You have heard the loud version. AI is going to take every job, end every industry and probably the world along with it. You have also heard the other loud version, where it solves everything by Thursday and anyone not using it is finished.
Both are exhausting, and neither describes what actually happens when a small business starts using this stuff. The real story is much more boring, and much better: a category of software that used to be out of reach became cheap enough to bother with.
I build these things for a living, so treat me as an interested party. That is exactly why the numbers below are specific rather than exciting. You can check them.
What actually changed
The thing that changed is not that computers got clever. It is that a handful of capabilities that previously needed a specialist team, a pile of training data and six months stopped needing any of that.
Reading a document and pulling out the useful fields. Answering a question in plain English. Sorting a thousand pieces of feedback into themes. Turning a rambling voice note into a written summary. Every one of those used to be a project. Now each is a few lines of code calling something that already works.
Pick a job below and see what it would have taken a decade ago against what it takes now.
The same job, then and now
None of these capabilities are new. What changed is that they stopped being something you build and became something you call. You are no longer paying to create the ability to read an invoice. You are paying for the twenty lines of code that point it at your invoices.
The part where I have to be careful
You will see people claim that software now costs nothing. That is not true, and repeating it would make everything else here less believable.
What is true is narrower and still remarkable. A single, well-defined capability that used to cost tens or hundreds of thousands can now often be built for under a thousand pounds. A thing that reads your supplier invoices and drops the totals into a spreadsheet. A thing that drafts a first reply to the same question you answer forty times a week.
What has not collapsed:
- Knowing what to build. Most failed projects were not badly built. They were the wrong thing, built well.
- Wiring it into how you actually work. The capability is cheap. Getting it talking to your accounts package, your inbox and the spreadsheet three people depend on is where the hours go.
- Keeping it alive. Software needs looking after. Budget 15% to 20% of the build cost a year, as our buy, subscribe or build guide sets out.
- Anything genuinely large. A marketplace with payments and two sides is still a serious build. Ask Old Hands.
My lot took proper months to build, and it would still take months. Escrow payments and a booking calendar two sides rely on do not come out of a box. The cheap bit is the clever bit, not the plumbing.
The fears, taken seriously
Dismissing worries about AI is its own kind of nonsense, and it is usually done by people selling something. Most of the fears point at something real. They are just aimed slightly to the left of where small businesses actually stand.
Pick whichever one nags at you most.
Every answer from the questions above, if you would rather read them all at once.
In a small business, it usually replaces the worst hour of the day
If you answered: It will replace people
The honest position is that this is a real question at the scale of whole industries, and a much smaller one inside a six-person company. Small businesses are not overstaffed. They are usually one person doing four jobs badly because there is no time.
What actually goes is the re-keying, the hunting for an email from March, the writing of the same reply for the fortieth time. In practice that has meant fewer terrible Sunday evenings rather than fewer people.
Where it does bear on hiring, it tends to shift the answer from 'hire an admin' to 'automate the admin and hire someone who grows the business'. Our cost of hiring calculator puts real numbers on that comparison.
Yes. Which is why you do not point it at anything you cannot check
If you answered: It makes things up
This one is not a myth and you should not let anyone tell you it is. These systems produce confident, fluent, wrong answers, and the fluency is exactly what makes the wrongness dangerous.
The way round it is unglamorous: only give it work where a mistake is visible immediately. Drafting a reply a person reads before sending is safe. Silently filing numbers into your accounts is not.
Our guide on where AI earns its place has the four-part test we use for deciding whether a task qualifies.
A fair concern, with a boring fix
If you answered: My data is not safe
The risk is real but rarely dramatic. It is almost never a hacker. It is someone pasting a customer list into a free tool at half past four on a Friday.
The fix is a one-page policy: which tools are approved, what must never be pasted in, and who checks output before it reaches a customer. Six lines is genuinely enough.
Beyond that, business-tier API access does not train on your data by default, and anything sensitive should be running through your own systems rather than a consumer chat window.
The valuations might be. The capability is not going back in the box
If you answered: It is a bubble
There is very probably a bubble in what these companies are worth. Plenty of them will not exist in five years, and being sceptical about the funding rounds is entirely sensible.
That is a different question from whether the technology works. The dot com bubble burst spectacularly and online shopping did not go away. The capability to read a document or summarise a thread is now cheap, widely available from several suppliers, and not dependent on any one of them surviving.
The practical lesson is about lock-in rather than abstinence. Build so you could switch supplier, and do not bet the business on one vendor's pricing staying put.
You are almost certainly early
If you answered: I have left it too late
The loud minority posting about their workflows are not representative. Most small businesses have done nothing beyond one person trying a chat window a few times.
Arriving late is also cheaper. The tools are better, the prices have fallen, and the early adopters have already discovered which ideas do not work, mostly at their own expense.
If you want a two-minute read on where you actually stand, the AI readiness scorecard will tell you, and it is not trying to sell you anything.
What this means for a business your size
Here is the bit the doom coverage keeps missing. For twenty years, decent internal software was something only large companies had. Not because small businesses did not want it, but because £80,000 for a system that reads your invoices is an impossible sentence when your whole year's profit is £60,000.
That gap has closed. The things that a competitor with two hundred staff had built into their operation are now within reach of a business with six. That is not a threat to small businesses. It is the single biggest levelling of the field in decades.
It is the same story on my side of the fence. A brand strategy and a month of planned content used to mean an agency retainer. Now it means answering some questions about your business properly. The expensive part was never the typing.
Where to start, without wasting a year
The businesses that get value do the same unglamorous thing. They pick one job, measure it, and check whether the change actually helped.
Find the task you resent
Frequent, roughly the same shape each time, currently manual, and easy to check when it goes wrong. That last one rules out anything touching your accounts unsupervised.
Time it before you change anything
Two weeks of honest measurement. Without a baseline you will be arguing about impressions in six months, and impressions always favour whoever bought the tool.
Run both ways in parallel
Do it the old way and the new way side by side for a fortnight. You are looking for how it fails, not for a success rate.
Decide on the evidence
If it has not clearly won, stop. A clean no after five weeks is a far better outcome than a year of vague dissatisfaction and a subscription nobody cancels.
Buying a tool before choosing a job for it. It was the pattern with CRMs, it was the pattern with intranets, and it is the pattern now. The subscription is never the expensive part. The year spent looking for something to justify it is.
The genuinely cheerful conclusion
The apocalyptic coverage is mostly about very large questions: labour markets, national economies, what happens if the models keep improving. Those are real arguments and worth having.
They are also not the question in front of you. Yours is smaller and much more answerable: is there a job in this business that eats hours, follows a pattern, and could now be done for a fraction of what it would have cost five years ago?
For most businesses the answer is yes, several. And the cost of finding out has fallen so far that being wrong is no longer expensive. That is the actual news, and it is good.
If you want to talk one through, that is literally what I am for. Bring the task that annoys you most and we will work out whether it is worth automating, including the times when the honest answer is no.
The readiness scorecard tells you where you stand. The guide on where AI earns its place covers which tasks qualify. And if you would rather just describe the problem to a human, hello@verislabs.co.uk reaches one.