Why so many AI projects never arrive
Adoption is growing fast. In Italy, 16.4% of businesses with at least 10 employees used AI in 2025, twice as many as the year before (Istat, December 2025). Getting results is another matter. According to MIT research published in 2025, around 95% of the generative AI pilots it studied had no measurable impact on the bottom line (Fortune, August 2025).
The technology is rarely the problem: the models are mature and the demos work. What stops projects is vague goals, missing data and no plan for getting the result into the hands of the people doing the work. The same research notes that the highest returns come from internal processes such as admin and operations, rather than from the most visible projects.
1. Start from a problem, not from AI
“Let's use AI” is not a goal. “Answer customer requests within an hour instead of a day” is. A good starting point fits in one sentence and comes with a number you can measure today: time, errors, lost requests, hours spent.
2. Pick a small first project
The first project has to prove its value quickly. The best candidates happen often, take up someone's time today and have a result that is easy to check: triaging incoming email, extracting data from documents, drafting replies. If the task is always the same, you may not need AI at all: AI Agent or Automation? explains how to tell the two apart.
3. Check the data before you build
Among Italian businesses that considered AI but didn't adopt it, 45.2% cite poor availability or quality of data as an obstacle (Istat, 2025). Before writing any code, answer three questions:
- Does the information you need already exist? Where is it?
- Can a system reach it, or does it live in inboxes and scattered spreadsheets?
- Is there personal or confidential data that needs special care?
If the data isn't there, the first step is organising it. Often an automation that gathers it in the right place is enough, as we explain in What to Automate First.
4. Get to real users early
A demo has to work once, in front of a friendly audience. A system in production has to work every day, on cases nobody anticipated. That is why the first version should be in the hands of the people who will really use it within a few weeks, with a narrow scope and someone checking the results.
At this stage AI works alongside people: it proposes, a person approves. Once the results are stable, you decide what it can do on its own.
5. Measure and decide
The number chosen in step one is the yardstick. After a month, compare it with where you started. If it improved, extend the system to more cases or processes. If it didn't, find out why and fix it, or stop the project before spending more: stopping early is as good an outcome as carrying on.
Who looks after it after launch
Every AI system needs someone to look after it: reading the cases it gets wrong, updating the instructions and documents it draws on, and making sure costs stay under control. It can be someone on your team or the partner who built it, but it has to be decided before you start.
Signs the project is going off track
- After a month, nobody can say whether it is working.
- The scope keeps growing before the first part is in use.
- Users go back to the old way whenever they can.
- The prototype is always “nearly ready”.
If you want to set up your first project the right way, get in touch: we start from the problem and tell you honestly whether AI is the answer.
Frequently asked questions
How much does a first AI project cost?
It depends on the process and the systems involved. A well-defined first project costs far less than a broad programme, and it is the safest way to check the return before investing more.
Do we need in-house technical skills?
No. You need someone who knows the process well and has time to try it. The technical side can be handled by a partner.
Is our data safe?
It depends on how the system is built. You choose providers that don't use your data to train their models, limit access to what is needed, and keep sensitive data out of the flow when it isn't required.
