What AI implementation costs and what drives the price

"How much will it cost?" is the first question at almost every AI meeting. The honest answer is "it depends", which helps nobody. So here is what the price actually depends on, which costs you see up front and which show up later, and how to judge the budget before you sign anything.

First: what problem are you solving

The price does not depend on "how much AI" you want but on what the solution has to do. There are three very different levels.

A ready-made tool for your people. Business subscriptions such as ChatGPT Business or Microsoft 365 Copilot cost roughly €18 to €27 per user per month (between 20 and 30 US dollars at published prices as of November 2026, less with annual billing). They suit writing, summarising and working with an employee's own documents. There is almost no implementation cost, but also no connection to your systems.

AI connected to your data and systems. For example an assistant that answers from your internal policies, or processing of incoming invoices that writes into your ERP. Here most of the cost is not the model but the analysis, the data preparation and the integration.

Your own or a fine-tuned model. Rarely needed: when the task is very specific, the volumes are large or the data must not leave your infrastructure. It is the most expensive level and only worth it for a clear economic reason.

The lines in the budget

Whatever you choose, an AI project budget is made of the same lines. It is worth seeing all of them early, because the ones most often underestimated are not the obvious ones.

Process analysis. What happens today, what it costs and where the errors are. Without it there is nothing to compare the result against.

Data preparation. Documents in different formats, duplicate records, missing fields. Gartner predicts that through 2026 organisations will abandon 60% of AI projects that are not supported by AI-ready data. Cleaning costs money, but skipping it costs more.

Integration. The connection to your ERP, CRM, email or document archive, access rights, and writing results back. This is usually the largest one-off line.

Model usage. Paid by the volume of text (in tokens) or as a per-user subscription.

Maintenance and monitoring. Models get updated, and providers retire older versions on a published schedule. Someone has to watch answer quality and keep the solution up to date.

Your people's time. The experts who explain the process and check results at the start are a cost too, even if it never arrives as an invoice.

What the model itself costs: an example

Model prices look small because they are quoted per million tokens (a token is roughly part of a word). Take an example: you process 1,000 documents a day, each about 3,000 tokens, and each answer is about 500 tokens.

That is 3 million input and half a million output tokens a day. With a mid-range model (2 dollars per million input and 10 dollars per million output tokens) it comes to about 11 dollars a day, or about €300 a month. With a small model (0.10 and 0.50 dollars) it is under 1 dollar a day, or about €15 a month. If you do not need the result immediately, batch processing at some providers halves the price.

The takeaway: for most business tasks the model bill is the smallest line in the budget. What matters far more is whether a smaller, cheaper model can do the job. You find that out by testing on your own real examples, not by assuming.

Pricing models

Companies that implement AI usually offer three ways to pay. Each has its place.

Fixed price. Works when the scope is clear. An honest fixed price comes after a short analysis, not before it. A quote for a whole AI project after one conversation means the risk is hidden somewhere in the price.

Time and materials. You pay for the time actually worked. Suits a task that is still taking shape. It should come with a cap and regular reports.

Subscription or managed service. A monthly fee for a working solution, support and monitoring. Convenient if you have no team of your own to maintain it.

A good practice is to split the project into stages: first an analysis and a small pilot with a measurable result, then a decision whether to continue. Your risk stays small and every next decision is made with real data.

Why projects end up costing more than planned

In 2024 Gartner predicted that at least 30% of generative AI projects would be abandoned after proof of concept by the end of 2025. The reasons given were poor data quality, inadequate risk controls, escalating costs and unclear business value. Three of the four are visible already when the budget is planned.

In our experience costs run over most often when the data turns out messier than expected, when nobody in the business owns the result, and when the pilot was tested on hand-picked "clean" examples instead of the real flow.

When you do not need a project

If the task is writing emails, summarising meetings or translating, and it needs no access to your systems, a ready-made business subscription is probably enough. If the volume is small, say a few documents a week, automation rarely pays off. And if the process is not clear today even to the people doing it, AI will not sort it out. Sort out the process first.

How to approach it

Before you ask for a quote, write down three things: which task you want to lighten, how many times a month it happens and how long it takes each time. With those three numbers any serious partner can tell you whether it makes sense, before you even talk about price.

If you would like to go through your case together, write to us. We will tell you honestly, even if the answer is "you do not need this".