Model Deprecation Procedure
Define the process for retiring AI models from production, including notification, data handling, audit trail preservation, and transition planning.
Objective
Ensure AI models are decommissioned in an orderly manner that preserves accountability, protects audit records, and minimizes disruption.
Maturity Levels
Initial
Models are retired informally with no documented procedure.
Developing
Engineers handle deprecation ad hoc; data and audit records are sometimes lost in the process.
Defined
A documented deprecation procedure covers data retention, audit trail preservation, stakeholder notification, and transition timelines.
Managed
Deprecations are tracked in a central log; compliance with retention requirements is verified before shutdown.
Optimizing
Deprecation is planned at deployment time; data and audit obligations are documented in the model card and carried through to retirement.
Evidence Requirements
What an auditor or assessor would expect to see for this control.
- —Deprecation notice records showing stakeholders were informed with the required lead time
- —Migration plan documentation and records showing dependent users were supported through migration to a replacement system
- —Data retention decision records for training data, model artifacts, and decision logs post-deprecation
- —Decommission confirmation records including final shutdown date, infrastructure teardown, and access revocation
- —Post-deprecation audit confirming all active dependencies and credentials were severed
Implementation Notes
Key steps
- Before retiring a model, inventory its associated data: training sets, decision logs, audit trails, and model weights — each may have different retention requirements.
- Notify stakeholders (internal teams, affected business processes, potentially regulators) before deprecation, not after.
- Preserve audit logs from deprecated models for the full required retention period even after the model itself is shut down.
- For vendor model deprecations, start your migration process when the vendor announces deprecation — end-of-life dates are rarely extended.
Example Implementation
ML team deprecating an NLP classifier after migrating to a new generative model
Model Deprecation Checklist — Intent Classifier v1.x
Pre-deprecation inventory:
- Training dataset: customer-support-intents-2024-v2 — retain 7 years per AI Act; archived to cold storage
- Decision logs (18 months of production): retained in compliance log store; TTL set to 2033-05-01
- Model weights: archived to model registry cold tier; not deleted (reproducibility requirement)
- Model card: archived with weights
Stakeholder notifications sent:
- Customer Support Product Lead — 30-day notice given 2026-04-01
- Integration team (3 downstream consumers) — migration guide shared
- Compliance team — deprecation and retention actions documented
Compliance verification: DPO confirmed no active data subject requests or litigation holds affecting this model's logs before proceeding
Shutdown date: 2026-05-15 Traffic cutover: 2026-05-01 (14-day parallel run period) Decommission confirmed by: Engineering Lead + Compliance Lead
Control Details
- Control ID
- CHM-005
- Domain
- Change Management
- Typical owner
- AI Engineering / Compliance
- Implementation effort
- Low effort
- Agent-relevant
- No