When not to automate: four tasks to leave alone
By Sam Kim · · 6 min read

How do I know when a task isn't worth automating?
Leave it alone if it happens rarely, if each case needs a judgement call, if the way you do it is still changing, or if the real cause is a person or a role rather than the steps. Automation pays back on work that is frequent, repeatable and stable. A rare task never earns back the setup time, a judgement call needs someone accountable for it, a changing process breaks the automation every month, and a people problem comes back in a new form.
Why would you not automate something you could?
Because every automation has running costs. Someone has to set it up, someone has to notice when it fails, and someone has to fix it when a supplier changes a form or a tool changes its settings. Every automation eventually breaks, so the question is always whether the time it saves is bigger than the time it takes to look after.
There's a second cost that's easier to miss. A task that runs by itself stops being looked at. If it was the wrong task, or done the wrong way, nobody finds out for months.
Is the task too rare to be worth it?
If it happens less than weekly, it usually is. A job you do four times a year doesn't give an automation enough runs to earn back the setup, and by the next run something has often changed.
You can check this on your own numbers. It takes five minutes:
- Time the task by hand, in minutes.
- Estimate how much of that the automation would actually remove. Be honest; there's usually a check left over.
- Multiply the minutes saved by the number of times it happens in a year, then divide by 60. That's hours saved a year.
- Estimate setup hours, plus maintenance at an hour or two a month.
If the setup and a year of maintenance come to more than the hours saved, leave it. End-of-financial-year jobs, annual reviews and the yearly insurance renewal mostly fail this test. A written checklist does more for them than any automation would.
Does the task turn on a judgement call?
Then a person should make the call. AI can collect the facts, draft the options and flag what looks unusual. It shouldn't decide whether to take on a client, waive a fee, write off a debt or tell someone their claim has no merit.
The test is accountability. If the decision goes wrong, someone in the business has to explain it. If that person can't say why the answer was right, the judgement has moved to the tool without anyone deciding it should. What AI won't do covers the wider version of this: a task is worth handing over when a wrong answer is cheap to catch before it leaves the building.
The useful split is to hand over the gathering and keep the deciding. A quote for a standard job can be assembled automatically. Whether to quote a job you don't want, or a client who paid late last time, stays with you.
Is the process still changing?
Then automating it locks in this month's version. If you're still adjusting how you onboard clients, quote jobs or handle bookings, any automation built now gets rebuilt every time you change your mind.
Signs a process isn't settled yet:
- Two people on the team do it differently, and both think theirs is the right way.
- Exceptions come up most weeks and get handled by whoever is around.
- You changed the steps in the last two months, or you're planning to.
- Nobody could write it down on one page without asking someone else.
The fix is cheap. Run it by hand with a written checklist for a few weeks and note every exception. When the checklist stops changing, it's ready. Often the checklist alone removes most of the delay, because the slow part was people working out the steps each time. Handover documentation is a good place to start that writing.
Is it really a people problem?
This is the one that wastes the most money. A process looks slow, so you automate it, and it stays slow. The steps were never the problem.
Some common versions:
- Work waits for one person. Jobs sit until the owner or a senior person approves them. Automating the reminders just sends that person more reminders. The fix is deciding what they no longer need to approve, which is the partner bottleneck problem.
- Nobody owns the task. Follow-ups get missed because three people each assume someone else is doing it. An automation can send the follow-up, but it can't decide who handles the reply.
- Someone doesn't want to do it. A step gets skipped because the person responsible thinks it's pointless. Sometimes they're right and the step should go. Sometimes they need to be told it matters. Either way, that's a conversation.
- The team doesn't trust the system. If people keep their own spreadsheet alongside the official one, automating the official one changes nothing.
One question helps here: if you hired a sharp new person tomorrow and gave them this task with clear instructions, would it get faster? If yes, the steps are the problem and automation may help. If no, because they'd hit the same wall, the wall is a person, a role or a decision.
What does that leave worth automating?
More than you might think. Once you take out the rare tasks, the judgement calls, the unsettled processes and the people problems, what's left is usually the frequent, dull, predictable work: confirming appointments, chasing missing documents, sending review requests, moving details from one system to another, filing.
That's where a small team gets real hours back, and where people stay in the loop by checking a sample of what went out instead of doing all of it by hand. Measuring whether it worked shows how to tell once it's running.
If you'd like a second view on which of your tasks fall on which side of the line, that's what the AI Tools Assessment is for.
What to do next
- List the five tasks you'd most like to automate.
- Put each one through the four questions: is it rare, does it need judgement, is it still changing, is it a people problem.
- For anything that passes, run the payback calculation on your own numbers.
- For anything still changing, write a checklist and run it by hand for a month.
- For anything that's really a people problem, have the conversation before you buy anything.
General information, not legal, tax or financial advice · Drafted with AI tools to my written rules · Terms