A quiet reading corner in a modern office at blue hour, an armchair and floor lamp by tall glazing

The other half of the test

When not to automate something

Do not automate work that is rare, work that is about to change, work nobody owns, or work where the difficulty is a decision rather than a sequence. Each of those looks like a candidate from the outside, and each of them costs more to automate than to leave alone.

Last checked 19 September 2026

This is half the answer, and it is the half nobody publishes #

Almost every page on this subject is written by somebody who is paid when you automate something, which is why almost every page concludes that you should. We are in exactly that position, so it is worth being direct: a good proportion of the jobs businesses bring to us should be left alone, and saying so early is cheaper for everyone than finding out later.

The positive version of this test, the two axes of repeatability and judgement, is set out in how to tell which jobs are worth automating. This page is the list of things that fail it, in the order we see them.

1. Work that happens rarely #

The economics are unforgiving here and they are not really about software. Any handover costs something to specify, build and check, and that cost is paid once. What repays it is frequency. A job that runs four hundred times a year is a different proposition from the same job running four times, even if each occurrence is identical.

The trap is that annual jobs feel enormous because they arrive all at once and disrupt everything. Year end, an audit, an insurance renewal. The disruption is real and the total hours are usually modest.

2. Work that is about to change #

A process being reshaped by new software, a new regulation, a new premises or a new member of staff should be left until it settles. Building against a moving process means building twice, and the second build is rarely cheaper than the first because the first one has to be unpicked.

The same applies to a business that is genuinely growing. A process that works at the current volume and will not work at three times the volume is not a process worth casting in software; it is a process worth redesigning. There is a separate hazard in things changing underneath a finished build, which we cover in what happens when the tools change underneath you.

3. Work nobody owns #

If no single person can say how a job is supposed to run, there is nothing to hand over. What gets built in that situation is somebody's best guess at the process, and when it produces an outcome two people disagree about, there is no one to adjudicate.

The symptom to watch for is a job that three people describe three different ways, each convinced theirs is the standard version. That is not a reason to give up on the job. It is a sign that the useful work is agreeing the process, and that the agreement is worth having whether or not anything is ever automated.

4. Judgement wearing routine's clothes #

Some work looks mechanical from a distance and is not. Deciding whether to chase a particular client again this week is not a rule, it is a read on a relationship. Deciding which of two urgent jobs the van goes to is not a rule, it is a read on which customer will forgive the delay.

These are the decisions that keep a business alive and they are the last thing anyone should try to hand over.

The useful move is to split rather than abandon. Gathering the information the decision needs is usually rule shaped and usually the bulk of the elapsed time. The decision itself stays where it is.

5. Work where being wrong is expensive and unrecoverable #

Anything that sends, spends or deletes deserves particular caution, and some of it should simply stay manual.

A wrongly filed document is an annoyance. A wrongly sent message to a client is a relationship. The reasoning behind where that line sits is in what should always wait for a person.

6. A configuration problem in disguise #

A surprising share of repeated admin exists because two systems that could talk to each other do not, or because a setting was never switched on, or because somebody built a workaround years ago for a limitation that no longer exists. None of that needs a model and none of it needs a consultancy. It needs somebody to spend an afternoon in the settings.

It is worth checking this before anything else, because it is the cheapest possible fix and because discovering it after a build is an expensive way to learn.

What to do with the jobs that fail the test #

Failing the test does not mean living with the job forever. It usually means one of three things.

  • Simplify it

    Plenty of work that cannot be automated can be reduced by changing what the business asks for in the first place.

  • Stop doing it

  • Hire for it

    Plenty more is genuinely a job for a person.

If the conclusion is a person, the relevant question becomes what that person's week should look like, which is a different exercise from this one and a more useful one than buying software you do not need.

Questions people ask

What is the most common job businesses try to automate and should not?
The one that annoys them most. Irritation is a reliable signal that something is wrong and an unreliable signal about the cause. A lot of the most hated work is hated because it is unpredictable and time pressured, and unpredictable time pressured work is exactly what automation handles worst.
Is a job too small to automate?
Frequently, yes. Volume is what pays for the setup. A task done twice a year can be genuinely awful and still never repay the work of handing it over, and no amount of good software changes that arithmetic.
What if the process is a mess? Will automating it clean it up?
No. Automating a broken process produces the same broken outcomes faster and with less visibility, because the person who used to notice the problems is no longer in the loop. Fix the process first, then decide whether it still needs automating.
Should I automate something a member of staff is about to leave behind?
Not immediately. When somebody leaves, a great deal of undocumented judgement leaves with them, and the version of the job you can describe is usually a simplified one. Run it manually for a while first, so you find out what the job really involved before you cast it in software.

Free audit

Find out what the repeated work costs you

The audit counts every job the business repeats, ranks them by the hours they eat, and maps the five worth automating first. It is free, you keep everything, and there is no obligation at the end of it.

AI automation, explained

Plain answers to what businesses ask before they automate anything.

All of ai automation, explained

Read next