# Model deprecation runbook · [Product] **Owner:** [PM] **Applied AI lead:** [name] **Last reviewed:** [Date] ## Standing preparation (before any notice) - [ ] Model register: every model dependency by feature, provider, channel, region, version or alias, and customer tier - [ ] Each provider's published lifecycle states, notice periods and "not sooner than" dates recorded - [ ] Subscriptions to provider notices; status fields checked on a schedule - [ ] Eval suite per feature, including regression sets from past incidents - [ ] Model abstraction so a swap is a configuration change plus evals - [ ] Decision per feature: pinned version or alias, and why ## When a notice arrives | Step | Action | Owner | Done | |---|---|---|---| | 1 | Log the notice: model, channel, region, retirement date | | | | 2 | Set the internal deadline at least [30] days before retirement | | | | 3 | List affected features, customers and contracts (including AI risk documentation commitments) | | | | 4 | Shortlist candidate replacements | | | | 5 | Run the eval suite on representative data; compare quality, latency and cost together | | | | 6 | Adjust prompts, tools and parameters for the chosen candidate | | | | 7 | Shadow or canary with a limited set of accounts | | | | 8 | Notify customers if behavior, data handling, region or price changes; update AI documentation | | | | 9 | Cut over with a tested rollback path | | | | 10 | Record results; update the model register and documentation | | | ## Candidate comparison | Candidate | Quality (eval pass rate) | Regression set result | Latency | Cost per task | Data handling or region change | Decision | |---|---|---|---|---|---|---| | | | | | | | | ## Escalation If no candidate meets the quality bar before the internal deadline: [options, e.g., degrade the feature gracefully, restrict it to supervised mode, request a provider extension where offered, notify customers].