How much does AI cost? The guides won't tell you.

It is the first question anybody sensible asks, and it is remarkably hard to get answered: what is this actually going to cost me?

So I did what you would do. Yesterday I searched for what AI implementation costs a small business and read the first page of results properly. Here is what is actually there, because it is worth knowing before you take any of it as a benchmark.

Almost all of it is written by people selling AI

Not some of it. Nearly all of it. The guides that come up are published by software development agencies, AI consultancies and implementation firms — companies whose business is you getting in touch after reading the page. That is not a scandal; it is how the internet works, and some of them are perfectly good firms. But it means the market’s reference prices are set entirely by the sellers, with no buyer-side counterweight at all.

The second thing is more telling. The cost figures on those pages generally carry no source. I read one in full — a certified IT services provider, entirely legitimate — which gives data preparation, infrastructure, integration and maintenance ranges in confident detail, none of them attributed to any study, survey or dataset. They read as industry knowledge. They may well be somebody’s honest experience. They are not evidence.

And they disagree with each other

Which you only notice if you read several, so here is the comparison.

On what the extras add to a quoted price: one guide says hidden costs add 150–250%. Another says they push the total 200–400% higher. On what it costs to keep running: one says annual maintenance runs 15–30% of the build cost; another says ongoing annual costs are 20–40% of the initial spend. On the total for a first year: one page puts a small organisation at $5,000 to $50,000, another puts small businesses at $25,000 to $150,000.

Same question. Same month. A thirtyfold spread on the headline number, depending which link you happened to click.

None of those figures is necessarily wrong. But they cannot all be right, and not one of them comes with a method you could check.

The tell: they source the fear, not the price

This is the detail I keep thinking about.

On the page I read closely, the frightening statistics were attributed — the share of AI projects that fail to deliver value, the share abandoned over poor data, each with a research firm’s name against it. The prices were not.

So the anxiety is evidenced and the invoice is vibes. Whether or not anyone intends that, it is the shape of a page designed to make you worried enough to call and vague enough that no quote can later be held against it.

Why an honest general number can’t exist

To be fair to everybody: the reason nobody can give you a straight answer is that there isn’t one.

The cost of putting AI into a business is dominated by the state of the business, not the state of the AI. Two firms of the same size, buying the same thing, can differ by a factor of ten because one keeps its customer records in one place and the other keeps them in four spreadsheets, a filing cabinet and somebody’s head. The AI is close to a commodity. Your data and your processes are the variable, and nobody can price those from a web page.

Which is the useful insight hiding under all this: if a supplier gives you a confident number before looking at your systems, they are guessing, and you will pay for the difference later.

Five questions that produce a real figure

1. What does year two cost? The single most revealing question. Build costs are quoted eagerly; running costs are where the surprise lives. If they can’t answer, you have a build quote pretending to be a price.

2. Who fixes it when it stops working, and is that included? It will stop working. Something upstream will change. Establish now whose problem that is.

3. What has to be true about my data before this works — and who does that bit? This is where budgets actually go. If the answer is airy, the tidying will land on you, unbudgeted, in the middle of the project.

4. What happens to the price if we use it twice as much? Usage-priced things have a habit of being cheap in the pilot and startling at scale.

5. Can I speak to a customer roughly my size who has run this for over a year? Not a logo on a slide. A person, a year in, who can tell you what the second year cost them.

An honest supplier answers all five without flinching. The answers give you a real number for your business, which is the only number that was ever going to be any use.

What argues against the cynical reading

Two things, and they matter.

First, the cost categories in those guides are genuinely right, even where the percentages are decoration. Data preparation, integration with what you already run, training people, and ongoing maintenance are exactly the four things that get forgotten, and the guides are doing you a favour by listing them. Take the checklist; leave the arithmetic.

Second, unsourced does not mean untrue. A firm that has done thirty of these has real knowledge, and it is unreasonable to demand a peer-reviewed citation for a rule of thumb. The problem is not that the numbers are made up. It is that they are presented with the authority of research while carrying the reliability of a hunch, and a buyer cannot tell the two apart.

And one more the guides all miss: the human time to check the output. It never goes away, it rarely appears in anyone’s model, and over a couple of years it is frequently the largest recurring cost of the lot. We have written before about judging tools on the cost of getting a task done properly rather than the sticker price, and this is the same argument one level up.

The short version

Treat every published AI price as marketing until proven otherwise, including any you find here. Get quotes that name the year-two figure. And be suspicious of confidence that arrives before anybody has looked at your data.

Method: on 7 August 2026 I searched for AI implementation costs for a small business and read the first page of results. The characterisation above — that the pages are overwhelmingly published by firms selling implementation services, that their cost figures are generally unattributed, and that where sources are named they tend to support failure statistics rather than prices — is my own observation of those pages, not a claim taken from any of them. The conflicting ranges quoted are reproduced as those pages state them and are attributed to nobody, because none of them attributes them either. No figure in this post is offered as a benchmark.