What You Should Automate First

The job you thought of first is the wrong one

Say the word automation in an office and watch what happens. Somebody immediately names the monster: the quarterly report that pulls numbers out of four systems, needs three sign offs and eats a full day. Everybody hates it, so it feels like the obvious candidate.

It is almost always the wrong place to start. It happens four times a year, which gives it four chances a year to repay the afternoon you spent building it, and by the second one you will have forgotten how the thing works. Meanwhile the small job you do eleven times a day carries on being done by hand, because it never felt important enough to mention.

Choosing well is most of the skill in this course. The building part is easier than people expect. The choosing part is where automations quietly die. So before you connect a single app, put your candidates through four questions.

Four questions, in this order

How often does it happen? Count real occurrences across one normal month. Not how often it feels like, which is always higher. Under about four times a month, leave it alone for now. Setup is a fixed cost and frequency is the only thing that repays it.

How much does it annoy you? This is not a soft factor and you should not be embarrassed to score it. A task you resent is a task you delay, and a delayed task costs more than the minutes inside it. The enquiry that sits unread until Thursday is no longer a two minute job, it is a two minute job plus an apology.

How bad is a wrong result? The first two questions ask whether the job is worth building. This one asks whether you should build it at all. If a mistake means something lands in the wrong folder, shrug and fix it. If a mistake means a customer receives a message meant for a different customer, that job needs a person standing inside it, which is lesson five.

How long does checking take? The quiet killer. If confirming the output takes nearly as long as producing it, you have not saved anything. You have swapped doing for proofreading, and most people find proofreading the worse job. The best first candidates are the ones where a three second glance tells you whether the result is sound.

Write one line per job you are considering. Then score each out of 5.
FREQUENCY     how many times a month it really happens
ANNOYANCE     how much you dislike doing it
DAMAGE        how bad a wrong result would be
CHECKING      how long confirming the result takes
Score = (frequency x annoyance) - (checking x 2)
Then apply the veto. Anything scoring 4 or 5 on DAMAGE does not go
first, however high it scores overall. Park it for later, with a
human approval step built into it.
Build the highest scorer that survives the veto. Only that one.

The two traps

Trap one is the rare job. It looks worth automating because it is horrible, and horrible is memorable. But an automation that runs twice a year never earns back its setup, and worse, it rots in between. Systems change, somebody renames a folder, a form gains a field. The next time it fires you are untangling something you last thought about six months ago. Rare jobs deserve a checklist, not an automation.

Trap two is the expensive job. Refunds, invoices, contract terms, anything that goes out under your name. These are tempting precisely because they are high value, which is the same reason they are dangerous. The right move is not to avoid them forever. It is to save them until you have run something harmless for a month and learned how your setup actually behaves when real data hits it.

Five shapes that make good first candidates

ROUTING       something arrives, decide who or where it goes
SORTING       take a pile of mixed things, put each one in a category
SUMMARISING   turn something long into something short a human will read
NOTIFYING     tell the right person that a thing happened
COPYING       take information from one place and put it in another
What these five have in common:

they happen often
a person can tell in seconds whether the result is right
getting one wrong costs a correction, not an apology
What is missing from that list, deliberately:
sending, paying, deleting, publishing, promising.

Pick one, and only one

The strongest predictor of whether somebody still uses their automations three months later is how many they built in the first week. One, watched closely for a fortnight, teaches you more than five built in an afternoon. You will find out what your data actually looks like, which is never quite what you assumed, and you will build the second one twice as well.

Fill this in before you build anything. Handwritten is fine.
THE JOB
It happens when: ...........................
It happens about ...... times a month
Right now the person doing it is: ...........................
It takes them about ...... minutes each time
WHAT GOOD LOOKS LIKE
When it works, this is true: ...........................
I will know it is wrong because: ...........................
If it gets one wrong, the cost is: ...........................
THE HONEST BIT
Checking each result will take me ...... seconds.
If that is close to the minutes above, choose a different job.