It’s pronounced “Jetpack”

Nov 7, 2026

5 min read

When an AI Workflow Times Out, Did the Work Still Happen?

Plan AI workflow recovery before launch: keep progress visible, check uncertain outcomes, prevent duplicate actions, and give unfinished work a clear owner.

By JTPCK Team

An AI workflow reads a request, prepares a record, and sends it to another application. Then the connection times out. The screen says something went wrong. Should someone press the button again?

That depends on which step stopped and what the destination already accepted. A second attempt might finish the job, create a duplicate, or overwrite work a colleague completed in the meantime.

We build AI integrations that connect models to existing business software. We recommend reviewing interrupted work before launch, alongside the quality of the model’s answers. Staff need to know what happened and how to continue without guessing.

Give each stage its own status

Consider a hypothetical repair company using AI to turn incoming service requests into proposed job records. A coordinator reviews the proposal before it enters the scheduling system. This is an invented workflow, not a JTPCK client account.

We would track the stages separately:

  • Request received.
  • Proposal prepared.
  • Coordinator review needed.
  • Approved for submission.
  • Submission attempted, result not yet confirmed.
  • Destination record confirmed.

Keep rejected and cancelled work visible too. A record awaiting review is different from one that may already have been created elsewhere.

Save progress in the application rather than relying on the conversation to remember it. The coordinator should be able to close the screen and return without losing the request or triggering another submission.

Treat an unknown result as unknown

A timeout limits how long one system waits. It does not prove that the other system did nothing. AWS’s retry guidance explicitly warns that an operation can have effects even when the caller receives a timeout or failure.

For our repair company, the scheduling system may have created the job just before the response was lost. Calling the operation again without checking could create a second job.

We recommend retaining the request identifier, intended destination, approved values, and any destination reference received. Use the destination’s supported lookup or reconciliation process to establish the result. If that evidence is unavailable, leave the item unresolved for a person.

“Not confirmed” is a useful status. Relabeling it “not submitted” would turn an information gap into a false fact.

Make a retry refer to the same operation

Some APIs support an idempotency key: an identifier that lets the destination recognize another attempt at the same operation. When supported and used correctly, this can prevent a repeated request from creating another record.

Stripe documents this behavior, including different handling across its API versions. Its API v1 guidance also describes parameter matching and eventual removal of stored keys. Those details matter; a key is not a permanent, universal duplicate-prevention guarantee.

For the hypothetical job record, retain the approved operation and its key before the first attempt. A retry should use that same operation, not ask the model to generate a fresh interpretation of the original message.

A revised appointment or different customer is a changed proposal. It needs its own checks and appropriate approval. Our guide to human approval in AI workflows covers keeping that approval tied to the action.

Where the destination offers no reliable retry or lookup mechanism, keep submission manual until there is a defensible recovery process.

Do not lose the work between two systems

Another failure can occur before a remote request is sent. The application saves an approved proposal, then crashes before recording the task that will submit it. The screen remembers the approval, but nothing is scheduled to happen next.

AWS describes the transactional outbox pattern for this gap. The application saves a database change and its pending event in the same transaction. A separate process later delivers the event. Delivery can still repeat, so the receiver must handle duplicates.

A business owner does not need to choose that implementation. Ask for a demonstration of the result: after an interruption, the approved item is still accounted for, with a visible pending, uncertain, or confirmed outcome. It should not disappear between those states.

The right mechanism depends on the existing software. A reliable task queue already built into the application may be enough.

Put limits on automatic recovery

Immediate, unlimited retries can worsen an overloaded service. AWS recommends spacing attempts, adding variation so clients do not all retry together, and limiting how many attempts occur. Check retry behavior across the whole integration, including any automatic retries in client libraries.

Longer outages need another response. Microsoft’s circuit-breaker pattern temporarily stops calls to a dependency that is likely to keep failing. The application still needs somewhere to put the unfinished work.

For our repair company, we would show the coordinator the affected requests, how long they have been waiting, and the next permitted action. Set an owner and an escalation time appropriate to the business. Do not leave every item behind the same indefinite spinner.

Also define the fallback before an outage. Switching to another model or service may change where customer information goes. Recovery should stay within the approved data destinations and action permissions.

Test the manual queue and the restart

Use synthetic requests and a test destination. Interrupt the workflow after a proposal is approved, after submission begins, and after the destination creates a record but before the application saves confirmation.

Then rehearse what staff actually do: reopen the item, click twice, resume after a restart, or take over manually. If someone creates the job by hand, the automation needs to recognize that resolution before it resumes.

Check that cancelled work remains cancelled and that a changed proposal does not inherit approval for an older version. Verify both the source queue and the destination records. A reassuring status message alone cannot establish that there is no duplicate.

Keep a short recovery guide beside the queue: what staff can safely retry, what needs a destination check, and who owns unresolved cases. For a low-volume workflow, a clear manual handoff may be more practical than elaborate automatic recovery.

Tell us which task you want to automate and where its result needs to go. The useful first design includes what happens when the connection breaks halfway through.

ai ai-integrations workflow-automation reliability

Keep Reading

More from the studio.

All Posts