Adoption gets attention; retirement does not. That is why decommissioning is where most of the residual AI risk in mature organisations actually lives.
The economics of organisational attention reward launches and punish endings. A new AI tool generates announcements, dashboards and momentum. A retiring one generates forms, conversations and a quiet ticket nobody wants. The predictable outcome is that tools are left running long after their value has gone, or are switched off carelessly, or are replaced without anyone confirming that the data has actually left.
ISO/IEC 42001 explicitly includes the disposal of AI systems in the lifecycle controls for the same reason ISO/IEC 27001 has always included asset disposal: this is where confidentiality fails after the project is forgotten (ISO/IEC 2023).
A defensible AI retirement covers six categories. Skipping any one of them produces a predictable class of incident.
1. Data return and deletion. Every dataset you provided — prompts, documents, fine-tuning corpora, evaluation sets — needs to be returned to you in a usable form, and then deleted by the supplier under contractual certification. Without a destruction certificate naming the systems and dates, you cannot evidence that the data is gone.
2. Model artefacts. Fine-tuned weights, embeddings, vector indexes and any custom retrieval stores derived from your data must be destroyed alongside the raw inputs. Embeddings have repeatedly been shown to allow partial reconstruction of source content (Song and Raghunathan 2020); they are not anonymous.
3. Downstream dependencies. Inventory every system, dashboard, automation and human workflow that consumed the retiring tool's output. Some will break loudly; some will degrade silently. The silent ones are dangerous.
4. Identity and access. API keys, service accounts, OAuth grants and on-behalf-of tokens issued to the retiring tool must all be revoked. A surprising number of historic breaches trace back to forgotten credentials of decommissioned systems (Verizon 2024).
5. People and process. The humans whose work depended on the tool need an alternative path before, not after, the switch off. Training, communication and fallback procedures matter more than any technical step.
6. Records. The retirement itself is an event. The risk register, the asset register, the supplier register and the change record all need a closing entry. For regulated industries, the model card, the evaluation history and the incident log should be archived in a form readable years later.
The cleanest decommissioning is a tool you procured. The hardest is an AI feature that arrived inside a SaaS product, was used by a few people and now needs to be turned off — but the SaaS contract makes no distinction between the AI feature and the rest of the product. Three steps usually work:
ISO-STANDARD.app carries each AI system through its full lifecycle — inventory, classification, controls, monitoring and retirement — with the evidence trail a regulator would expect.
ISO-STANDARD.app ships a ready-to-adopt ISO 42001 workspace with the risk register, controls catalogue, policies and audit-ready exports already wired together — no spreadsheet sprawl, no consultant lock-in.
Prefer a conversation? Email hello@iso-standard.app — a real human responds within one business day.
