Two years ago we were told this would either transform the business or end it. It did neither. What it actually did was more specific and more useful than either version of the prediction, and worth writing down honestly, because most of what gets said about AI in small businesses is written by people selling it.
Here is what it took over, what it made worse, and the test we now run before anything else goes on the card.
The three jobs it genuinely took over
First drafts. Not finished work — drafts. Proposals, standard emails, job adverts, the covering note that goes with a quote. The first version of anything now takes minutes rather than an hour, and the quality of that first version is roughly what a competent person produces when they are tired. That turns out to be a perfectly good starting point, because editing something mediocre is far faster than staring at an empty page.
Notes from calls. Transcription and summarising was the least glamorous change and the one that stuck hardest. Nobody in the business now writes up a client call by hand, and the notes are more accurate than the ones we used to write, because they are not filtered through what the person taking them thought was important.
The second pass on structured admin. Pulling figures out of a pile of documents into a spreadsheet, cross-checking two lists against each other, reformatting a supplier's export so it will import somewhere else. Repetitive, rules-based, verifiable work. That last word is doing a lot of the lifting.
The two it made worse
Client email, when we let it run at scale. Replies got longer, more polite and completely empty of information. A three-line answer to a direct question became four paragraphs of throat-clearing. Two clients said, gently, that our emails had started sounding like everyone else's, which is precisely the opposite of the thing a small firm is selling.
And research, on the one occasion nobody checked it. A figure went into a quote that was confidently stated and a year out of date. It was caught at proof stage, which meant the total cost was an afternoon and a slightly awkward internal conversation rather than a wrong price going to a client. But the lesson landed: the failure mode is not obvious nonsense, it is plausible nonsense delivered in the same tone as the correct answer.
Every hour it saved us in drafting came back as roughly half an hour of checking. That is still a good trade — but it is a trade, not a free hour.
The bit we underestimated: checking is the job now
The mental model most owners start with is that AI does the work and you get the time back. What actually happens is that the shape of the work changes. Producing goes down, verifying goes up. If the output is going outside the business — to a client, a lender, HMRC, a customer — somebody has to read it properly, and the person doing that reading has to know enough to spot what is wrong.
Which means the tools are most valuable in the hands of the person who could have done the job themselves, and least valuable in the hands of the person who could not. That is the reverse of how they are usually sold.
The test we apply before paying for anything else
Three questions, in this order.
Which named task does this do, how often does it happen, and how long does it take now? If nobody can answer that without hand-waving, the answer is no. Who checks the output, and how long does that take? Add that time back before calculating any saving. And what happens when we cancel — is the work still doable, or have we quietly built a process that only exists inside somebody's subscription?
Then the arithmetic, which is genuinely simple. A tool at £20 a month costs £240 a year. If it saves half an hour a week, that is 26 hours; at an internal cost of £45 an hour, about £1,170 of time. Obviously worth it. A tool at £200 a month costs £2,400 a year, and if it saves fifteen minutes a week — 13 hours, about £585 — it is destroying value in a way that no demo will ever show you. Most of the disappointment we see in other businesses comes from never doing that second calculation, which is the same failure that drives the subscription trap generally.
What it has not replaced
Judgement about which client to take on and which to decline. The relationship, which is still the main reason anybody buys from a business our size rather than a bigger one. The decision about what the business should be doing at all, which is the thing owners are supposed to spend their time on and rarely do.
And accountability. When a piece of work is wrong, the answer is never that the tool got it wrong. It is that we sent it.
Where we landed
Two subscriptions, not nine. One shared account rather than everyone expensing their own. A written rule about what never goes into a chat window — client data, anything with someone's personal details in it, payroll, anything covered by a confidentiality clause. A named person who reviews anything client-facing before it leaves. And a standing review every six months of what each tool has actually replaced, which so far has cancelled more subscriptions than it has approved.
The honest summary is that it made a small business meaningfully faster at the parts of the work nobody enjoys, and changed nothing about the parts that decide whether the business survives. Which, on reflection, is what most good tools do. The harder question — what you should stop doing yourself and hand to somebody, machine or otherwise — is the one covered in delegation without losing control, and it has not changed in two years.
Common questions
Is AI actually worth paying for in a small business?
It depends entirely on whether you can name the task it replaces. Tools that take over frequent, repetitive, verifiable work — drafting, transcription, reformatting data, summarising documents — usually pay for themselves quickly, because the time saved is real and recurring. Tools bought because the category sounds important almost never do. The arithmetic is worth doing explicitly: annual subscription cost against hours saved a week multiplied by 52, valued at what an hour of your team's time actually costs. Then subtract the time somebody now spends checking the output, because that time is real too and it is the number most buyers leave out.
What should you never put into an AI tool?
Anything you would not be comfortable seeing outside the business: client data, personal details of staff or customers, payroll information, anything covered by a confidentiality clause or an NDA, and anything a client has given you on the understanding it stays with you. Personal data put into a third-party tool is still personal data you are responsible for under UK GDPR, and consumer-grade tools may retain or process it in ways you have not assessed. The practical answer for most small firms is a short written rule listing what is banned, one shared business account rather than personal logins, and a named person who decides when something is borderline.
Will AI replace staff in a small business?
In our experience it replaces tasks rather than people, and mostly the tasks nobody wanted. What changes is the mix of what a role contains: less producing, more checking and more of the work that needs judgement. That has a real implication for hiring — the value of someone who can spot when an answer is wrong goes up, and the value of someone who can only produce a competent first draft goes down. The businesses that get caught out are the ones that cut a role first and discover afterwards that nobody left knows enough to check the output.
How do you tell whether an AI subscription is earning its keep?
Set a review date when you buy it, six months out, and write down at the time what task it is supposed to replace and how long that task currently takes. At the review, ask whether that specific task actually got faster and whether the output still needs the same amount of checking. If nobody can point to the task, cancel it — a tool that everyone likes but nobody can attribute a saving to is a subscription, not an investment. Reviewing the whole stack on a fixed date also stops the slow accumulation of small monthly charges that never get individually questioned.
