DEV Community

Cover image for A stopped automation looks exactly like a quiet week
Brian Pawl
Brian Pawl

Posted on • Originally published at nuwaybizsolutions.com

A stopped automation looks exactly like a quiet week

Originally published on the NuWay Biz Solutions blog.

✦ Cover image: Made with ChatGPT — we're transparent about AI. See the exact prompt on the original post.

Zapier will switch one of your automations off once it errors 95% of the time it runs, and only after it has run more than twenty times in the past week. That is a reasonable threshold for a company running a few hundred million of these things. Below it, the automation stays on.

Zapier turns an automation off after it errors 95% of the time it runs

Sit with what that means for one business. Something that fails half the time is, by that measure, working fine.

And that is the generous case, because at least a failure is an event. It gets counted. Zapier will email you about it immediately, every time, without you configuring anything, because that is the default setting.

The ones that hurt you never produce an event at all.

The automation stops running, and nothing errors

An error notification needs an error. An error needs a run. Follow that backwards and you get the whole problem: an automation that has stopped triggering generates no runs, so no errors, so no emails, so no alarm anywhere in the system.

It is not hiding the failure from you. There is no failure to report. As far as the tool is concerned, nothing happened, which is exactly what it would report on a genuinely slow week.

An automation that has stopped running looks exactly like an automation with nothing to do.

This happens more mundanely than people expect. Somebody at a software company changes something on a Tuesday and the trigger stops matching. A password change or an expired login quietly drops the connection. Somebody turns an automation off to test something and never turns it back on. Or you hit the ceiling on your plan, at which point Zapier holds actions across every automation in the account until the next billing cycle.

Notice that only some of those are anybody's fault, and none of them make a sound.

Sets off the alarm

  • A step runs and returns an error
  • The same step errors on nearly every run for a week
  • A connected account needs reconnecting

Sets off nothing

  • The trigger stops firing, so nothing runs at all
  • It runs and writes an empty field
  • It runs twice and creates the record twice
  • It fails a third of the time, forever

It runs, and writes the wrong thing

The second failure mode is worse, because it produces a success. Somebody renames a field on one side, and the automation keeps firing on schedule and writes a blank into the other side. Every run goes green. Your customer list slowly fills with records missing the phone number.

Nothing in that chain is broken in a way software can detect. The trigger triggered, the step ran, the record was created. A tool that moves data between two apps has no opinion about whether the data was any good, and asking it to have one is asking it to understand your business.

The third mode is the same idea pointing the other way: it runs twice and you get two of everything. We spent a whole article on what that costs you when it lands in your books, so it is enough to say here that a double is also a success, twice.

The pattern behind all three: your tools can tell you a step failed. They cannot tell you the result was wrong, and they cannot tell you about work that never happened.

Every one of these tools has the same gap

Zapier is named throughout because it publishes its thresholds and its defaults, which is more transparency than most tools offer. Every other tool that moves data when something happens has the same blind spot, whatever the logo.

The check is the thing you are missing

The fix is not a better platform. If your automations are mostly working, keep them. Moving everything to a new tool buys you a fresh set of quirks and the identical blind spot.

What you are missing is a check: something that confirms the work actually happened, and happened correctly, without relying on the tool that did the work to grade itself.

At the smallest scale that check is a person and a calendar reminder.

Count both sides over the same window. Forty-one people filled in the form last week. Are there forty-one new rows in the customer list? That single comparison catches all three failure modes at once, which no amount of staring at the automation history will do, because the history can only show you the runs that happened.

Painterly editorial illustration: a brass balance scale, its beam tilted clearly out of level. On the low pan a cobalt glass bowl heaped with pale stones; on the high pan a matching bowl holding only a few stones, one of them red. Slate-navy shadow on one side, warm cream light on the other. No people, no legible text. Two counts that don't agree.

✦ Made with ChatGPT — we're transparent about AI. See the exact prompt on the original post.

The ten-minute version

Pick the automation that touches your money or your customers. Count what went in over the last full week at the source, then count what came out at the destination. Write both numbers down. If they match, you have learned your automation is alive, which you did not actually know this morning. If they do not, you have found the thing before your customer did.

Above a certain size, that check has to run on its own. Something that expects work to happen and complains when it sees none. The rule to hold onto is that no errors does not mean nothing is wrong: you want proof of work completed, not the absence of complaints.

Not everything needs this. If an automation moves something you would notice missing by lunchtime, your business is already the check, and adding a second one is theatre. The ones that need watching are the quiet ones: the automations feeding a number somebody trusts, where nobody is waiting on the result and nothing looks wrong until a decision has been made on it.

It's the same story as the spreadsheet that had been quietly wrong for a year. A number gets believed because it appeared where numbers appear. Automating the work does not remove that risk. It removes the person who used to trip over it.

  • [ ] I know which of my automations touch money or customer records
  • [ ] I would find out within a day if one stopped running
  • [ ] Somebody checks that both sides still agree, on a schedule
  • [ ] The error emails go to a person who reads them, not a shared inbox
  • [ ] I could rebuild the important ones if the person who set them up left
  • [ ] I know what each one is supposed to produce in a normal week

Four or more means your automations are supervised, which puts you ahead of most. Look at the ones you couldn't tick and start with whichever touches money.

Start with one automation. The one that would embarrass you in front of a customer if it had been dead since March.

Count both sides. Then decide whether you knew the answer before you counted.

If you've got a handful of automations running and nobody's sure which ones are still doing their job, tell us what they're meant to be doing and we'll tell you which ones are worth putting a check on. Some of them won't be.


Practical AI. Clear process. Real business value.

— Brian, NuWay Biz Solutions

P.S. This question gets much sharper the moment somebody tries to sell you an AI agent that runs on top of all this. That one's coming.

Top comments (0)