What AI automation actually costs a small business

Nobody can price it from a one-line description. Here is what moves the number, and how to work out yours before anyone quotes you.

Ask three suppliers what AI automation costs a small business and you get three honest, different ranges. The price has almost nothing to do with the AI in it. It is set by how many of your systems it touches, and how badly those systems behave.

Why is there no single answer to what AI automation costs a small business?

Because "AI automation" is a label, not a thing. It covers a rule that files an attachment in the right folder and a system that reads a thousand supplier invoices a month and books them. Not the same job.

Two automations with the same one-sentence description can be different projects entirely. "Automate our invoice processing" is a sentence, not a scope. The price lives in what it leaves out: how many places the invoices arrive from, how many layouts, what happens to the ones the machine is unsure about.

Which makes "how much does business automation cost" the wrong first question. The answerable one is narrower: what does it cost to automate this task, in your systems, at your volume.

A figure on an agency's pricing page is the price of their smallest project, not of yours.

What drives the price, the AI or the integrations?

The integrations, nearly every time. Model calls are close to a rounding error in a small-business automation; the work sits in the plumbing on either side of them.

Most of what gets sold as AI automation is workflow automation with a model doing one step of it. Price it as a workflow, then: reading from wherever the data lives, writing to wherever it has to land, and coping with the days it arrives wrong.

Six things move the number:

• How many systems it touches. One in, one out is cheap. Every extra one adds a connection, credentials and a new way to fail.
• Whether they have a real API. A modern tool with documented endpoints is an afternoon; a twenty-year-old package whose database you must not write to is weeks.
• How messy the input is. Structured data is cheap to read. Photos of receipts and threads where the answer is in the fourth reply are not.
• How much it may act alone. Drafting for a person to approve is far cheaper than acting unsupervised, which needs limits, checks and a trail of what it did.
• How many exceptions you cover. The cases that follow the rule are most of the value and a fraction of the work; chasing the stubborn few is where budgets go.
• Who can change it afterwards. Something only your supplier can edit is cheaper to build and more expensive to own.

Which model it runs on is not on that list. Swapping one for another is usually a line of configuration.

What is the cheapest automation you can build?

The one where data gets retyped from one system into another. It happens constantly, the steps never change, and it needs no judgement — so little or no model, and little to go wrong.

It should be first because it is small enough to finish, and finished automations compound where half-built ones just add maintenance. We set out the full order in our note on what to automate first: retyping, then reports assembled by hand, then routing, then drafting, and decisions last.

What does a first pilot cost, and how long does it take?

A range against a bounded scope, not a fixed price — and weeks, not months, to run it. A price quoted before anyone has mapped the workflow is a guess, and a first automation quoted in months has the wrong scope.

What you want from a pilot is one workflow, automated end to end, running beside the manual process so you can compare the two.

Ours starts from €3,500, excl. BTW: the audit plus your first flow live.

We publish that shape on our AI and automation page: about a week mapping the work with the people who do it, two to three weeks building the pilot, then a roll-out that extends it to the rest of the workflow. Four to eight weeks in total, typically, before the first automation runs in production.

On price, compare the shape rather than the total: how many hours, at what rate, over how many weeks, and what happens if it runs long. Ask whether the mapping week is billed, and what you own at the end of it. The map should be yours either way.

What does it cost to keep an automation running?

Something, every month, forever. It is the line most quotes leave out, and the one that decides whether the thing still works in a year.

Three costs arrive after go-live:

• Usage. Model calls, API calls, whatever it runs on. Usually modest at these volumes, but it scales with use, so learn the cost per run before the work grows.
• Monitoring. Something has to notice when it stops, and "someone will spot it" is not that something. A silent failure is worse than no automation, because people have stopped checking.
• Drift. A supplier changes an invoice layout. A tool ships a new API version. Someone adds a field to a form. The automation was correct when it was built and is quietly becoming wrong.

Budget for it as a small standing cost. Ours starts from €250/month, excl. BTW. Before you sign the build, ask what the monthly arrangement covers, how fast someone responds when it breaks, and what a change costs.

How do you know the ROI clears the cost before you build?

Do the arithmetic on your own numbers, before anyone quotes you. You need two figures: what the task costs today, and what it would cost to run automated.

The first is hours the task takes each week, times the loaded hourly cost of the person doing it, times fifty-two. Say three hours a week at €40 fully loaded: a bit over €6,000 a year, before errors. Then add what the mistakes cost — the invoice paid twice, the enquiry nobody answered, the order typed in wrong.

Compare that against the build price plus a year of running costs. Paying for itself inside a year, on a task that is not going away, is usually worth doing. Three years means the task is too small or the scope too big — and the usual fix is a smaller scope.

If the build does not return more than it costs, do not start it.

One caution about saved hours: they only become money if the time goes to something that earns. Four hours a week handed back to five people is real relief and rarely a headcount.

What should you be asked before anyone quotes you?

Six questions. If nobody has asked them, the number you are holding is a guess.

• Which systems the data comes from and lands in, and whether you can get API access to each.
• How often the task runs, and how long it takes each time.
• What the exceptions look like, and roughly how many there are.
• What happens today when it goes wrong, and who notices.
• What the machine may decide alone, and what a person must approve.
• Who owns and edits the automation once it is live.

A supplier who prices a fixed build off a paragraph has either padded the number to cover the unknowns, or will come back for them later at your expense. Neither of those is a quote.

Those six answers are most of what a thirty-minute call is for, and that is exactly what we run: bring the person who does the work today, describe the problem rather than the solution you arrived with, and leave with a yes, a no, or the name of someone better suited.

If that is your week too, thirty minutes is enough to find out.

Book 30 minutes →