An AI workflow needs a useful answer to a simple question: what happens when one of its dependencies stops responding? A model may be unavailable, a search index may be rebuilding, or a connected business system may reject a request. Those situations should not leave staff guessing whether work was lost, completed or duplicated.
A fallback is an alternative operating path with clear limits. It is not a promise that another model will produce an equivalent result. The objective is to preserve the task, protect the underlying records and give someone enough information to continue responsibly. Designing that path before launch is usually easier than inventing it during an outage.
Separate the failures you need to handle
Start by mapping the workflow into meaningful steps: intake, retrieval, generation, validation, approval and execution. Record what each step reads, what it can change and what evidence confirms completion. A timeout during drafting is different from a timeout after a payment or record update has been submitted.
Use failure categories that lead to different decisions. An unavailable dependency may justify a delayed retry. Invalid credentials require operational attention. Missing source evidence calls for a clarification or manual investigation. A rejected business rule should not be treated as a temporary technical problem that more retries will solve.
Also distinguish unavailable from slow. Set a deadline for the whole task, not only individual network requests. A chain of apparently reasonable waits can consume the entire service window. Once the deadline is reached, the workflow should move into a documented state rather than continuing invisibly in the background.
Preserve the task before attempting recovery
Give each accepted task a stable identifier and persist its approved inputs before calling external services. Include the request time, requester, intended outcome and relevant record references. Avoid storing unnecessary sensitive material simply because it might help later. The recovery record should be useful without becoming an uncontrolled copy of every source system.
Track progress at boundaries where it matters. For example, a quotation workflow might record that product evidence was retrieved, a draft was generated and approval remains outstanding. This allows recovery from the last safe checkpoint instead of starting over and potentially generating a second customer-facing document.
The user-facing status must reflect this record. Use distinctions such as waiting, needs review and completed only when the underlying evidence supports them. A cheerful success message after a request was merely queued creates extra support work when the expected outcome never arrives.
Make retries bounded and deliberate
Retries are appropriate only when another attempt can safely succeed. Set a maximum attempt count, a total time budget and a delay policy. Add variation to retry timing so many queued tasks do not all hit a recovering dependency at once. Record the attempts under the original task rather than presenting each attempt as new work.
Before repeating a write, determine whether the earlier request may already have taken effect. Use an idempotency mechanism supported by the destination or look up the intended result using a reliable business identifier. If the outcome is uncertain, route the item for reconciliation instead of blindly submitting the same action again.
Do not ask the model to decide whether a duplicate is harmless. That decision belongs in application logic and the destination system's transaction design. The generated explanation can help a reviewer understand the situation, but it should not replace the checks that prevent repeated changes.
Choose an alternative that preserves the controls
Possible fallbacks include a delayed queue, a simpler deterministic process, an approved secondary provider or a human handover. Pick the smallest alternative that still meets the task's requirements. If retrieval is unavailable, a knowledge assistant should not silently switch to unsupported general knowledge and present it as company policy.
A secondary model needs its own evaluation. Differences in formatting, instruction following and tool behaviour can affect downstream steps even when the wording looks acceptable. Keep the same validation and approval gates. If the alternative has not passed the relevant tests, limit it to drafting or use the manual path.
For a human handover, provide the original request, completed steps, evidence references, unresolved decisions and safe next action. Avoid dumping raw logs on the person. A concise recovery packet is more useful than a technical transcript that requires them to reconstruct the workflow from scratch.
Communicate without exposing internals
Tell the requester what has happened to their task and whether they need to do anything. A useful message might explain that the request is saved, an external service is unavailable and the team will continue through a review queue. Do not promise a recovery time unless someone has established one.
Separate customer communication from operational diagnostics. Customers do not need authentication errors, internal endpoint names or model responses containing source material. Support staff need a task reference and a trustworthy status. Engineers need detailed, access-controlled logs with correlation identifiers and dependency information.
Provide a way to check the task later. This reduces duplicate submissions from people who reasonably assume nothing happened. If a requester cancels while the workflow is waiting, propagate that decision before the next execution attempt and verify that no downstream action is already in progress.
Test the awkward recovery moments
Exercise failures at more than the first network call. Stop retrieval after some documents are returned. Let generation succeed but fail validation. Simulate a successful destination write followed by a lost response. Restart the worker while an approval is pending. These cases reveal whether your checkpoints represent real business states.
Measure recovery outcomes as well as error counts. Track tasks safely completed, tasks needing manual reconciliation, duplicate actions prevented and time spent in uncertain states. A system can report a low technical error rate while leaving a small but costly group of tasks stranded without ownership.
Run a rehearsal with the people who will receive escalations. Ask them to recover a sample task using only the information the workflow provides. Missing permissions, unclear instructions and inaccessible evidence often become visible immediately. Update the handover packet rather than assuming training will compensate for missing information.
Restore normal operation carefully
When a dependency returns, avoid releasing the entire backlog at once. Apply a controlled processing rate and prioritise by business need, not merely arrival order. Recheck cancelled requests, expired approvals and source records that may have changed while the task waited.
Keep a short recovery record covering the failure window, affected tasks, manual interventions and remaining exceptions. Use it to improve the workflow, not to claim that every failure can be eliminated. Reliable automation includes a credible plan for the moments when automation cannot continue.
For AI automation projects, fallback design belongs alongside the main workflow rather than in a later support document. A useful first deliverable is a one-page map of failure states, owners and recovery actions. If the team cannot explain those paths clearly, the implementation is not yet ready to operate without close supervision.