How much should a small business automate with AI?

The question is usually asked as though the limit were budget or technical ability. In practice the limit is supervisory capacity, and it binds much earlier than either. Every automated process is a thing that can be quietly wrong, and quietly wrong processes accumulate in exactly the businesses that adopted enthusiastically. A firm running four automations it checks is in better shape than one running twelve it does not.

The clearest way to see this is to price an automation properly. The build cost is visible and one-off. The running cost is a subscription and some usage. The cost nobody enters is the recurring obligation: someone must notice when it stops, notice when its output degrades, and update it when something upstream changes. That is perhaps an hour a month per automation in a settled state, and considerably more in the first quarter. Twelve automations is therefore most of a day a month before anything goes wrong.

This suggests a sequencing rule that is more useful than a target. Automate the highest-frequency job first, then stop and run it for a month before building anything else. Frequency matters because the return is per occurrence, so the same build effort pays back several times faster on a daily task than a monthly one. Stopping matters because the first automation teaches you what the ongoing obligation actually costs in your business, and that number is what should govern the next decision.

There is a category worth automating early that is often skipped because it feels unambitious: the observation layer. A daily summary of what arrived, what was answered, what is outstanding. It has no permission surface, its failure is visible because you expect it each morning, and it makes the rest of the estate legible. Businesses that build this first tend to make better decisions about what to automate next, because they can see where the work actually is.

There is also a category to leave alone regardless of capacity. Anything where the exception is the point — the complaint, the unusual request, the customer who is upset — should stay manual, not because a system could not handle the common case, but because the value of those interactions is concentrated in the cases the automation would misclassify. Automating the easy ninety per cent of a job whose whole value sits in the difficult tenth is a net loss that shows up months later as churn.

For a business of a handful of people, a realistic settled state is a small number of automations, each with an owner, each with a check, and each with a written note of what it does and what to do when it stops. That is a less impressive answer than a transformation programme and it is the one that is still working in a year, which is the only comparison that matters.