AI capabilities age quickly. Providers deprecate models, policies evolve, data sources change, and better options appear. The question is not whether you will deprecate AI components - it is whether you will do it deliberately or in a scramble.
Deprecation is a lifecycle control
Sunsetting should be treated as part of your operating model, alongside rollouts and incidents. A good deprecation process reduces risk and protects user trust.
Start with a registry and clear ownership
You cannot retire what you cannot inventory. Maintain a registry that maps: model/provider versions, prompts, tools, and which products and tenants rely on them (see model registries and service catalogs).
Define deprecation triggers
Common triggers include:
- Vendor end-of-life notices or region changes.
- Security findings or unacceptable incident patterns (see incident response).
- Better quality or cost economics elsewhere (see vendor exit strategy).
- Policy updates that require different controls (see policy layering).
Use staged timelines, not surprise switches
A practical deprecation timeline often includes:
- Notice. Announce the change, the reason, and the impact.
- Dual-run window. Run old and new side-by-side for critical workflows to validate outcomes.
- Migration assistance. Provide guidance, updated prompts, or changed integrations.
- Cutover and rollback path. Switch with a clear rollback plan if quality degrades.
- Removal. Remove the old path, clean up routing rules, and revoke access.
Feature flags and routing are your main levers for safe transitions (see feature flags and routing and failover).
Communicate changes like product releases
Users care about behaviour. Even if the output is "better", tone, refusals and citations may change. Use release notes and clear messaging so customers are not surprised (see AI release notes).
Stabilise if signals degrade
When migration signals worsen (higher escalation, more refusals, more incidents), pause and stabilise instead of pushing through. A short change freeze can protect trust (see change freeze playbooks).
Deprecation done well is invisible to users. Deprecation done poorly becomes an incident.