The most expensive automation failure is not an error. It is silence.
Why error monitoring misses it
A workflow that stops running produces no error. No log line, no failed run, no red icon. It simply is not there, and monitoring designed to react to failures has nothing to react to.
Platforms document error handling as a first class concern for good reason, and it is worth doing. It does not cover this case.
Why multi location makes it worse
In one office, somebody notices within days: the thing they expected did not happen.
Across thirty locations, nobody does. Each location assumes it is a local problem, three of them mention it separately to different people, and nobody compares. A failure affecting a tenth of your network can run for a month.
The usual causes
An expired credential. Tokens expire, an account is disabled, a password rotates. The connection is now dead and the platform quietly stops.
A trigger that stopped firing. An upstream change, a renamed field, a permission adjustment. The flow is fine and nothing reaches it.
A location never added. The most common one at scale. A branch was onboarded and nobody added it to the configuration, so it never had an automation to break.
A filter that excludes more than intended. A condition written for one case that now silently drops records nobody checks.
The three controls that catch it
1. Alert on absence. If this ran two hundred times last week and zero today, that is the alert. It does not depend on anything failing, which is exactly the point.
2. A heartbeat per location, not per workflow. Thirty locations means thirty expected signals. A single workflow level check reports healthy while three branches produce nothing.
3. A weekly look at the business result. Open ten records the system should have touched and check the outcome is present. Fifteen minutes, and it catches what no dashboard shows.
Make it part of the configuration
If a location is a configuration, add two fields to it: expected volume, and last successful run. Then the absence check is a query rather than a project.
This also solves the onboarding case, because a new location with no expected volume shows up immediately as unconfigured rather than as silent.
The definition to keep
A green execution log is not proof that anything worked. And the absence of a log is not proof of health.
The proof is the record: open it, and the business result is either there or it is not.