What happens when an automation fails? Checks your business needs
Business automation
The short answer
When automation fails, first establish what actually finished. A failed read, an action that partly completed and a timeout after a possible send need different responses. The workflow should leave unfinished work visible, alert someone who can act and allow a controlled pause. Retry only when repeating the action is safe. Reconcile the source and destination before resuming so delayed runs do not duplicate records, messages or work completed manually.
An error message is only the beginning of recovery. The practical question is whether a customer received something, a record was created, or a task is now missing, and how the team can establish that without guessing.
Distinguish failure, partial success and uncertain completion
| State | Example | First check |
|---|---|---|
| Known failure before action | A required record could not be read | Confirm no destination action was attempted |
| Partial success | A draft job exists but its internal notification failed | Find the existing draft before repeating the workflow |
| Uncertain completion | A send request timed out after being submitted | Inspect delivery or destination evidence before sending again |
| Business exception | A customer match is ambiguous | Obtain a decision rather than repeatedly retrying the same data |
A workflow-level failure does not prove that every step failed. Recovery needs to understand the individual actions, especially those that create something or communicate outside the business. The design should preserve enough information to make that distinction.
Make unfinished work visible in business terms
Keep a source reference, run reference, last confirmed step, destination reference where available and a clear status. A person should be able to find the affected job or request without interpreting a long technical log.
Statuses should distinguish waiting for approval, missing information, retry scheduled and needs review. A generic failed label may hide whether someone needs to correct data, renew access or inspect a possible duplicate. Successful completion should mean the business result was confirmed, not merely that the connection started.
Limit the private information stored in logs and alerts. A notification can identify the record and link to an appropriately protected view rather than repeating the full customer submission. Check that the people receiving alerts can actually open the relevant record and take the intended action.
Retry temporary problems without repeating consequences
Microsoft's error-handling guidance describes retry policies for temporary failures. A useful policy limits attempts and spaces them out. It should not repeatedly hammer an unavailable service or treat invalid data as a temporary outage.
Creating a record or sending a message requires more care than repeating a read. Where the destination supports it, a stable action identifier can help make a retry safe. The implementation must check what the destination actually guarantees, not assume that a platform's retry button prevents duplicates.
Stripe documents duplicate webhook events and tracking processed event IDs. This is a concrete example of repeated delivery to plan for. Your connection needs its own agreed way to recognize repeat events and avoid repeating the business action.
An illustrative recovery: the draft exists, the alert does not
Imagine a workflow that creates a draft job from an accepted request and then sends an internal review notification. The notification step fails. This illustrative example shows why rerunning the entire sequence can be the wrong response.
- Pause affected processing if additional runs could create more uncertainty.
- Find the accepted request and inspect the destination for a draft using its stable reference.
- Confirm the draft contains the expected details and has not been changed by staff.
- Record that draft creation succeeded; identify the notification as the unfinished step.
- Recover the notification only after checking whether it was already delivered or acted on.
- Reconcile the request as handled so a delayed event cannot create another draft.
If completion cannot be established, keep the item in review. A person may need to inspect the destination or speak with the team. Guessing that nothing happened is not a recovery plan.
Plan the alert, the pause and the manual alternative
Agree who checks urgent alerts and how less urgent exceptions are reviewed. Match the response to the consequence. A delayed internal summary may wait; a workflow that could send incorrect customer confirmations needs a faster stop. Avoid so many repetitive alerts that the meaningful one is overlooked.
The team needs a documented way to pause the connection and continue essential work manually. Record manual actions against the source reference and mark them for reconciliation. Otherwise the automation may repeat those actions when access returns or a backlog is replayed.
Also decide what happens if the usual person is absent or cannot access the automation account. Recovery should not depend on a private login or undocumented knowledge. Keep a brief recovery checklist close to the workflow instructions, with credentials stored separately and securely.
Test recovery, including work that never started
Before launch, test missing data, expired access, an unavailable destination, repeated events and partial completion. Use a controlled test environment or privacy-safe test records. Demonstrate the pause and manual process as well as the successful route.
Reconcile eligible source records against destination results after an interruption and at a frequency appropriate to the task. An automation can miss an event entirely, leaving no failed run to alert on. Compare expected work with actual records, not just the number of green status indicators.
Our automation projects include agreeing failure cases and the response within the workflow's scope. Bring the consequence you are most concerned about. Recovery requirements help determine what connection is sensible and what checks need to exist before it runs.
A few follow-up questions
Can every failed automation just retry automatically?
No. Missing permission, invalid data and an ambiguous customer match usually need a correction or decision. An uncertain send may already have succeeded. Automatic retry should be limited to actions whose repetition has been shown to be safe.
Is a failure email enough?
Only if it provides a usable route to the affected work and someone checks it. Important workflows also need visible unresolved items and reconciliation, because an event that never triggered may not generate a failure email.
Will reconnecting an account recover the backlog?
That depends on the product and workflow. Check whether missed events are retained, whether runs resume automatically and which actions staff completed manually during the interruption. Reconcile before replaying anything with external consequences.
Sources and further reading
Plan the failure path before the connection runs
Tell us which tools and actions are involved, and what a missed or repeated action would mean for the business. We can discuss the checks and recovery steps the workflow needs.
Discuss reliable automation See what our business automation service includes