How to choose an artificial intelligence agency. Nine questions to ask before signing.
This guide is written by an agency. The questions apply to us too, and at the end you will find our answers.
You choose badly when you compare prices of different things. Two quotes for «an AI system» can describe a subscription to a ready-made product and a system built on your processes. The questions make offers comparable before you look at the price.
The questions that matter are nine and they touch four things: who keeps what gets built, how the result is measured, what the machine does on its own, and what the AI Act and the GDPR require.
Every question has an expected answer, written before signing. An answer that comes only verbally, or only after the contract, leaves that item uncovered.
Why the choice is hard for an SME
Because judging an artificial intelligence proposal requires skills that companies often lack. According to Istat, Imprese e ICT, year 2025, «the lack of adequate skills holds back AI adoption in almost 60% of the companies that evaluated but then did not make AI investments».
Meanwhile the gap is growing. In 2025, 15.7% of small and medium-sized enterprises with at least 10 employees use at least one artificial intelligence technology, against 53.1% of large ones, and a year earlier they were 7.7% and 32.5%. Whoever arrives now buys in a fast-moving market, where the right questions protect more than technical knowledge.
The nine questions, and the answer you should get
Nine questions, each with the answer you should get. To judge them you don't need to know the technology: it is enough that the answer is written, precise and verifiable. A vague answer on one point amounts to an uncovered point, and it is worth asking again before signing.
| Question | Why it matters | The answer you should get |
|---|---|---|
| Who owns the data, system and accounts at the end?ownership | A system built on your processes is worth something only if it stays yours. If accounts and data are in the provider's name, changing provider means starting from scratch. |
A written clause: the data stays yours, what is built for you is yours upon final payment, the accounts are in the company's name. |
| What happens if we stop paying?the exit | The answer tells you whether you are buying a system or renting it. The fee covers maintenance and services, and it is right that it exists: what matters is what remains when it ends. |
What keeps working, what stops, within how many days you receive data and exports, and in which format. |
| Which number says it worked, and what is it today?the measurement | Without a number measured beforehand, the after cannot be checked and every result becomes an opinion. The starting sum can also be done on your own, as for the cost of replying by hand. |
A single indicator, for example the average first-response time, measured before starting, and the date when it is measured again. |
| Who builds it, and who steps in when it stops?the people | Knowing who puts their hands on the system tells you who will be able to fix it. A name and a response time are worth more than a description of the team. |
The name of whoever builds, the response time for a fault, and who takes over if that person is not there. |
| What does the system do on its own, and what goes through a person?control | A wrong message to a customer costs more than the time saved. The boundary between automatic and approved is decided beforehand, in writing, and can be moved. |
The list of automatic actions and of those that wait for approval, and who can change that boundary. |
| Does the system talk to your customers?AI Act, article 50 | From 2 August 2026 a system that interacts directly with people must let them know it is an artificial intelligence, as required by article 50 of Regulation (EU) 2024/1689. |
Where and how the system declares it, with the exact text of the message. |
| Who processes personal data, and where is it kept?GDPR, article 28 | If the provider, or the models it uses, process your customers' data on your behalf, you need the agreement provided for by article 28 of the GDPR. |
The agreement appointing the data processor, the list of providers used, AI models included, and the country where the data is kept. |
| Who pays when the model changes price?dependency | AI systems rest on third-party models, which change versions and price lists. If the usage cost rises someone pays it, and it is worth knowing who beforehand. |
Which models are used, who bears the usage costs, and what happens to the fee if those costs rise. |
| What does the price cover, and when does the fee start?the scope | Two quotes can be compared only if they describe the same work: what is built, what is left out, by when. Otherwise you are comparing prices of different things. |
A written scope with a fixed price, and the date from which the fee starts, tied to the system working. |
1. Who owns the data, system and accounts at the end of the work? 2. If we stop paying, what keeps working, and in what format do we receive the data? 3. Which number measures the result, what is it today, and when is it measured again? 4. Who builds it, and how quickly do they step in if the system stops? 5. What does the system do on its own, and what waits for approval? 6. If it talks to customers, where and how does it declare that it is an artificial intelligence? 7. Will you send us the agreement provided for by article 28 of the GDPR and the list of providers, models included? 8. Who pays if the model's usage cost rises? 9. What is the scope, what is the price, and when does the fee start?
Three signs that a proposal needs rereading
A result percentage without a starting number. A promise like «70% less time» means something only if someone has measured today's time. If the measurement is missing, the percentage remains a promise, and you should ask how it will be checked.
