In 2026, 68% of Italian SMEs say they want to "do something with AI" in the next 12 months. The problem: less than 20% of those who have already tried are satisfied with the results. The difference isn't in the budget. It isn't in the technology. It's in the method.
This guide isn't for those who want to do an "AI project". It's for those who have slow processes, recurring errors, training that's hard to scale, and who want to solve these issues with concrete, measurable tools.
Why most AI projects in SMEs fail
Before understanding how it's done, it helps to understand why it doesn't work. We analyzed AI projects in Italian SMEs that produced no results, and the cause is almost always one of these three:
Mistake 1 · They buy technology without a process
The company buys an AI software license, runs a generic training course, and expects things to improve on their own. The tool exists, but no one uses it systematically, because no one has redesigned the process around it.
Mistake 2 · They start with the wrong project
Enthusiasm leads to choosing the most "visible" or most technically interesting project, not the one with the greatest operational impact. Time and resources are spent on automations that touch 5% of the work, ignoring the 40% that could be optimized right away.
Mistake 3 · No KPI defined up front
Without success metrics defined before starting, it's impossible to know whether the project worked. And without that certainty, the organization doesn't adopt the tool with conviction.
"Italian SMEs face two opposing pressures: the AI hype that promises everything, and the large consulting firms that cost too much and deliver slides."
The starting point: understanding where you are (AI Readiness)
Before any implementation, a company needs to know precisely where it stands on its AI adoption journey. This is called an AI Readiness Assessment, and it isn't a formality: it's the foundation on which everything else is built.
A serious assessment answers these questions:
- Which business processes have the greatest potential for optimization with AI?
- What is the organization's current level of digital maturity?
- Does the team have the minimum skills to adopt and maintain the tools?
- What are the 3 interventions with the best impact-to-effort ratio?
The deliverable of a well-done assessment isn't an 80-page report. It's a prioritized roadmap and an action plan with timelines, owners and KPIs. Something you execute, not something you file away.
The 3-stage method: from assessment to autonomy
An AI journey that works isn't a single project, it's a progressive process with three distinct stages, each with specific, measurable deliverables.
Stage 1 · AI Readiness: diagnosis and roadmap
It always starts with a diagnosis. The goal is to identify, among all business processes, those where AI can make the most difference in the least time. It's not about evaluating the technology, it's about understanding the business.
Output: Prioritized roadmap with the 3 high-impact interventions, an action plan with timelines and owners, agreed success KPIs.
Stage 2 · Implementation: hands-on support
Implementation can't be delegated. The internal team must be a protagonist, not a spectator. Only this way does it acquire the skills needed to manage and improve the tools on its own.
In this phase we work side by side with the team: we implement, test, measure, correct. The first concrete results typically arrive between 30 and 90 days from the start.
Output: Tools working in production, a trained team, redesigned processes, baseline metrics.
Stage 3 · Scale: validation and autonomy
Once the first processes work, you scale. Proof of concepts validate the process redesign on a wider scale. Compliance-by-design is integrated. And above all: the team is autonomous, it doesn't depend on the consultant to make things work.
Output: Scaled processes, full operational autonomy of the internal team, a plan for future evolution.
The sign that an AI journey is working isn't that the company uses AI. It's that the team no longer thinks of it as something special, it has become part of the normal way of working.
How long it realistically takes
One of the most frequent questions is about timing. The honest answer:
| Phase | Traditional approach | ONP method |
|---|---|---|
| Assessment | 6–8 weeks | 2–3 weeks |
| First results | 6–12 months | 30–90 days |
| Team autonomy | 12–18 months (if ever) | 3–6 months |
| Measurable ROI | 18–24 months | 4–6 months |
The difference isn't technological, it's methodological. Starting from high-impact processes, working with hands-on support, and defining KPIs before starting dramatically compresses the timeline.
How to measure the return on investment
The ROI of an AI project isn't measured "in the value of the AI", it's measured against the specific problem you wanted to solve. Before starting any project, you have to answer these questions:
- What is the current cost of the problem? (person-hours, errors, delays, lost customers)
- By how much do we expect to reduce it? (a realistic percentage, not an optimistic one)
- In what time frame? (30 days? 90 days?)
- Who verifies the numbers? (a person responsible for tracking)
Only with these answers can you assess whether an investment makes sense, and prove that it worked. Without this structure, any AI project is an act of faith.
Want to understand where to start in your company?
An AI Readiness Assessment with ONP starts from €5,000 · with guaranteed deliverables.Where to start, concretely
The first step isn't choosing a tool. It isn't training the team. It's identifying the process with the greatest impact, the one that, if optimized, frees up the most time, reduces the most errors, or unlocks the most growth.
In almost all Italian SMEs, the most frequent candidates are:
- Order management and data entry (manufacturing, distribution)
- Document drafting and regulatory research (professional services)
- Customer service and request handling (retail, distribution)
- Staff onboarding and training (any sector)
- Reporting and data analysis (any sector with more than 50 employees)
Once the process is identified, you quantify the problem, define the KPIs, and decide whether to run a formal assessment or start directly with a proof of concept.
What matters is not waiting until you have "the perfect project". AI is learned by doing, and the companies that have already started, even with a small intervention, are building a competitive advantage that in 2027 will be hard to catch up on.