Which parts of a small business can actually be automated with AI?
A realistic assessment of what AI automation does and does not suit in a small business, how to choose the first thing to automate, and what it costs to run.
The parts of a small business that suit AI automation are the repetitive, rule-shaped tasks with a clear input, a clear output, and a tolerable failure mode — routing incoming enquiries, drafting quotes from a fixed price list, extracting details from invoices, reformatting data between two systems, and producing first drafts of routine content. The parts that do not suit it involve judgement, relationships, or accountability for being wrong. The distinction matters more than the technology: most disappointing automation projects fail not because the model was inadequate but because the task was never a good candidate. A useful filter is whether you could write down the rule clearly enough for a new starter to follow it.
How to choose the first thing
Count three numbers for each candidate task: how often it happens, how long it takes each time, and how bad it would be to get it wrong.
High frequency, meaningful duration, low risk is the sweet spot. A quarter-hour task done daily is worth automating. A two-hour task done twice a year is not, however much you dislike it — you will spend longer specifying it than doing it.
Then check that the inputs are actually machine-readable. A workflow that begins with a handwritten note or a phone call needs a different first step.
What an automation actually looks like
Usually a small and rather boring piece of software. Something triggers it — an email arriving, a form submission, a schedule. It gathers the real source data, calls a language model for the part that genuinely needs language understanding, and then does something deterministic with the result: files it, sends it for approval, updates a record.
The important design choice is where the human sits. Anything customer-facing should be drafted by the automation and approved by a person, at least until it has earned trust. Anything internal and reversible can run unattended.
Keeping it from inventing things
Constrain it. Give the model the actual source data rather than asking it to recall facts. Scope each automation narrowly. Log inputs and outputs so you can audit what happened. Build a visible failure mode, so that when something upstream changes the process stops and tells you rather than quietly producing wrong output.
If a task cannot be checked — because nobody would notice an error — that is a strong argument against automating it at all.
What it costs
Two numbers: a build cost, and a running cost. For a small business workflow the running cost is often a few pounds a month in API usage, sometimes less. The build cost depends entirely on how many systems have to be persuaded to talk to each other.
Both should be quoted before work begins, alongside an estimate of hours returned per month. If the arithmetic does not clearly favour building it, the honest recommendation is not to.
What to be sceptical of
Anything sold as a general-purpose assistant that will “transform your operations”. In a small business there is rarely a whole role’s worth of automatable work, and promises pitched at that scale usually end in an expensive tool nobody opens. The realistic and genuinely valuable outcome is narrower: a few hours a week returned to people who have better things to do with them.