The most expensive automation is not the one that fails. It is the one that works, on a job that should not have been automated, and quietly makes the business worse for two years.
The frequency test
Count how many times a year the job actually happens. Under about twenty and you are almost certainly better off doing it by hand, because every automation carries a maintenance cost that arrives whether you used it or not. Systems rot, tools change their interfaces, and something built for a job you do quarterly will be broken the next time you need it.
The consequence test
Ask what the worst realistic outcome is if it gets it wrong and nobody notices for a week. If the answer is embarrassment, proceed. If the answer is a legal problem, a lost client or an unrecoverable amount of money, the job may still be worth automating but the approval line has to sit in front of it, which changes the economics.
Automating a decision is very different from automating the work that leads up to it. The first is usually a mistake. The second is almost always the win.
The relationship test
Would the person on the other end mind if they knew? Not whether they would find out, whether they would mind. A rent reminder, nobody minds. A condolence, an apology, a difficult conversation about money owed by somebody genuinely struggling: those are the moments the relationship is actually made of, and handing them to a machine saves twenty minutes and costs the client.
The stability test
Is this how you will still be working in eighteen months? Businesses about to change systems, restructure a team or move into a new service line should not be entrenching the current shape in software. Wait, or build only the parts that survive the change.
What to do with the ones that fail
Usually one of three things. Delete the job entirely, which is more often available than people expect, particularly with reports nobody reads. Simplify it so a person can do it in a fraction of the time. Or automate the gathering around it and leave the judgement where it is, which is the answer surprisingly often.