One request arrives, but two tasks appear. Two people receive the same approval, or a customer gets the same message twice. Staff start checking the automation manually because they no longer trust its output.
A duplicate task often points to a gap in how the workflow recognizes work it has already handled. Fixing that gap requires tracing the request through the whole process, including retries, concurrent runs, and partial failures.
This guide gives business owners and operations leads a way to define reliable behavior with the person maintaining a Power Automate cloud flow. The purchase-request example is hypothetical, and the design choices depend on the connectors and data store in use.
1. Find where the duplicate first appears
Start with one affected request. Record its source ID, the output task or approval IDs, the relevant times, and the flow runs associated with it. Compare the input and action history for those runs. Keep the original records available while investigating.
Ask whether the system received two separate submissions, whether two flows handled one submission, or whether a single run repeated an action. These situations need different repairs. Staff may also have copied a task manually after an apparent delay, so include the business process in the investigation.
Microsoft recommends designing flows to tolerate duplicate inputs and provides troubleshooting guidance for repeated trigger activity and actions. That principle is called idempotency. In practical terms, processing the same request again should preserve the intended business result.
Microsoft guidance on trigger problems and duplicate processing
2. Define what counts as the same work
A reliable design needs a stable identity for the source record. A request number or system record ID is usually more dependable than a person's name, an email subject, or the current time.
Then define the action being protected. Request PR-1042 might legitimately produce an approval, a fulfillment task, and a receipt confirmation. A key such as PR-1042:manager-approval:v1 identifies one specific operation. A separate key represents the fulfillment task.
Decide what should happen if the request changes after approval. A revised amount may require a new approval version, while a spelling correction may leave the existing approval valid. The process owner should define that rule before the flow builder implements it.
- Duplicate input. The same source request and operation arrive again.
- Legitimate revision. A business change requires a new operation with a new version.
- Separate request. Similar content belongs to a different source record and should proceed independently.
Write these distinctions into the workflow specification. A flow cannot infer whether a second purchase is accidental merely because its amount matches the first.
3. Store processing state where the next run can find it
Keep a durable processing record for each protected operation. Useful fields include the operation key, source record ID, status, output record ID, last attempt time, and the flow run reference. Restrict access according to the information being stored.
Possible states include Received, Processing, Completed, and Needs Review. Assign meanings to them. Completed should mean the required result was verified and its identifier was recorded. Needs Review should tell an operator where to investigate and who owns the decision.
Use a storage system capable of enforcing uniqueness or an equivalent atomic claim appropriate to the workload. The implementation must handle competing attempts to create the same operation record. Searching for a key and then inserting a row leaves a race if two runs both search before either inserts.
A builder might use an enforced unique key in a suitable list or database, or a destination API's supported idempotency feature. Connector support, licensing, and failure behavior need checking before selecting a pattern. The business requirement is that competing runs cannot both believe they exclusively own the same operation.
4. Plan for the gap between creating an output and recording success
Suppose the flow successfully creates a task, but the next action fails before saving the task ID to the processing record. A later retry sees no recorded output and creates another task. A ledger by itself has not solved this partial failure.
When possible, include the source operation key in the destination record and make that destination enforce uniqueness or accept an idempotency key. Otherwise, define a reconciliation step that checks whether the output already exists before recovery proceeds.
For an external action with uncertain completion, such as an email send that times out, the system may need a human review path. The reviewer should be able to determine whether the intended side effect occurred. Decide this behavior explicitly, especially when an automatic retry could send another message or create another approval.
Microsoft's error-handling guidance describes Run after paths and ways to capture flow-run information. Those controls can route an uncertain result into a useful exception record with the source ID, action name, and run reference.
Microsoft guidance on error handling and run information
5. Use trigger rules and concurrency with a clear purpose
Define when the process should start. If a flow reacts to updates on its own source record, distinguish a business submission from a housekeeping update. Test whether the flow's status changes cause it to run again.
Concurrency controls determine how many runs can execute together. They can help control contention, but serializing one flow does not cover a copied flow, another application writing to the same destination, or a later replay. All writers must follow the same uniqueness rules.
Microsoft documents different trigger behavior, including polling and webhook triggers. A restart can also have consequences for pending events. Review the specific trigger before turning a problematic flow off and on as part of recovery.
Microsoft guidance on trigger behavior and concurrency
Record why each control exists. An unexplained concurrency setting is hard for the next maintainer to evaluate when request volume grows.
6. Work through a purchase-request example
For PR-1042, the business rule is one manager approval for each submitted request version. The flow obtains the source ID and version, constructs the operation key, and tries to claim that operation in the processing store.
- If the operation is already Completed, use its recorded output reference and finish according to the documented duplicate-handling rule.
- If another run is actively Processing it, follow the waiting or exception route defined for this process.
- If this run successfully claims the operation, create the approval using the supported protection or correlation mechanism.
- Save the output ID and verified status. If that save fails, route the operation to reconciliation before another attempt can recreate the approval.
- If the request amount changes enough to require approval again, create a new version under the business rule and explain what happens to the earlier approval.
A maintainer must also define how to handle a stale Processing record after a failed run. Review the output and run history before releasing it for another attempt. A fixed delay alone cannot prove that the first attempt produced no result.
7. Test failure and replay scenarios before release
A successful submission is one test case. Build a small acceptance set that represents the situations most likely to damage trust.
- Submit the same event twice and confirm the expected output count.
- Start competing runs for the same operation and confirm that only the authorized attempt proceeds.
- Interrupt processing after output creation and before the success record is saved.
- Replay a completed request and inspect its stored output reference.
- Change a field that requires a revised approval, then change one that does not.
- Verify the exception notification and ask the backup operator to follow the recovery instructions.
Record the source and output IDs for each test. Have the process owner confirm that the result matches the business rule, including when an operation should pause for review.
8. Measure reliability through the business result
Track duplicate outputs, unresolved operations, reconciliation age, and manual corrections. A green run history can coexist with incorrect work, so periodically compare completed source requests with their intended outputs.
The handoff should explain ownership, alerts, recovery, and how to change the key or revision rules. Include a backup maintainer and a safe test process. Our workflow ownership guide covers continuity when a maintainer leaves.
If duplicate processing keeps returning, J3 can help trace the failure, define the operation rules, repair the agreed workflow, and verify recovery. Learn about Power Automate workflow repair or start with business process and automation preparation.
Related reading
- SharePoint Permissions Cleanup Without Disrupting Staff
- Technology Change Management for Small Businesses: A Practical Checklist
Need help applying this?
Get duplicate workflow actions under control.
Trace the affected requests, repair the agreed behavior, and document how to recover safely.