AI automation for small businesses. What pays off and what does not.

Almost any process in a business can be automated. The question is not whether, but which one first, and which one to leave alone. Here is how to tell the difference.

The question that comes before all the others

Not every process deserves to be automated. The sum is simple and rarely done: how often a month, how long each time, and how bad is a mistake.

Something that takes five minutes twelve times a year is one hour of work. Spend two days building for that and you have kept yourself busy. Something that costs twenty minutes a day is a hundred hours a year, and there almost any effort pays for itself.

The third figure matters most: what happens when it goes wrong. A misfiled advertising email costs one click. A missed deadline costs money, sometimes a whole case. So you do not automate the most laborious thing first, you automate the most expensive one.

What automates well

Anything with a clear rule that happens often. The common factor is not the technology but the question: could a new member of staff take this over after one explanation?

• Collecting receipts: from mailboxes, from messages, from scanned post, and matching each one to the bank entry it belongs to.
• Finding and reporting deadlines: remind somebody on the due date and you have reminded them too late. Useful warnings come 180, 90, 30 and 7 days ahead.
• Reading incoming post: identifying sender, subject, type and deadline, including from a scan.
• Gathering messages from every channel in one place: email, WhatsApp, SMS, fax, letters, phone calls.
• Reconciling payments: which invoice is paid, which is not, which receipt is missing.
• Recurring reports: the same question every morning, the same figure from the same source.

What automates badly

Anything where the rule lives in somebody's head and changes with the situation. A price that depends on the relationship with the customer. The decision whether to take a job on. A conversation that is about trust.

There is one hard boundary, and it does not run along difficulty. It runs along reversibility. What can be undone, a machine may do. What is out, is out: a message sent, a payment released, a file deleted.

That is why the default for anything leaving the building is a draft. The machine writes, the person reads, the person sends.

From one person to thirty

This page is written for businesses between one person and around thirty: planning practices, architecture practices, the trades, small-scale manufacturing, services. Not because larger firms do not have the problem, but because they get a different answer to it.

A corporation has a department that tidies up processes and a budget for six-figure software. In a business of eight, the boss does the quotations, the bookkeeping and the purchasing, and does them in the evening. There every hour won back counts twice, because otherwise it comes out of your own time.

The second difference is the rule itself. With five people it lives in five heads and in no file at all. Write it down so a machine can follow it and you have written it down as a by-product, and that is worth more than the automation the first time somebody is off sick.

How to start without overreaching

Not with a big system, but with one job. The approach that works has four steps.

• Write down for a week what keeps coming back. Not estimate it, write it down.
• Pick out the three most expensive, counted as hours a year multiplied by the cost of getting it wrong.
• Build the first one and let it run until it is boring.
• Only then the next. Build three at once and you have three halves.

Common questions

Do I need my own IT department for this?

No. What is described here runs on an ordinary computer. More important than technology is that somebody knows the rules the business actually works by.

What does an automation cost?

The question is better asked the other way round: what does it cost to leave it. Twenty minutes a day is a hundred hours a year. What the build costs depends on the job and can be worked out in advance.

Do I lose control over my own processes?

Only if the rules are written down nowhere. A good automation records what it does and why, which makes it more transparent than a habit three people handle differently.

Does my data go to a provider?

That is a decision, not a necessity. There are set-ups where the data stays on your own computer and only individual requests go outside.