Services
Custom SoftwareCRMs, portals, dashboards + internal tools. Automation + IntegrationsWorkflows, APIs + business logic. DataScraping, enrichment + monitoring. AIAgents, assistants + knowledge systems. WebWebsites, performance + conversion.
Company
AboutWorkJournal
Start a project →

How to Recover Failed Automation Runs Without Duplicates

When an accepted order times out on its way to scheduling, check whether the job already exists before deciding what to retry.

NM
Written byNathan MackenzieFounder · ZappFlow
Practical notes on software, automation, data, web and AI.

An accepted order should become a job on the scheduling board, but the automation timed out. Before you retry, search for the job using the original order reference. If the job exists, finish only the work that is missing. If you cannot reliably tell whether it exists, hold the run for manual review.

A timeout means the automation stopped waiting for a response. The job system may have saved the record before that happened. The run can therefore show a failed job creation step even though the job exists, while later steps, such as writing the job number back to the order, never ran. Zapier’s guidance on duplicate data describes how replaying an action after a timeout can create a second record.

The idea in one viewVISUAL GUIDE
NOT EVERY DECISION SHOULD BE AUTOMATICWHO DECIDES

After a timeout, check before you retry

AUTOMATE · Retry safely
Retry only if no job exists and no attempt is pending.

Clear rule, clear input and a safe next action.

REVIEW · Job exists
Link the existing job and finish only the missing step.

The system can prepare the decision, but a person should check it.

PERSON · Not sure
Hold the run and ask a staff member to check.

Judgement, sensitivity or consequence makes human ownership the right choice.

Search using the original order reference before choosing a path.

Start with the run history, not the Retry button

    • Open the accepted order. Record its original reference and confirm it still needs a job. A cancelled order or one already handled by staff should not be sent through again.
    • Open the failed run. Note the run ID, time, error message, last confirmed successful step and any steps marked skipped or not run. Check whether another run for the same order has succeeded or is still pending.
    • Check who is handling the order. Give one person responsibility for recovery and check whether an automatic retry is queued. Do not start a manual retry while another attempt could still create the job.

    Check the job system using the original order reference

    Search the job system for the exact reference the automation sent. A customer name is useful for confirming a match, but it is not a reliable lookup on its own: the same customer may have several orders. If you find a match, check the job details against the order and note the job ID.

    No search result is not necessarily proof that no job was created. Your view may exclude some records, the search index may lag behind new jobs, or the reference may not have been saved in a searchable field. Check the relevant views and permissions, allow for any known delay and use a more reliable record lookup if one is available. If the reference was not stored or several jobs match it, pass the case to someone who can check the destination records before anyone creates another job.

    Worked example: one timed-out order, three recovery decisions

    In this illustrative example, order SO-1042 was accepted. The run history shows a timeout at Create job and says Write job number back to order did not run. The blank job number on the order confirms the link is missing; it does not tell staff whether the job was saved.

    Recovery decision table

    • No job was created → Retry the unfinished step. A reliable check of the job records confirms there is no job for SO-1042, and no attempt is still pending. Once the cause has cleared or been corrected, retry job creation without repeating earlier successful actions. Verify the new job and write its ID back to the order if the retry does not do that for you.
    • A job was created before the timeout → Link it and finish the missing work. The job system shows one matching job for SO-1042, with a creation time consistent with the original attempt. Verify the details and use that job’s ID to update the order. If staff expect a scheduling notification, check whether it already went out before sending one.
    • Staff cannot reliably tell → Hold for manual review. The normal search returns nothing, but new jobs can appear late and staff cannot check the underlying records. Assign someone to confirm the outcome, keep the order visible on the recovery list and do not retry job creation while the outcome is uncertain.
    WHAT HAPPENS WHEN SOMETHING GOES WRONGBACKUP PLAN

    Recover a Timed-Out Order Without Duplicating Its Job

    01 · PROBLEM
    Order timed out

    The order was accepted, but job creation timed out. The job may still exist.

    →
    02 · FALLBACK
    Check before retrying

    Search by the original order reference; verify any match. If unsure, pause for a manual check.

    →
    03 · RECOVERED
    Choose the safe next step

    No job: retry creation. One matching job: link its ID. Unclear result: hold for review.

    OUTCOME
    Confirm the order links to the right job and finish any follow-up.
    Verified recovery

    Check what your retry button actually repeats

    Before using a replay option, confirm where it starts, whether it can use an existing job ID and which actions it will repeat. Zapier’s replay documentation distinguishes replaying failed steps from replaying an entire run; a full replay can repeat steps that already succeeded. Even replaying only the failed step is unsafe if that step was a timed-out create and you have not resolved whether it saved a job.

    If your tool cannot resume from the right point, use an approved manual fix for the missing action where possible. A temporary connection problem may justify a limited retry once it is safe. Invalid order data or missing permissions need correcting first. Microsoft’s retry guidance recommends finite retry limits for transient faults and warns that repeating actions can have unintended side effects.

    Close the recovery record only after checking both systems: the job exists for the right order, the order points to the right job ID and any required follow-up is complete. Record what was checked, what was retried or fixed manually and who owns any unresolved work.

    Make the next timeout easier to recover

    If this happens regularly, start by making the original order reference reliably searchable in the job system. Then check whether the destination can enforce one job per order or accept an idempotency key, a value you send again with a retry so the receiving system can recognise the same create request. It only protects you if the destination supports and honours it, and the retry uses the same key. AWS’s guidance on idempotent operations explains how this makes a repeat request safe after a missing response.

    A search before creation helps staff reconcile a failure, but it does not prevent two overlapping attempts from both finding nothing and creating a job. When assessing changes to your automation and integrations, test the failure you actually need to recover from: the job saves, the response never arrives and staff need to finish the order without creating a second job.

    ZappFlow · practical next step

    Fix an automation your team keeps rescuing

    Tell ZappFlow which systems are connected, what one failed run was meant to do and what staff found in each system. A redacted example can help us assess whether a configuration change is enough or the workflow needs a safer way to recover.

    Discuss your failed workflow