How much artificial intelligence costs in a company: four spending ranges and what decides them

Bringing artificial intelligence to a business process that already exists costs between 5,000 and 10,000 euros. Building a model from scratch costs between 50,000 and 100,000, for a minimal working version. In between there are two smaller items that nobody usually puts in a quote: the hardware, when the model has to run in house, and the consumption, when the model belongs to somebody else.

The difference between the first range and the second does not depend on how ambitious the project is. It depends on three choices, and they are best made in this order.

What is done	Order of magnitude
An existing model is used, integrated into a process and into the management system	5,000 to 10,000 euros
Hardware is added because the model has to stay inside the company	1,000 to 3,000 euros
The model is paid per use, with nothing installed	from fractions of a cent per operation
A model of your own is built, trained on your own data	50,000 to 100,000 euros for a minimal version

If you want to know which row your case falls into before asking for a quote, write to us: half an hour and the process you have in mind are enough.

The first choice: existing model or model of your own

Almost every company that calls us thinks it has to build something from scratch. Almost none of them actually does.

Building your own model means gathering a large quantity of data, classifying it one item at a time so that it becomes usable, designing the architecture, training it until the results hold, and then exposing it so that the rest of the software can use it. Each of those steps has a cost, and the heaviest ones are people: whoever works on the data and whoever builds the model. That is the path that leads to the six-figure range.

Using a model that already exists means skipping the whole training part and paying only for the work of connecting it to your own trade. Here the spending is not in the model, it is in the integration: understanding which piece of data is needed, where it sits today, how the result comes back into the system used every day.

The question that decides between the two paths is a single one, and it is technical only in appearance: does the model have to learn something that concerns your company alone? If the answer is no, and in the vast majority of cases it is no, an existing model is used.

The second choice: where it runs

A model can run on somebody else's server, and then it is paid per use, or inside the company, and then the hardware is bought or rented.

The order of magnitude changes considerably depending on the task. To read text inside an image, for example from a scanned document, the performance required is minimal and an ordinary computer is enough. To generate images or long texts, dedicated hardware is needed, and there the spending starts at a few thousand euros. For proprietary models, the ones used only through the service of whoever made them, no hardware is bought at all.

Consumption, where there is any, is the item that frightens people most and weighs least. It is measured in fractions of a cent per operation, and becomes relevant only at very high volumes.

The real reason for keeping the model in house is almost never the cost. It is that certain data must not leave, and it is a concern that Italian companies state openly: among those that assessed artificial intelligence without then investing in it, 43.2 per cent name privacy and data protection among the obstacles, according to the Imprese e ICT survey by the Italian national statistics institute ISTAT, 2025 edition, on a sample of 26,246 companies with at least ten employees.

The third choice: how far to integrate

This is the item that moves a quote more than any other, and the one that is talked about least.

A model that answers inside a separate window, where somebody copies and pastes, costs little and changes little. A model that reads the data from the management system, processes it and writes the result back into the management system costs more and changes somebody's work. Between those two lies the difference between a demonstration and a tool.

In Italian manufacturing, 59.9 per cent of companies with at least ten employees already use a management system, again according to the same ISTAT survey. Which means that in most cases the point of attachment already exists, and the question is not whether to connect, but how far.

What really costs and does not appear in the quote

The data. Not its quantity: its shape.

A model that has to answer about the status of an order needs that status to sit somewhere in readable form, kept up to date by the people doing the work, with a meaning that stays stable over time. If that piece of data lives today in three spreadsheets and in the heads of two people, the project is not an artificial intelligence project: it is first a tidying-up project, and that is where the cost is.

This is not our opinion. In the same ISTAT survey, the unavailability or poor quality of the necessary data is the third declared obstacle, at 45.2 per cent, ahead of privacy and costs. The first, at 58.6 per cent, is the lack of internal skills. High cost comes fifth, at 43.0 per cent.

Put differently: when a project of this kind goes wrong, it is almost never because it cost too much.

Two cases, with the figures

Custom image generation, with the model installed by the client. Realistic photographs of dishes were needed for a recipe book. The model chosen was free and open, so the cost of use was zero, but it had to be configured and it needed suitable graphics cards. Total planned spending, hardware and consulting together: 2,000 euros.

Integration of an existing model into software already in use. Counting only the hours that concern the artificial intelligence part, and not the rest of the product, the spending fell between 4,000 and 5,000 euros. The model's consumption, in operation, is worth a few cents per operation.

These are two useful extremes: in the first case the spending is almost entirely hardware, in the second almost entirely work.

How to read a quote

Three questions, and they serve to tell whether whoever wrote it has understood the process.

The first: which data it uses, and where that data sits today. If the quote does not say, whoever wrote it has not been to look yet.

The second: what happens when the model gets it wrong. It always gets it wrong, in a share of cases. A serious project states in advance who checks, when, and what happens to the piece of work that came out wrong.

The third: what remains if tomorrow the model's supplier changes price or terms. The honest answer is not "nothing changes", it is knowing how much work switching supplier would take.

The process on which the calculation can be made in advance

It is the most boring and most repetitive process you have, the one whose cost in hours you already know. It is the scope that fits inside the 5,000 to 10,000 euro range, and the only one on which it can be said in advance whether it is worth it, because the term of comparison already exists.

If you already have the process in mind and want a reasoned estimate, write to us: half an hour, and we will talk it through.