A price without a scope. «All included» is a convenient formula for whoever sells. For whoever buys, what matters is the list of what is built and what is left out, because that is where extra costs arise once work has started.
A proposal without a name. A document signed by «the team» leaves open the question of who builds and who answers. A name and a direct contact are the simplest way to know whom to call when something stops.
The same questions, asked of Itria
Itria's answers are already written in the Terms, in the principles and on the AI transparency page, and they stand as a commitment. Below are the verbatim sentences, question by question, and where to find them.
| Question | What is written |
|---|---|
| Ownershipquestions 1 and 2 | «Upon payment of what is due, the client remains the owner of their data and holder of what has been built for them.» And: «Every engagement ends with the client owning the system.» Terms, point 8. |
| The measurementquestion 3 | «No work is proposed before looking at where the company loses time or revenue today.» Terms, point 7. |
| The peoplequestion 4 | Itria «operates in person» with Mattia Esposito, who builds and answers. About. |
| Controlquestions 5 and 6 | «The machine prepares, but a person always decides.» Ethics. Where a reply goes out on its own, the owner decides it channel by channel. AI transparency. |
| Data and modelsquestions 7 and 8 | «We do not use a client's data to train models.» The third-party model is named, with the agreement that excludes training. AI transparency. Who bears the usage costs is written into the scope of each engagement. |
| The pricequestion 9 | «The price is fixed and agreed beforehand.» And: «no fee that starts before the system works.» Terms, point 7. |
We take on few clients at a time, because each system is built by one person. That is why the questions on this page are worth asking us too, before starting.
Questions and answers
How do you choose an artificial intelligence agency for an SME?
With nine questions, asked before signing and answered in writing. They concern four things: who keeps the data, the system and the accounts; which number measures the result and what it is today; what the system does on its own and what goes through a person; what the AI Act and the GDPR require.
The questions make different offers comparable before the price. A vague answer, or one that comes only after the contract, leaves that item uncovered.
Who owns the system built by an AI agency?
It depends on the contract, and it is the first thing to get in writing. The most common split separates what is yours, that is the data, configurations and workflows built for you, from what stays with the agency, that is the methods and generic tools it uses with everyone.
It also matters in whose name the accounts of the services used are registered. If they are in the provider's name, changing provider means starting from scratch.
Does an AI system for a company have to comply with the AI Act?
Yes, for the parts that concern it. From 2 August 2026 the transparency obligations of article 50 of Regulation (EU) 2024/1689 apply: a system that interacts directly with people must let them know it is an artificial intelligence.
From 2 February 2025 article 4 also applies, rewritten by Regulation (EU) 2026/1744: whoever uses AI takes measures aimed at supporting the AI literacy of those who use it, as explained on the page on the training required by the AI Act. If the system processes personal data on the company's behalf, the GDPR applies, with the agreement of article 28.
Is a monthly fee or a one-off price better for an AI system?
It depends on what the fee covers. An AI system has costs that continue after delivery: models are paid per use, services have a subscription, and someone has to step in when something changes. The fee pays for these items.
Two things to get in writing: what it covers, item by item, and when it starts. A fee that starts before the system works pays for work not yet delivered.
How do you compare two quotes for an artificial intelligence system?
By bringing them to the same scope: what is built, what is left out, by when, and with which number the result is checked. If a quote does not say it, ask before comparing prices.
Then compare the exits: what stays yours if the relationship ends, and in what format you receive the data. A lower price with a more expensive exit costs more over time.
Notes on sources
- Istat, Imprese e ICT, year 2025, published on 15 December 2025. From here come the 16.4% of companies with at least 10 employees that use AI (8.2% in 2024), the 15.7% of SMEs against 53.1% of large companies, and the sentence on skills, quoted verbatim in translation. The survey covers only companies with at least 10 employees: microenterprises are left out.
- Regulation (EU) 2024/1689 on artificial intelligence, article 50, paragraph 1, and article 4, replaced by Regulation (EU) 2026/1744.
- Regulation (EU) 2016/679, the GDPR, article 28 on the data processor.
- The sentences in the section on Itria's answers are verbatim and come from the Terms (points 7 and 8), from Ethics, from AI transparency and from About. Istat and Itria's pages were opened and read on 25 September 2026; the texts of the regulations are those already checked for this site's AI Act pages.
Bring the nine questions to the first meeting. Ours too.
The first step with Itria is a fifteen-minute video call: we start from where the company loses time or revenue today, and the questions on this page are asked there, before any quote. Write us a line about what weighs on you. We take the first step: what a customer sees when they search for you, and what we found there. Even if we end up not working together.