An automation that breaks loudly is a nuisance. An automation that breaks quietly is the expensive one, because by the time somebody notices, it's been wrong for a month and you have to work out which records to trust.
These are the five ways it happens, in roughly the order they happen.
1. A connection dropped and nothing said so
Somebody changed a password, or an app asked to be reconnected and the email went to whoever set it up two years ago. The automation stops. Nothing else happens.
This is the commonest one by a distance, and it's invisible, because a job that isn't running produces no evidence that it isn't running.
What stops it: an alert on the second consecutive failure, sent to somebody who still works there. And a heartbeat, so that something running zero times today is itself the alarm. Most connector tools can do the first and not the second.
2. It ran out of tasks
Connector tools bill by volume. Zapier's entry paid plan is $19.99 US a month billed annually and includes 750 tasks, where every step of every run counts as one. Make's Core plan is $9 a month for 10,000 credits, counted differently again.
A five-step automation running thirty times a week is 150 tasks a week, or about 650 a month. That's inside 750 until a busy couple of weeks, and then it stops partway through the month with no warning.
What stops it: watch the usage, not just the runs. And do the arithmetic at three times today's volume before you rely on it, because the plan you're on is sized for the business you had when you set it up.
3. It did the same thing twice
A step timed out, the tool retried it, and the first attempt had actually worked. Now there are two invoices, or two orders, or the customer got the same email twice.
This is the one that damages trust fastest, because a duplicate is visible to the customer in a way that a missing record isn't.
What stops it: each step has to be safe to repeat. That means checking whether the record already exists before creating it, and it's ordinary practice in software. In a connector tool it's usually possible and rarely set up.
4. The other app changed
It ran perfectly for a year. Then the app on the other end changed a field name, or started returning dates in a different format, or added a step to its login.
The automation carries on and writes the wrong thing into the right place, which is worse than stopping. Blank names, dates a year out, amounts in the wrong currency.
What stops it: checking that what came back is the shape you expected, and failing loudly when it isn't. A good automation refuses to guess. Most simple ones happily carry on.
5. A case nobody thought of at setup
You built it around how the job normally goes. Then a customer pays in two instalments, or a job gets cancelled and rebooked, or somebody's name has an apostrophe in it.
The automation was never told what to do, so it does the closest thing, and now one person's records are wrong in a way nobody looks for.
What stops it: send the unknowns to a person instead of guessing. Anything that doesn't match a case you planned for goes in a pile somebody looks at on a Friday. That pile is also the best list you'll ever get of what to build next.
What makes one hold up
The pattern across all five is the same. They aren't logic failures. The logic is usually right.
They're failures of everything around the logic: knowing when it stopped, knowing when it's wrong, coping with a step that half-worked, and having somewhere to put a case nobody predicted.
That's the honest difference between a connector automation and something built. Not that one is cleverer. It's that the built one is allowed to have error handling, retries that are safe, a log you can read, and an alarm that goes off.
Connector tools are genuinely good and worth starting with. The moment to move is when it's carrying something you'd be embarrassed to get wrong, or when the list of exceptions is longer than the list of rules.