existence, used to establish risk assessment, documentation, and governance controls. The primary divergence lies in how broadly these stages are defined and whether they emphasize technical processes or operational oversight. The AgID draft guidelines on AI adoption in the PA and Guideline E-23 — Model Risk Management (Office of the Superintendent of Financial Institutions) (final) establish broad end-to-end frameworks running from design and development through monitoring and eventual decommissioning, with Guideline E-23 explicitly incorporating institutional governance gates such as independent review, approval, and change management. Conversely, the National Artificial Intelligence Strategy (Ulusal Yapay Zekâ Stratejisi) 2021-2025 outlines a narrower technical progression of training, validation, and deployment, while the Government-wide Vision on Generative AI (Overheidsbrede visie generatieve AI) specifically tailors its lifecycle phases to generative models by focusing on pre-training, fine-tuning, and deployment.
Synthesised from the 4 statutory definitions below. Each one is quoted in full, with its source.
How jurisdictions differ
What is distinctive in each definition, binding law in force first. Showing 2 of the 4 instruments that define this term; all are listed under Definitions below.
| Jurisdiction | Instrument | Status | What is distinctive |
|---|---|---|---|
| Canada | Guideline E-23 — Model Risk Management (Office of the Superintendent of Financial Institutions) (final) | Awaiting Entry | Adds independent review/validation, approval, change management, retirement and documented controls at each stage to the sequence. |
| Netherlands | Government-wide Vision on Generative AI | Adopted | Limits lifecycle to three generative AI stages—pre-training, fine-tuning, deployment—and structures risk identification, documentation and governance measures. |
Definitions (4)
The end-to-end stages of an AI model's existence as defined by the guidelines, encompassing design, dataset curation, development, testing, deployment, monitoring, maintenance, updating and decommissioning; the lifecycle framework informs risk assessments, documentation requirements and controls applied at each stage.
The sequence of stages for a model including design, development, testing, independent review/validation, approval, deployment, monitoring, maintenance/retraining (where applicable), change management and decommissioning/retirement, with documented controls at each stage.
The lifecycle concept enumerating model development phases (training, validation, deployment) used in the Strategy's glossary to align technical, testing, documentation and monitoring practices across institutions and to frame risk assessment and governance measures throughout an AI model's operational life.
The stages through which a generative model progresses: pre‑training (initial large‑scale training on broad data), fine‑tuning (task- or domain‑specific adaptation using additional data or objectives), and deployment (operational use in services or products). The vision uses these phases to structure risk identification, documentation, and governance measures.
Not to be confused with
Model lifecycle stages limits to three phases—training, validation, deployment—while model lifecycle definitions include additional stages like monitoring and decommissioning.
AI lifecycle covers stages from planning through decommissioning for AI generally, whereas model lifecycle focuses on model-specific phases like development and deployment.
AI system lifecycle starts from data to monitoring for systems, differing from model lifecycle's emphasis on design, testing and decommissioning in some definitions.
Questions and answers
What stages are included in model lifecycle under OSFI Guideline E-23?
It includes design, development, testing, independent review/validation, approval, deployment, monitoring, maintenance/retraining, change management and decommissioning/retirement. Documented controls apply at each stage.
How does the Dutch Vision define generative AI model lifecycle?
It covers pre-training on broad data, fine-tuning for specific tasks or domains, and deployment in services. These phases structure risk identification, documentation and governance.
What is the model lifecycle in Italian AI guidelines for public administration?
It encompasses design, dataset curation, development, testing, deployment, monitoring, maintenance, updating and decommissioning. The framework informs risk assessments, documentation and controls at each stage.
Related Terms
Life cycle
The sequence of phases governing an AI system from design to decommissioning....
Model lifecycle stages
Phases of model development: training, validation, deployment....
AI life cycle
Stages from scoping through decommissioning....
AI lifecycle
Stages of AI from planning through decommissioning....
AI system lifecycle
Stages of AI development from data to monitoring....