AI decommissioning procedure

What it does

It closes the lifecycle deliberately. A retired system leaves things behind: processes that relied on it, copies of data, model files, credentials, logs, and duties that outlast it, such as explaining past decisions. The procedure accounts for each. The system owner is accountable; it operates whenever a system or model version is withdrawn.

How to implement it

• Identify every process and system that consumes the outputs, from the inventory.
• Agree the replacement or the manual process and the date of switch-over.
• Dispose securely of training data, weights and embeddings that are no longer needed.
• Revoke credentials and remove access.
• Retain the records required to explain and evidence past decisions, for the period that applies.
• Update the inventory and notify affected users.

Evidence it operates

Design: a decommissioning procedure and checklist. Operation: for a retired system, a completed checklist; evidence of data and artefact disposal; revoked credentials; retained records located and accessible; inventory status updated.

Where it fits

The final stage of the lifecycle, also applied when a model version is replaced. It is the primary control for uncontrolled AI retirement and completes AI change management.

Limits

Disposal is only as complete as the inventory of copies. Data held by a supplier depends on the contract and on the supplier's own practice. A model's influence can persist after retirement, in decisions already made and in other models trained on its outputs. Retirement is also when knowledge of the system is lost, so its documentation and the names of those who understood it should be kept with the records.