An LLM model deprecation should trigger a scoped migration: locate the workflows using it, test a supported replacement against their acceptance criteria, and agree an operating fallback before retirement. Your maintenance agreement should name who performs that work and who accepts the result. A clause covering API availability alone leaves changes in business behaviour unresolved.
The difficult boundary arrives after shutdown. Reverting to the previous model then depends on a service the provider has withdrawn. Budget for a supported alternative or a bounded manual process that keeps the essential work moving.
Establish which deadline applies to your deployment
Anthropic’s model lifecycle guidance distinguishes deprecated models, which remain functional, from retired models, whose requests fail. It also states that partner-operated platforms such as Amazon Bedrock and Google Cloud set their own retirement schedules.
That distinction belongs in the maintenance record. A retirement date copied from the direct API documentation does not establish the date for a deployment bought through another platform.
Record the provider, serving platform, exact configured model identifier and applicable notice. Include the person responsible for checking updates. Treat the configuration as evidence of what the application requests; confirm the serving platform’s documented rules where aliases or routing sit between that identifier and the actual model.
Then locate the consumers. Search application configuration and deployment settings, but also inspect scheduled jobs, fallback routes and low-volume tools. Anthropic documents a Console usage export broken down by API key and model. That provides a useful activity check alongside configuration inspection. A dormant monthly workflow can escape a short usage sample, so retain both views.
Buy acceptance work alongside the model change
Changing the model identifier establishes that the application can attempt a request to the replacement. The owner still has to establish whether the resulting work is usable.
Consider a hypothetical professional-services firm using AI to prepare account briefs. Its replacement model returns fluent summaries and the integration records successful requests. During review, the account team finds that unresolved commercial questions have become confident recommendations. The transport migration succeeded while the briefing behaviour changed.
Specify the evidence required to accept that migration:
| Maintenance deliverable | Evidence the buyer receives | Acceptance owner |
|---|---|---|
| Affected-workflow inventory | Configured targets, serving platforms and responsible teams | Technical owner |
| Replacement compatibility check | Supported request options and successful handling of the response by the application | Engineering lead |
| Behaviour comparison | Old and replacement outputs on the same retained cases, with reviewer decisions | Workflow owner |
| Operational trial | Observed completion time, failures and review effort under the proposed configuration | Service owner |
| Retirement fallback rehearsal | Essential work completed with the retiring target disabled | Business continuity owner |
The worksheet is a proposed purchasing and acceptance tool, not a report of a migration Model Operator has executed.
Use the original pilot success criteria where they still describe the job. Add cases covering changes made since launch. Keep the inputs and scoring rules stable during the comparison; otherwise a model change and a test change become difficult to separate.
Anthropic’s lifecycle guidance recommends testing replacement models on your applications before retirement and links to its model-specific migration guides. Use the guide for the actual replacement to check compatibility. A recommended replacement identifies a migration destination; your own cases establish suitability for the workflow.
Put a boundary around remedial work
Ask how much application change the supplier has agreed to absorb under “model updates included”. Resolve that cost boundary while there is still time to negotiate.
Separate routine replacement work from application redesign in the agreement. A useful scope states which workflows are covered, who supplies representative cases, how many review rounds are included and what happens if the replacement fails acceptance. Record the escalation date early enough for another operating route to be prepared.
For the account-brief example, removing an unsupported request option belongs to compatibility work. Redesigning how the application distinguishes an unresolved question from an approved recommendation changes product behaviour. The agreement should establish how that work is authorised and priced before an approaching shutdown compresses the decision.
Reserve review time with the internal workflow owner before booking the cutover. In the briefing example, that person must inspect how the replacement handles an unanswered pricing question and reject any invented commitment. Until that review happens, the supplier has evidence of a working connection but no acceptance decision for the account team. Put the review date and its owner on the migration schedule.
Rehearse the service with the old model unavailable
A pre-retirement canary can use the old model as a temporary escape route. Record when that route expires.
For the post-retirement plan, choose a fallback whose availability and acceptance have been checked independently. Depending on the work, this can mean another supported model with narrower authority, a deterministic process for the essential subset, or a manual queue with an assigned owner and capacity limit.
Run a controlled rehearsal in a test environment with the retiring model disabled. Check whether the system still sends traffic to it through retries or fallback configuration. Follow a representative case through the supported route and inspect the business result. Include an exception that the fallback must decline or hand to a person.
This is a proposed acceptance exercise. It requires the team’s actual environment and has not been executed here. A successful rehearsal establishes evidence for the cases tested, rather than a guarantee of uninterrupted service.
Keep model replacement separate from recovering actions already taken. If an agent has changed a CRM record or sent a message, selecting another model does not reverse that effect. The AI agent rollback plan covers the action ledger, reconciliation and compensation decisions needed at that boundary.
Assign the decision before the retirement window closes
The buyer’s final decision is whether the replacement can serve the agreed work, whether a narrower service is acceptable, or whether the workflow should pause. Name the person authorised to make that decision and the evidence they will receive.
Model Operator is a Dubai-based AI product studio led by Alexander Chrisostomou. Its AI Initiative Consulting work can help leadership teams map workflow dependencies and agree implementation priorities; discovery determines the scope. No model choice removes the need for an internal owner who can accept the resulting work.
Bring the affected workflow and its maintenance terms to a Model Operator build conversation. Use that session to agree the migration boundary and the operating route that must remain available after retirement.