# Workflow automation maintenance: failed runs, retries and support scope

> Scope automation support around failed-run alerts, safe replay, credential expiry and named owners. Separate an incident from a new feature.

Source: https://www.fyadvisers.com/blog/workflow-automation-maintenance-failed-runs/  
Author: Fiona Wang, FYAdvisers  
Published: 2026-11-08  
Last updated: 2026-11-08

A workflow is not finished when the demonstration succeeds. Credentials expire, APIs return too many requests, and someone changes a source field. Maintenance is the agreement about how a failed item is found, who responds and how work resumes without a duplicate transaction.

[System integration](/services/system-integration/) already depends on visible failures. This page is the operator runbook and the support boundary. What the work costs to keep running belongs on the [quote worksheet](/blog/workflow-automation-cost-singapore/).

## Four kinds of failure

| Kind | Examples | What a retry should do |
|---|---|---|
| Transient | Timeout, HTTP 429, a brief outage | Bounded retry, preferably exponential back-off, then a person |
| Invalid business data | Missing supplier, unbalanced total | No retry. A queue with the reason |
| Permission | Expired token, revoked scope | Alert the owner. Do not hammer the API |
| Downstream accepted, caller unsure | The bill posted, the flow timed out | Look up the business key. Do not post again |

Microsoft’s [error-handling guidance](https://learn.microsoft.com/en-us/power-automate/guidance/coding-guidelines/error-handling) documents run-after branches for failure, timeout, skip and success, scopes as a try/catch pattern, and retry policies with a fixed or exponential interval. It prefers exponential retry for transient faults, and it warns that excessive logging and alerting can itself become the problem. None of that makes a write idempotent. Idempotency is a property you design: the same business key, applied twice, has the same end state as applying it once.

## A key and a processed-item record

Choose a key the downstream system stores: invoice number within a supplier, purchase-order number, claim id, job id. Write a local row that says “accepted downstream” only after you have seen that key in the downstream system, not merely after your HTTP client returned. If the client times out, the recovery path reads the downstream system by that key.

The [overdue reminder](/blog/overdue-invoice-reminder-workflow-singapore/) version of this bug is a second email. The accounts version is a second bill. The [long-running approval](/blog/power-automate-long-running-approval-workflow/) version is a second valid decision. Same pattern.

## Logs a person can use

Log a correlation id, the stage, the result and a short error summary. Do not paste a full identity document, a bank account or a customer file into the alert. The PDPC’s [Key Concepts guidelines](https://www.pdpc.gov.sg/-/media/Files/PDPC/PDF-Files/Advisory-Guidelines/AG-on-Key-Concepts/Advisory-Guidelines-on-Key-Concepts-in-the-PDPA-17-May-2022.pdf) expect reasonable protection of personal data. A monitoring channel is a disclosure. Keep the payload in the restricted store and put a link, or only the id, in the message.

Alerts go to a named owner, with an escalation if that person does not acknowledge them. Agree the hours and the target in the contract. Do not describe a “standard SLA” this article has not measured. A manual queue is part of the design. If the queue can be ignored, the automation is optimistic.

## A concrete limit: Xero’s API

Xero’s [developer limits FAQ](https://developer.xero.com/faq/limits) describes, for calls against a tenant, a concurrent limit of 5, a minute limit of 60 and a daily limit of 5,000, plus an app-wide minute limit of 10,000. Exceeding a limit returns HTTP 429, and the FAQ describes remaining-allowance headers. Those figures are that FAQ’s. Recheck them for the app and the API you will call. They are a reason to batch and to back off, not a number to hard-code and forget. Other products publish different ceilings. Ask for the limit that applies to the connection on this project.

Microsoft also notes that a cloud flow which is not triggered for 90 days may be turned off, with exceptions for premium or Process capacity and with notice to owners. A flow that only runs at year-end can be switched off by inactivity. Put that on the runbook: a calendar event to open it, not a surprise in January.

## A run that never starts is not an error email

The four failure types above assume a run existed. A trigger that stops arriving does not create a failed step, so it may not create an alert. On Zapier, the notifications article updated 29 May 2026 says error emails go to the account email by default, that you can set the frequency to Never (which Zapier does not recommend), and that no error email is sent when an error handler runs. A company that inherited the builder’s inbox, or that turned alerts off to reduce noise, can look healthy while nothing runs. Check, on a schedule a person owns, that the expected record appeared. The [handover checklist](/blog/who-owns-your-automations-handover-checklist/) is how that inbox stays a company inbox. This is not a second article on monitoring. It is the gap between “we have error email” and “we would notice a silence”.

## What the retainer includes

Include credential rotation, a documented response to a connector change, dependency updates you have agreed, backups of definitions and of the business records, a restore drill, and a transfer of ownership when the maintainer changes. The [offboarding](/blog/employee-offboarding-access-revocation-workflow/) case, where the flow’s owner leaves, is a maintenance test as much as an HR test.

Exclude new features. Write the line between “the invoice did not post” and “please also send reminders”. The first is the retainer. The second is a change, priced on the [worksheet](/blog/workflow-automation-cost-singapore/).

## Tests

1. The remote system stores the document and the caller never receives the success. Replay does not create a second document.
2. A token expires outside the covered hours. The queue holds the item and a named person sees it when cover starts. Nothing is retried for the rest of the night.
3. The same invalid supplier id arrives repeatedly. It is quarantined once, not retried without limit.

Illustrative scenario, not a client engagement: an invoice post times out after the accounting system has accepted it. The operator searches the supplier and invoice number, finds the posted bill and marks the local row complete. The run button, pressed without that search, would have been a duplicate. The runbook’s first line is the search.

To turn the last invisible failure into a queue with an owner, use the contact form or WhatsApp +65 9071 3976 ([chat](https://wa.me/6590713976)).

## Frequently asked questions

### Who receives an alert, and what if they are away?

A named role, with a deputy, on a channel they actually watch. A shared inbox with no owner is how failures age. Response expectations belong in the support terms. This guide does not invent a standard response time.

### How is safe replay demonstrated?

With the acceptance case: the downstream system has already stored the business key, and the workflow did not record success. Replaying looks up the key and finishes the local record. It does not post a second document. Watching a happy-path rerun is not this test.

### What should a maintenance fee exclude?

New steps, a new system, a new legal entity, a changed approval matrix and historical cleanup, unless the contract says otherwise. Include credential rotation, a failed connector, a vendor API change you agreed to track, and a restore test.

### Should every error be retried?

No. Retry transient faults with a bounded, usually exponential, back-off. Do not retry a validation error, a permission denial or a payment that may already have been accepted. Quarantine those for a person.
