In 2025 the share of Italian companies with at least 10 employees using artificial intelligence doubled, from 8.2% to 16.4%, according to the Istat report Imprese e Ict 2025. The gap by company size remains wide: AI has reached 53.1% of businesses with at least 250 employees and 15.7% of SMEs. The most telling figure, however, is a different one. A third of the companies adopting it (33.4%) report no business purpose at all, and 83.3% of these are small firms with 10 to 49 people, as Key4biz sums up.
So many companies are trying out tools before deciding where to use them. Meanwhile the market is moving fast: according to the Osservatorio Artificial Intelligence at Politecnico di Milano, it grew by 50% in 2025, reaching 1.8 billion euros. The eight questions below are for people who run operations, finance or IT and need to choose the first process to automate without burning the budget on a pilot that nobody ends up using.
Start with the problem
The authors of the analysis published by Key4biz, Gambaro and Lo Turco, suggest starting from a specific business problem and dropping the vague goal of "adopting AI". Their examples are very concrete: preparing a quote faster, predicting a breakdown, reducing waste, planning orders better, eliminating repetitive manual tasks.
They also describe where they struggled most. Choosing the tool mattered little; the real work was mapping processes, improving data quality, assigning responsibilities and getting different skill sets to work together. This matches our experience, because when we assess an AI workflow with a company, much of the time goes into understanding how the process works today, exceptions included.
To use the checklist, pick three candidate processes. They could be sorting the emails that land in the general inbox, entering supplier invoices into the management system and preparing quotes. Then ask yourself the same questions for each of them.
The 8 questions to assess a candidate process
1. Volume and repetitiveness
Automating business processes with AI pays off when a task recurs often and follows similar rules. An info@ inbox that receives hundreds of messages a week, each of which has to be read and forwarded to the right department, is an excellent candidate: an LLM classifies the request, extracts the useful data (customer, order number, urgency) and routes it. The quarterly report that takes the controller two days of thinking is a much weaker one. Count how many times the task comes up in a month and multiply by the minutes it takes: that gives you the hours at stake.
2. Data quality and access
This is where the starting questions proposed by Gambaro and Lo Turco come in handy: what data do we have, where is it, who produces it, how reliable is it, and can we connect it. Scarce or poor-quality data also appears among the obstacles reported to Istat by companies that considered AI without adopting it.
With documents the start is usually easy, because current models can read invoices, delivery notes and orders even from scans. Quotes are a different matter. You need up-to-date price lists, discount rules and a history of past offers, which often lives in scattered Excel files or in the memory of your most experienced salesperson. The same authors note that a lot of useful information sits in people's experience and in no database at all.
3. Integration with ERP and management systems
An AI agent that reads an invoice and leaves a person to retype it into the management system saves very little. The gain comes when the data flows into the ERP, CRM or ticketing system on its own. Check whether your system exposes documented APIs, accepts file imports or at least has a reachable database. With a closed, outdated system the integration can cost more than the AI part, and at that point it is worth looking at the signs that it is time to replace your management software.
4. Cost of errors
Every model gets things wrong some of the time, so the useful question is about the consequences. An email sent to the wrong department is fixed with a forward. A wrong amount in a quote that has already gone out, or an incorrect VAT rate in the accounts, costs far more. For your first projects, choose processes where errors are recoverable, or build in a human approval step: the agent prepares the draft, and the operator checks and confirms it.
5. Measurable KPIs
If nobody knows today how long a task takes, tomorrow it will be impossible to say whether AI has shortened it. Before you start, set two or three indicators and measure them on the current situation: average response time to a request, minutes to record a document, quotes completed per week, share of manual corrections. That 33.4% of companies with no stated purpose, which Istat reads as a sign of adoption that is still experimental, is exactly the situation to avoid.
There are some figures on expected results. A Banca d'Italia study from March 2026 (Ropele and Tagliabracci), reported by Key4biz, links AI adoption to a 5.2% increase in value added per employee and an 11.9% increase in gross operating margin per worker. The sample covers companies with at least 50 employees, and adoption is self-reported by the companies. These figures should therefore be read as a trend, and they may look very different for, say, a 15-person company.
6. People involved
Lack of skills is the main obstacle cited by those who considered AI and then gave up, accounting for 58.6% of responses. A project needs at least three roles: someone who knows the process in detail, someone with the authority to make decisions and someone who will use the tool every day. Involve the people who do the work by hand today right from the analysis phase, because they know the exceptions and shortcuts that no written procedure records.
7. Regulatory constraints
Regulatory uncertainty and privacy are two more obstacles in the Istat survey. Personal data of customers or employees, decisions that affect people and regulated industries all lengthen timelines and require checks with the DPO or legal counsel. The AI Act treats some of these uses as high risk, with heavier obligations; for disclosure obligations towards users, you will find practical examples in our article on the AI Act and transparency obligations for tech and marketing. As a first project, an internal process is simpler: one based on non-sensitive company data, where AI prepares the work and the decision stays with a person.
8. Time to a prototype
The last question is about speed: how many weeks will it take before you see the system working on your real data. A good first candidate produces a usable prototype in 2-4 weeks, the same timeframe we set for prototypes in our AI consulting projects. If you need six months of data cleanup and three integrations before getting anything, that process should be tackled later, once the team already has a successful project behind it.
A scorecard to rank your processes
Give each candidate process a score from 1 to 5 for each question, where 5 always means the situation most favourable to automation. A spreadsheet with one row per process and one column per criterion is more than enough.
- Volume and repetitiveness: 1 if the task comes up a few times a month, 5 if it recurs every day with stable rules.
- Data: 1 if the information is scattered or in people's heads, 5 if it is digital, organised and accessible.
- Integration: 1 if the management system is closed, 5 if it has ready-made APIs or imports.
- Cost of errors: 1 if a mistake reaches the customer or the accounts, 5 if it can be fixed in a few seconds.
- KPIs: 1 if you measure nothing today, 5 if you already have baseline times and volumes.
- People: 1 if there is no owner, 5 if you have a process lead and users willing to try.
- Regulatory constraints: 1 for sensitive data or decisions about people, 5 for internal, non-personal data.
- Time to prototype: 1 if months of preparation are needed, 5 if you can test within a month.
The maximum score is 40. As a rule of thumb, above 30 you have a strong candidate, between 20 and 30 a process that needs some groundwork (usually on data or integration), and below 20 a project to postpone. If two processes end up close, look at the scores for data and for the consequences of errors, which carry more weight than the others in a first project.
In our example, email triage scores high on volume, severity of errors and time to prototype. Quotes may be worth more financially, but lose points precisely on data and errors. It makes sense to start with the first and move on to the second with a few months of experience and a cleaner history.
From score to first prototype
Once you have chosen the process, write a one-page brief before talking about models or vendors.
- Describe the process as it works today, with its steps, the systems involved and recurring exceptions.
- Record the baseline KPIs measured over at least three weeks.
- List the data sources and who is responsible for each.
- State who approves the agent's output and in which cases a person steps in.
- Set the date by which the prototype will be evaluated and the minimum result needed to continue.
With this brief in hand, the conversation with a vendor or with your internal team shifts to your process and your numbers. If your company lacks a technical figure who can lead the project, a Fractional CTO can fill that role part time.
If you would rather do this work together with us, you can start with DevInterface's free AI assessment. We take the processes you have identified, assess them against your data and systems, and point out the one that makes the most sense for building your first prototype.
