How much AI costs a small business (and the lesson of the St Albans millstones)

"How much does AI cost?" It's the question we hear most, and the short answer is: from a few dozen euros a month to several tens of thousands of euros, depending on what you want to achieve. The useful answer is longer, and it's mostly about the costs that don't appear on the price list. Let's take it step by step.

The lord's mill and the St Albans millstones

In medieval Europe a watermill was expensive infrastructure: you needed a watercourse, a building, gears, millstones. Whoever could build one, usually the lord or a monastery, often required local people to grind their grain there, paying a share of the flour. Historians call this banal rights, or banalités: a monopoly on the service, with a "fee" every time you used it.

At St Albans, in England, this took a memorable turn. The townspeople, under the lordship of the abbey, claimed the right to grind at home with their own hand-mills. After one of many disputes won by the monks, the confiscated millstones were used as paving in the abbey. In June 1381, during the great English Peasants' Revolt, a group of rebels entered the cloister, lifted those stones, carried them out and smashed them; the abbey's chronicler, Thomas Walsingham, says the fragments were handed out among the people "like blessed bread" after Mass (the episode is reconstructed in this study of William Grindecobbe and the St Albans uprising).

You may ask: Professor, peasant revolts now? Yes, because those millstones say something precise about the cost of technology. The St Albans townspeople's problem wasn't the mill itself, a very useful machine. It was depending entirely on someone else's mill, on their terms, with no alternative.

Well, your company today faces the same choice, in a gentler form. Using AI "on subscription" inside other people's platforms is convenient and cheap at first. But the real cost is measured over years: how tightly that vendor binds you, how much it costs to leave, how much of your know-how ends up inside a tool you don't control.

The four spending levels

It helps to think in levels. The figures for levels 3 and 4 are indicative orders of magnitude, based on our experience with small-business projects: every real quote depends on processes, data and integrations.

For level 1 you can check official price lists directly, for example the Claude pricing page (Team plan at 20 dollars per seat per month billed annually, 25 billed monthly, as of this writing) or ChatGPT's. Prices change often: always check before deciding, and remember that list prices usually exclude VAT.

Pay-per-use: paying "by the sack of grain"

Levels 3 and 4 often include a variable cost: calls to models via API are billed by the amount of text processed (so-called tokens). At small-business volumes this is usually modest, but it must be estimated: an assistant reading a thousand documents a day costs differently from one answering ten emails. It's the modern version of the miller's share: you pay each time you use it.

Hidden costs (the ones that really matter)

The monthly fee is the most visible part and almost always the least important. These are the items companies underestimate:

• People's time. Mapping processes, trying tools, correcting early results. In the first weeks a project takes time before it gives any back.
• Data cleanup. Duplicate records, scattered documents, inconsistent file names. Often the longest job.
• Integration. Connecting AI to the ERP or CRM may require development, access rights, permissions.
• Training. A tool nobody knows how to use produces nothing. More in AI training for employees.
• Compliance. Vendor contracts, privacy notices, possibly a data protection impact assessment. See our guide on GDPR and artificial intelligence.
• Maintenance. Models changing version, prompts to update, automations that break when a connected application changes.
• Exit cost. What would it cost to switch vendors in two years? If your data and procedures are locked into a platform, the answer may be "a lot". Those are the millstones set into the abbey floor.

How to estimate the return before spending

The basic calculation is simple, and worth doing on a sheet before talking to any vendor.

1. Hours saved per month on the chosen task (measured, not guessed).
2. Multiply by the hourly cost of the people involved.
3. Subtract recurring costs (licenses, usage, maintenance).
4. Divide the upfront cost (project, training, internal time) by the net monthly saving: that's how many months it takes to break even.

Typical scenario: a distributor's order desk

Picture an electrical-supplies distributor with a three-person order desk that receives orders by email and retypes them into the ERP. Suppose, for the example, that manual entry takes 60 hours a month in total and that an automation with AI order reading saves 40 of them, at a company cost of 30 euros an hour: 1,200 euros a month of freed-up time. If recurring costs are 200 euros a month and the initial project costs 10,000 euros, break-even comes in about ten months.

These are hypothetical numbers chosen to show the method: yours will differ. But note two things. First, freed-up time isn't a saving unless it's used for something useful, like serving customers or reducing errors. Second, if the hours are guessed, the whole calculation is fragile: measure first.

Subscription, custom, or both?

There's no single answer, but there are practical criteria:

• Start at level 1 to learn where AI really helps your people. It's cheap and teaches a lot.
• Move to custom when you have a repetitive, measured process with enough volume to justify the investment.
• Insist on portability: your data, prompts and automation documentation must stay yours, in formats you can export.
• Avoid long contracts until you have field data. A 30–60 day pilot is worth more than any demo.

So by all means use the mill. But keep your own millstones at home.

Can incentives lower the bill?

In part, yes, at least in Italy. For example, the MIMIT Cloud and Cybersecurity voucher covers 50% of eligible spending on cloud services and subscription software, including suites with built-in AI features, up to a 20,000-euro grant, with applications from November 10, 2026, according to the Ministry's official page. All verified measures are in our guide to AI incentives for small businesses. One piece of advice: an incentive lowers the cost, but it doesn't turn a useless project into a useful one.

Further reading

This article is part of our complete guide to AI for small businesses. If you want a realistic estimate for your company, see our page for small and mid-sized businesses and our automation and AI strategy services, or book a free intro call. Bring your measured hours: we'll work out the rest together.