Monitoring automations

An automation that crashes you notice. One that quietly stops you notice after weeks – when someone asks about the report that stopped coming long ago. And one that keeps running but has delivered nonsense since a change to the source you may not notice at all.

Three ways to fail

• It no longer runs. Access expired, quota empty, paused by the provider.
• It runs and delivers something wrong. A column has a new name, a folder was moved, the model was updated.
• It runs and nobody reads the result. It costs money and creates the illusion that something is under control.

For each there is a simple countermeasure.

1. The heartbeat

Every run ends with a message – including the one where nothing came up. “Run of 10 Nov, 7:00: 0 new entries, source reachable” is not noise but the proof that a check took place. If the message is missing, you know the same day that something is wrong.

Keep the message short and always structured the same way:

Run: weekly report · 10 Nov 2026, 07:00
Source reachable: yes
Processed: 214 orders
Anomalies: none
Result: folder “Weekly reports”, “Weekly report CW 45”

Three numbers you can compare week by week show deviations faster than any text. If “Processed” jumps from 214 to 0, you do not need to read the report to know something is broken.

2. The canary

Against wrong results a known case helps: put an entry into the source whose correct evaluation you know – a test order, an email to yourself. If it does not show up in the result, or shows up wrong, the processing is no longer right.

Also check the canary after every change you did not make yourself: new model, new version of the tool, restructured source.

3. The budget

Every automation gets a cap: runs per day, cost per month or both. The platforms set limits of their own, but those protect the provider, not your wallet.

Two numbers are enough: what does one run cost, and how many may there be at most? An event trigger without a cap is an open invoice. How the cost of a run arises is explained in Understanding AI costs.

4. The expiry date

Automations get set up and never switched off again. Give each one a date four to twelve weeks out when you create it. On that day you answer one question: has anyone used the result in the meantime? If not, it is deleted – not paused.

The register

All of this presupposes that you know what is running. A single list, one row per automation:

“Owner” is a name, not a team. And the “May” column forces you to write the permissions out once – often that is the moment it becomes clear that an automation may do more than it needs.

The switch

For every automation someone must be able to say in under a minute where it is paused. Try it once while nothing is on fire. In an emergency you will otherwise be searching three interfaces for a button while the agent keeps working.

The last station of this path shows five workflows with clear access limits that you can take as templates. If you want to build yourself after that, the expert level picks up from here – with coding agents.