How much does AI cost for a small business?

How much does AI cost in practice

Short answer: it depends on what you are buying, and the range is wide. A subscription to a chatbot tool costs about as much as a streaming service, per person, per month. A custom AI system built into your operations, one that reads invoices, drafts replies, or checks stock against supplier data, can run into thousands of euros to build and then a smaller amount every month to run. If someone quotes you one flat number before asking what you actually need, be skeptical. There is no single answer to how much AI costs, only an answer for your specific case.

We hear some version of this question from almost every business owner we talk to. They have read that AI is cheap now, or that it is expensive, and both are true, depending on which part of the stack they are looking at. The two of us build these systems ourselves, so the price we quote reflects work we can point to line by line, not a markup on someone else's software.

Three places the cost shows up

Cost splits into three buckets, and each behaves differently.

Tooling. Off-the-shelf products, Copilot, ChatGPT Team, various point solutions, charge a flat fee per seat, per month. Predictable, low commitment, and easy to cancel if it does not work out. For most single-person tasks, drafting text, summarizing a document, answering a question about a spreadsheet, this is the whole answer, and you do not need a consultant for it.

Building. A custom project, an agent that handles your specific workflow, a system tied into your own data, is paid for once as a build. This is where a consultancy like ours does most of its work, and it is the part with the widest range, because scope varies so much. A small automation that saves one person two hours a week costs far less to build than a system that touches five departments, reads from three different data sources, and needs proper testing before anyone trusts it with real transactions. The build price also depends on how messy your existing data is. Clean, structured records are cheap to connect. Years of inconsistent spreadsheets and PDFs cost more to prepare than the AI part of the project does.

Running. Once a system is live, it keeps costing money. Most AI tools bill by usage: the underlying API meters tokens, so a system that handles more requests costs more that month, by definition. This is the part people forget to budget for. A prototype that costs almost nothing to demo can cost several hundred euros a month once it is processing your actual volume, and that number moves with your business, not with your invoice.

The cost nobody prices in: upkeep

A build is not a one-time expense that ends at launch. The model behind your system will get replaced by a newer version within a year or two, and providers do retire older models on a schedule. When that happens, the prompts and checks that made your system reliable need a review, sometimes a small rewrite. Budget a handful of hours every quarter for this. The ground underneath a model shifts even when your own workflow does not.

The same goes for monitoring. An automation that quietly misreads one field on an invoice, or drifts on edge cases nobody tested for, is worse than one that fails loudly. Part of what you are paying for in a proper build is someone checking its output on a schedule, not just switching it on and walking away.

Why a fixed timeline is hard to promise for agent work

One thing we watch closely: how badly AI agents themselves misjudge how long a task will take, because that misjudgment turns straight into cost when you are billed for compute time or consultant hours on top of it.

A study published this week looked at how well coding assistants such as Claude Code and OpenAI's Codex estimate their own work. Across 200 test tasks, both models were consistently over-optimistic beforehand and often wildly wrong in hindsight. Codex misjudged elapsed time by up to ten times the actual duration. Claude Code was off by about three times on average. Both systems also rated their own finished work roughly 20 percentage points better than it actually scored, even on tasks that had mostly failed.

Agents are still worth using for the right task. But price the work with a capped scope instead of an open-ended "let the agent run," so that estimation risk sits with the people building the system, not with your invoice. That is how we scope agent-driven builds for clients: a fixed discovery phase first, then a capped build phase, with the running cost quoted separately once we know the real volume.

Build versus buy

Before you commission anything, check whether an existing tool already does most of what you need. For drafting emails, summarizing documents, or searching your own files, it usually does, and a subscription beats commissioning a build every time.

Custom work earns its cost when the task is specific to your business in a way no off-the-shelf product accounts for: matching purchase orders against contracts with your own thresholds, checking incoming shipments against rules only your inkoop team knows, routing sales leads the way your team actually works rather than the way a generic CRM assumes. That is the kind of project where our AI services start, with a short discovery phase to figure out whether a build is worth it before we charge you for one.

What to ask before you sign anything

Ask any AI vendor three questions: what happens to the price if usage goes up, what happens if the first version does not work, and who owns the system once it is built. A consultancy that cannot answer those clearly, or that only quotes the build and skips the running cost, is leaving out the part of the bill that adds up over a year.

Ask, too, how they price the discovery phase. A short, fixed-price discovery step before any build tells you two things: whether the project is worth doing at all, and roughly what the build will cost, before anyone has spent real money on it.

And if what you actually need is a better subscription, not a custom system, we will tell you that too. Half the value of talking to people who build this stuff themselves is that they will say plainly when you do not need them yet.