What AI actually costs an SME

Nobody will give you a straight number, so here is one. The three costs of a first AI project, what drives each, and how to budget without a nasty surprise.

Almost every business owner we speak to has the same worry about AI, and it is a fair one: nobody will tell them what it costs. The demos are free, the pricing pages talk about credits and tokens, and the real number only appears after you have committed. Here is the plain version, in the categories that actually show up on your bill.

There are three costs to a first AI project, and they behave very differently. Confusing them is how budgets blow out.

Setup: the one-off cost of building it

This is the work to turn an idea into something that runs: understanding your process, connecting to the systems that hold your data, building the automation or tool, and testing it against real cases rather than tidy examples. For a focused first project, this is the largest line item and it happens once.

The honest driver of this cost is not the AI, it is the messiness of your existing setup. A clean process with data in one place is quick. A process that lives across three disconnected systems and a shared inbox takes longer, because most of the work is untangling that before any AI is involved.

Running: what it costs every month it operates

Most modern AI is billed by usage. Every time the system reads a document or drafts a reply, there is a small per-use charge from the underlying model provider, plus whatever the software it runs on costs to host. For a typical SME first project, this monthly figure is usually far smaller than owners fear, often in the range of a modest software subscription rather than a salary.

The trap is projects with no cap and no monitoring, where usage quietly climbs. A well designed system has limits built in and someone watching the number, so a surprise is caught in week one, not on the quarterly bill.

Support: keeping it working after launch

This is the cost people forget, and it is the one that decides whether the project still works in a year. Models change, your process changes, an edge case appears that nobody planned for. Support covers monitoring, small adjustments and someone to call when something looks wrong.

You can buy this as an ongoing managed arrangement or handle it in-house if you have the skills. What you cannot do is assume it is zero. Software that nobody maintains does not stay still, it drifts.

A realistic way to budget

Treat setup as a project, running and support as a subscription. Start with one process small enough that the setup cost is easy to justify against the hours it saves, then let the monthly costs prove themselves before you expand. A first project that pays for its own running and support inside a few months is a sound one. A first project that needs a leap of faith to justify is usually scoped too big.

If a vendor cannot break their quote into these three buckets, that is worth noticing. Vague pricing is usually a sign of vague scope, and vague scope is where budgets go to die.