An automation breaks in two ways. A hard failure stops the run, and the platform can alert you. A silent failure finishes the run and produces nothing, so no alert fires. Build an error workflow for the first and a check on the output for the second. The difference decides what you build.
This becomes real the moment a workflow runs in production and a colleague depends on it. Outages are normal. The useful question is not whether something will fall over, but how many days pass before anyone knows.
How do you notice an automation stopped running?
Detection is a thing you build, not a thing that happens to you. Three pieces do the work: an error workflow that fires on a failed execution[1], a counter that asks whether anything came out today, and a failure rate you glance at every few weeks[2]. Without them, the outage reaches your customer before it reaches you.
What an error workflow does, and what it does not
Some vocabulary first. In n8n, a workflow is a set of connected nodes that together automate a process[3]. Nodes decide when the workflow runs, fetch data, send it and transform it[4], and every production workflow needs at least one trigger node that determines when it starts[4]. One pass through that workflow is an execution, or a run.
When a run fails, a second workflow can take over. You point to that error workflow per workflow in the workflow settings, it runs when an execution fails, and it can send an email or a Slack message[5][6]. The error workflow has to start with the Error Trigger node, which collects the details of the failed workflow and the errors[1][6]. One error workflow can serve several workflows[6], and a workflow that contains an Error Trigger itself does not need to be published: it uses itself as its error workflow by default[7].
So far, so expected. The part that trips people up is testing. You cannot try an error workflow out by running the main workflow by hand: the Error Trigger only fires when an automated run throws an error[7]. A failure on your own terms also has to be declared. The Stop And Error node does that, showing your own error message, failing an execution on conditions you choose, and sending your own error information to the error workflow[8].
That is the line between technical monitoring and monitoring the content of the work. Executions usually fail for technical reasons, such as a mistake in the node settings or a workflow that runs out of memory[6]. An order arriving without a customer number is not an error to the platform. If you want an alert on that, you have to raise it yourself. WeAdapt builds workflow automation with both kinds of monitoring in place, because the second kind is what prevents the expensive mistakes.
While you are in the workflow settings, turn on saving failed production executions and set the timeout for a running execution[5]. Without saved executions there is nothing to read back afterwards.
Which signal means what
| signal | what it means | first move | who picks it up |
|---|---|---|---|
| Alert from the error workflow | The run stopped, usually on node settings or memory[6] | Open the failed execution in the executions list[6] | the named owner |
| Failure rate climbing | Failures are structural, not a one off[2] | Check whether the failed runs hit the same node | the named owner |
| Counter at zero | The workflow ran and delivered nothing: a silent failure | Check the input side, not the workflow | the named owner |
| No execution at all today | The trigger never fired, so there is no failure to report[6] | Check whether the workflow is still active | the named owner |
| A colleague says nothing arrived | Someone spotted it before your monitoring did | Process it by hand, then widen the monitoring | the backup |
| Work piling up in the source system | The automation processes less than comes in | Count the open cases and put them on a list | the named owner |
What is a silent failure in a workflow?
A silent failure is a run that succeeds while producing no result. No error workflow fires, because an error workflow only runs when an execution fails[1][6]. To the platform, a green run with an empty outcome is simply a green run.
This is not a platform defect. It is a gap in the monitoring, and it takes three shapes in practice:
- Nothing to process. The source returns an empty list, because a filter is too narrow or a connection hands back an empty set. The workflow walks through every node and ends cleanly.
- Processed into the wrong place. The run succeeds and the data lands in a field, a folder or a list that nobody opens.
- The trigger never fired. There is no execution at all, so there is no failed execution either, and nothing to send an alert about.
In all three, the only reliable measurement is the outcome itself. Put a second, small workflow next to the automation and let it run on a fixed schedule. The Schedule Trigger does exactly that: it starts a workflow at fixed intervals and times[9]. That check answers one question. How many items were processed today? If the answer is zero on a day that should not be zero, it sends the alert.
No error workflow can produce that alarm, and neither can the usage statistics of your platform. Insights in n8n reports total production executions, total failed production executions, the failure rate, the time saved as configured per workflow, and the average run time[2]. Those numbers answer whether runs fail, never whether runs deliver. Read both and the monitoring is complete: did it run, and did something come out?
In practice that turns into a handover rule. An automation is finished when it has an error workflow and a check on its output, with a threshold you picked yourself. Not because it looks thorough, but because the outage nobody sees is the one that costs the most. How that move from trial to daily use works is covered in AI pilot to production.
Who gets the alert when something goes wrong?
One person, by name, plus a second who takes over when the first is away. An error workflow can send an email or a Slack message[6], and where that message lands decides whether anything happens next. A shared channel without an owner is not a destination. Everybody reads it and nobody acts.
Three things belong in the message: which workflow, when it went wrong, and what the error was. The Error Trigger collects those details from the failed workflow[1], so it is a matter of passing them on rather than writing them yourself. Add one line about what the recipient should do now: handle it manually, or wait for the builder to look.
The same applies to silent failures, with one difference. There the check workflow reports what did not happen. “Nothing processed today, expected more than nothing” is a message somebody can act on. Sending an all clear when there is nothing wrong works against you, because after two weeks nobody opens it.
Who owns which automation, and what you agree about that at handover, is a separate conversation. It is covered in who owns an automation after launch. This article stays with detection, alerting and recovery. If you are not sure which of your processes actually need this level of monitoring, consultancy maps that out before anything gets built.
What do you do by hand while the automation is down?
You fall back on the process that existed before the automation, and you keep a record of what piles up. That only works if somebody still knows that process and if the list of open cases can be pulled from the source system. Both are arranged at handover, not during the outage.
The fallback path, in seven steps:
- Pause or deactivate the workflow, so it cannot run half of the work while people handle the other half by hand.
- Pull the list of open cases from the source system rather than from the automation. The automation is the part you cannot trust right now.
- Name one person who processes by hand and writes down what has been done.
- Tell the people waiting on the output that today will be slower. One short message saves an afternoon of answering questions.
- Look for the cause in the executions of that one workflow, or in the executions of every workflow you have access to; streaming logs to an external system is an option here[6].
- Switch the workflow back on only once you know what went wrong, and replay what was left behind using the list from step two.
- Check that the replay did not double anything: the same email twice, the same record twice.
One rule governs all seven steps, and most handover documents leave it out. Retire the manual process after the automation has survived its first outage, not after its first good week. A workflow that has never failed has monitoring nobody has seen work, and error workflows cannot be rehearsed by hand[7]. The price of keeping the fallback path until then is a one page instruction with a name on it.
If you cannot name the person who would spot your automation going quiet, the monitoring is the gap, not the workflow. WeAdapt goes through the detection, the alert and the fallback path, and builds in what is missing. Book a call, or start with the frequently asked questions about how a project runs.