Almost every conversation about AI starts in the same place. Someone has been told they should be using it, they do not know what that means for their business, and they are somewhere between anxious about falling behind and suspicious that the whole thing is noise.
Both instincts are reasonable. Here is a way to work through it that does not depend on believing anybody’s sales pitch.
Start with the process, not the tool
The question “should we use AI?” cannot be answered, because it is not a question about your business. It is a question about a category of technology.
The answerable version is narrower: which specific thing that we do repeatedly, that costs us time or money or accuracy, might be done better? That question has an answer, and quite often the answer turns out not to involve AI at all.
So the first thing worth examining is the work itself. What happens between an enquiry arriving and someone acting on it. Where things get typed twice. What breaks when a particular person is on holiday. Where people go quiet when asked how long something takes.
Three tiers, and most businesses need the cheapest
Once there is a real process on the table, the options usually fall into three tiers. They differ enormously in cost, risk and how much of your attention they need afterwards.
Tier one: plain automation, no AI
Rules-based work. When a form is submitted, create a record, notify the right person, and chase them in three days if nothing has happened. Nothing is being interpreted, so nothing can be misinterpreted.
This is unglamorous, cheap, and it covers a genuinely large share of what businesses are asking for when they say they want AI. It is predictable, it is debuggable, and when it breaks it breaks loudly instead of quietly producing something plausible and wrong.
Tier two: AI assisting a person
Now something is being interpreted. Summarising a long thread, drafting a first-pass reply, pulling structured detail out of an unstructured document, sorting enquiries by what they are about.
The critical property is that a person still decides. The AI produces a draft and a human approves, edits or bins it. That containment is what makes this tier safe enough for most businesses to adopt without much ceremony.
Tier three: AI acting on its own
The system takes actions without a person in the loop: replying, updating records, making decisions that have consequences.
This is what most of the noise is about, and it is the tier that warrants the most caution. Not because it never works, but because the cost of it being confidently wrong scales with how much it has been left to do unsupervised. It needs monitoring, a way to intervene, and a clear-eyed view of what happens on a bad day.
Very few small and mid-sized businesses need tier three. Quite a few have been quoted for it.
The blocker is almost always the data
Here is the part that rarely comes up in a sales conversation: if your information lives in one person’s inbox, three spreadsheets with different column names, and a shared drive nobody has tidied since 2019, no AI is going to resolve that. It will confidently produce answers built on whichever mess it happened to read.
This is why data comes before tools in any sensible assessment, and why the honest recommendation is sometimes “not yet, sort this out first.” That is a real answer, not a way of avoiding the work.
The upside is that getting your data in order pays off whether or not you ever adopt AI. It is not a wasted step.
What it actually costs
The sticker price is rarely the whole cost, and the surprises land in three places.
- Usage-based pricing. Most AI services charge by volume of text processed. Fine at ten documents a day, uncomfortable at ten thousand. Model it before you commit, using your real volumes rather than a demo.
- Per-seat fees. Bolt-on AI features are often priced per user per month. Multiply by headcount and by twelve before deciding it is cheap.
- The human cost of checking. If someone has to verify every output, you have not removed the work, you have changed its shape. Sometimes that is still worth it. Sometimes it very much is not.
A reasonable way to start
Pick one process. A small, annoying, well-understood one, where you can tell easily whether the output is right and where being wrong is cheap.
Run it for a month. Measure whether it actually saved time, rather than whether it felt impressive. Then decide about the next one.
What you learn from one honest small experiment is worth considerably more than a strategy document, and it costs a fraction of what a platform commitment does.
The short version
You are not behind. Most businesses talking confidently about AI adoption are running one pilot and a lot of enthusiasm.
Work out what is actually slow or expensive. Check whether the cheapest tier resolves it, because it often does. Be honest about whether your data is ready. Then spend money, in that order.
If a second opinion on something specific would help, describe what you have been quoted for. You will get a plain answer on whether it is worth it, including when the answer is that you do not need any outside help at all.
Two things are worth reading alongside this. If the hesitation is regulatory, the EU AI Act asks far less of a small business than the headlines suggest. If it is financial, the real cost is rarely the licence fee.
And if something is already underway and quietly going nowhere, that has a small number of recognisable causes, all of them fixable.
If you have a process in mind and cannot tell whether it is a real candidate, describe what happens today. The honest answer is often that it is not, and that is worth knowing before spending anything.
The questions that come next
These are the ones that follow, roughly in the order people hit them: which tool, what it costs, what an agent actually is, why these projects fail, and one account of what goes wrong when a tool is switched on before anyone checks what it can see.
The legal side is a separate question with its own answer. The EU AI Act, and what actually applies to a small business, covers that.
Common questions
Where should a small business actually start with AI?
With a process that annoys someone, not with a tool. The answerable question is which specific repeated thing costs you time, money or accuracy, and quite often the answer turns out not to involve AI.
How much should we budget to find out?
Less than most expect. A subscription for one named group doing one named task will tell you more than a pilot project, and it is cancellable if the answer is no.
What if we are already behind?
Almost nobody is as far ahead as the noise suggests. Businesses that moved early on a vague brief are frequently the ones now paying for tools nobody opens.