Your Business Automation Plan

Two, not seven

You have now seen seven recipes: sorting mail, routing enquiries, extracting from documents, chasing invoices, filling a content queue, handling video, and catching the failures that hide. The temptation at this point is to want all of them.

Resist it. Pick two.

The strongest predictor of whether a small business is still running its automations six months later is how many it started with. Two, built carefully and watched closely, teach you what your data actually looks like, which is never quite what you assumed. Seven, built in a weekend, become seven things nobody quite owns and nobody quite trusts, and by March they are all switched off.

Two also fits the honest amount of attention you have. Each one needs a digest read, a weekly sample and an owner. That is real time, every week, forever. Multiply it by seven and you can see why the seven approach ends the way it does.

Four decisions, all made before you build

Pick the two jobs. Frequency and annoyance decide which are worth it. A veto on anything expensive to get wrong decides which go first. Your best two candidates are almost certainly among the boring, daily, instantly checkable ones.

Name an owner for each. One person, by name. The owner reads the digest, runs the weekly sample and has the authority to switch it off without asking anyone. If you cannot name an owner, you cannot have the automation. That is not a rule about process, it is a description of what happens: unowned automations are the ones still running wrong nine weeks later.

Define success measurably. Not it saves time. How many items a week, how many needed correcting, how many minutes it actually gave back. Write the target down before you build, because afterwards you will unconsciously adjust it to whatever you got.

Agree a stop condition while you are calm. This is the one people find strange and it is the most useful thing in the lesson. Decide now, in a good mood, with nothing going wrong, what would make you turn this off. Because when something does go wrong, you will be embarrassed, invested, and inclined to explain it away. A stop condition written in advance makes that decision for you.

Fill one of these in for each of your two. Handwritten is fine.
THE JOB
It runs when: ...................................
It happens about ....... times a week
Right now it is done by: ........................
It takes about ....... minutes each time
THE OWNER
Name: ...........................................
Reads the digest: [daily / weekly]
Runs the sample: every ......... , five random items
Can switch it off without asking anyone: YES
SUCCESS, MEASURED
In a good week it handles ....... items
No more than ....... need correcting
It gives back about ....... minutes a week
I will know it is wrong because: ................
THE STOP CONDITION, AGREED TODAY
I switch this off if:

more than ....... in ten results need correcting, or
it produces something a customer sees that I would
not have sent, or
nobody has read the digest for two weeks, or
................................................
REVIEW DATE: ........... (put it in the diary now)

Build this outside LearnAI

One practical note. LearnAI has no accounts and stores nothing about you. There is nowhere here to save this plan, and that is deliberate.

So make the artifact somewhere you actually work: a document, a shared note, a page in whatever your team already opens, or a sheet of paper on the wall. Somewhere the owner will encounter it without going looking. A plan nobody sees again is the same as no plan, and this one needs to be readable by whoever inherits the job when the owner is on holiday.

Your first fortnight, then your first quarter

DAY 1        Switch on the first one only. Read every result
against its original. All of them.
DAYS 2 TO 3  Read half of them against their originals.
DAY 4        Check the unclear pile. Empty is suspicious.
Growing is a definition problem.
DAY 7        Count the low confidence results. More than one
in five means your categories overlap.
DAY 14       Ask the owner one question: are you reading the
digest, or skipping it? An ignored digest is a
failed automation, however accurate it is.
WEEK 3       Switch on the second one. Same process. Never
both in the same fortnight.
WEEKLY       Five random results, ten minutes, written down.
Keep the sheet.
REVIEW DAY   Three questions and nothing else:

Did it hit the number we agreed?
Did the owner actually run the samples?
Would we build it again knowing what we know?
Two noes means retire it. Not rebuild. Retire.

That last instruction is the hardest one to follow and it is worth more than any recipe in this course. Retiring an automation that is not earning its keep is not a failure. It is the thing that keeps the two you do trust worth trusting.

The handover note

One more page, and it takes five minutes. Every automation needs to be understandable by somebody who did not build it, because eventually somebody who did not build it will have to deal with it at the worst possible moment.

Keep this beside the plan. Update it whenever you change
anything. Plain words only, no tool jargon.
WHAT IT DOES, IN ONE SENTENCE
...................................................
WHAT STARTS IT
...................................................
WHERE THE RESULTS GO
...................................................
WHERE THE THINGS IT COULD NOT HANDLE GO
...................................................
HOW TO SWITCH IT OFF
Exact steps, so a stand in can do it without asking:
...................................................
WHAT HAS GONE WRONG BEFORE, AND WHAT WE ADDED
Date .......  What happened .......................
What we changed ....................
WHO TO ASK IF IT LOOKS WRONG
Name ...............  Second name .................

The last two boxes are the ones that pay you back. A record of past failures explains why the odd looking checks exist, and a second name means the automation does not become unsupervised the week its owner is away.