The short answer
An artificial intelligence (AI) project for an SME has three cost items: the initial project (analysis, development, integrations, training), model usage, which is paid for as it is used, and maintenance. The cost is driven by the integrations with ERP systems and data and by the quality of the data you start with, not by the model. A sensible figure exists only after looking at the process: that is why we do not publish prices.
- Three items: initial project, model usage, maintenance and evolution.
- The cost is driven by the integrations with ERP systems and data, not by the model.
- Model usage is paid for as it is used: for office tasks it is normally the smallest item.
- The costs you do not see: the time of the company's own people, training, putting the data in order.
- It is assessed on the time given back to people, measured before and after; the pilot serves to replace the estimate with a measurement.
- No prices in this guide: a serious quote starts from the process, with a written scope.
Note. This guide gives no prices and no comparisons in figures: every project is estimated on the real process, with a written scope. It is for reading a quote, not for producing one.
Why it is hard to get an answer
According to the Artificial Intelligence Observatory of the Politecnico di Milano, in Italy about one company in three struggles to estimate in advance the balance between the costs and benefits of an artificial intelligence (AI) project. And 76% of Italian SMEs, says the Digital Innovation in SMEs Observatory, have not invested in AI and do not plan to. The two things are linked: without an idea of the cost, nobody starts.
The figures from ISTAT, Italy’s national statistics institute, show the same gap: in 2025, 16.4% of Italian businesses with at least ten employees use AI, double the year before, but the figure goes from 15.7% of SMEs to 53.1% of large companies.
This guide does not give a figure, and not out of commercial caution: two projects that look the same can cost very different amounts, and the reason almost always lies in the systems to be connected and in the data you start with. It sets out the cost items, what drives them and how to read a quote. The figure comes afterwards, on the real process.
The three cost items
- The initial project: analysis, development, integration with the systems, testing, training of the people who will use it. It is paid for once.
- Model usage: each request to a language model is paid for according to the text read and written. It is paid for as it is used.
- Maintenance and evolution: processes change, systems are updated, models are replaced by new versions. It is paid for over time.
The three items carry different weights. For an office process in an SME, as a rule, the initial project is the largest item, model usage the smallest, and maintenance the one people forget to ask about.
What makes the cost of the project rise and fall
- The integrations. This is the item that weighs most. Reading and writing in the ERP (the company’s management software), the CRM, email, the document management system: the more closed the system, the more it costs. A cloud service with documented APIs is quick to connect; an ERP that exposes nothing requires custom work; a system installed on the company’s premises years ago, with no documentation, is the most variable item of all.
- The data you start with. Tidy, accessible data lowers the cost; data that is scattered, duplicated or only on paper raises it. We cover this in data and integrations.
- How many things are done together. One process at a time costs less, and the second project reuses what the first one built: connections, permissions, log.
- The control required. Role-based permissions, confirmations before writing, an action log: they are needed, and they are work.
- The complexity of the task. An assistant that answers on documents is simpler than an agent that takes several steps inside the ERP; a linear flow is simpler than a system with several agents.
- Who is involved in the company. An available point of contact shortens the timescale more than any technology.
How model usage is paid for
Language models are paid for by the token, that is, by fragments of text: a page of text is worth roughly a thousand tokens. You pay both for the text the model reads and for the text it writes, and providers publish their price lists per million tokens, which differ by model size: small models cost a fraction of large ones.
What the bill depends on, in order of weight:
- how many requests the company makes per month;
- how much text each request carries: an assistant on documents sends the model the question, the instructions and the passages retrieved from the documents; an agent that works in the ERP takes several steps for each task, and each step is paid for;
- which model: using the smallest model that does the job well is the simplest lever;
- how much is reused: the repeated parts of requests, such as the instructions, cost less if the provider keeps them in memory, and non-urgent jobs can be put off to the reduced-price bands.
In brief: for an SME’s office tasks model usage is normally the smallest item. It should still be estimated beforehand, with the volume assumptions written in the quote, and checked in the pilot: that is where the estimate becomes a measurement.
The alternative is to run a model on one of the company’s own machines, without paying per use: you pay for the machine, the energy and whoever maintains it, and the cost does not depend on usage. The choice is made on confidentiality and on the volume of work: we cover it in the guide AI in the cloud or on-premise.
What maintenance includes
An AI system is not a purchase you make and then forget. Processes change, ERP systems are updated, models are replaced by new versions, and every replacement has to be tested on real work.
The right question to ask before signing is: what does the recurring fee include? Fixes, model updates, small enhancements, user support, monitoring of usage costs. And what it does not include: new features count as evolution, with a scope and a price of their own.
The costs you do not see
- The time of the company’s own people: a few hours a week from those who know the process, for the whole duration of the project. It is the cost most often forgotten.
- Training: those who use the tool must know what to expect and how to check the results. It is also what the AI Act requires.
- Putting the data in order: often the first benefit, but it has to be budgeted for.
- The change: a process that changes needs someone to see it through.
How to tell whether it is worth it
The calculation is simple, and it must be done before and after.
- How many times a month is the process carried out, and how many minutes does it take today?
- What is the value of an hour’s work by the people who carry it out?
- How much of that time can realistically be given back?
- What does the system cost per year, counting the project cost spread over time, usage and maintenance?
An example in hours: a check that today takes 12 minutes, done 600 times a month, comes to 120 hours. If the system gives back two thirds of them, that is 80 hours a month, almost a thousand a year: multiplied by the hourly cost of the people, they give the figure to compare with the annual cost of the system.
The pilot serves exactly this purpose: to replace the estimate in point 3 with a measurement.
How to ask for a quote you can compare
- A written scope: what the system does and what it does not do.
- A defined price for that scope, not an open-ended hourly rate.
- The integrations listed one by one.
- The estimate of monthly model usage, with the volume assumptions.
- What the maintenance includes and how much it costs.
- Who owns the code, data and configurations at the end.
- How the result is measured.
Two quotes with the same written scope can be compared; two figures without a scope cannot.
How we work, and the incentives
At FT Support we publish no prices and no comparisons in figures: every project is estimated on the process, after the first conversation and the process map, that is, the two steps of our AI consulting for companies. Each step has a scope and a price defined before it starts, and no quote is by the hour: we explain this in the method. The first conversation carries no obligation and serves precisely to find out whether there is a process on which the sums add up.
For companies in Italy, part of the spending can be covered by incentives: we have collected them in the guide Incentives 2026 for software and AI in companies (in Italian).
Sources
- Artificial Intelligence Observatory, Politecnico di Milano, press release of February 2026: market and spread of projects (in Italian)
- Digital Innovation in SMEs Observatory, Politecnico di Milano, press release of 21 May 2026 (in Italian)
- ISTAT, Imprese e ICT, year 2025, report of 15 December 2025 (in Italian)
Checked on 12 October 2026. Rules and deadlines change: always check the official source before you decide.