Between announced failures and the race to adopt, how to really make AI work in the enterprise

The most interesting data, however, for understanding real adoption is the way in which companies use the technology. The Artificial Intelligence Observatory of the Politecnico di Milano reports that in Italy 81% of large companies have at least begun to evaluate or activate AI projects, but we remain below the European average (89%).

Our country remains last among those analysed in terms of the share of companies with at least one active AI project (59%), but among those that have started, a substantial proportion is already beyond proof of concept: a quarter of large companies declare widespread AI projects, more than Germany, France and the United Kingdom.

On GenAI 'off-the-shelf', the country appears even faster than others: more than half of large companies say they have purchased licences. And among those using them, 39% say they have experienced an increase in productivity.

The temptation is to believe that simply 'putting a tool in people's hands' will get results. But this is precisely where the paradox lurks: consumer AI and desktop GenAI can deliver immediate benefits, while projects that should really change processes require much more: adequate data, organisational integration, governance, metrics and skills.

The Polytechnic Observatory highlights the Italian split: in SMEs, adoption is 'far behind' compared to the large ones. 58% say they are at least interested in the subject, but there are few real projects: 7% of small and 15% of medium-sized enterprises declare AI projects started, the use of generative Ai stops at a modest 8%.

When AI enters these companies, it often does so for defensive and concrete reasons: operational efficiency, optimisation of production processes, waste reduction, predictive maintenance. But the most cited limitation is structural: immaturity in data management, i.e. the precondition without which AI adoption remains crippled. In many smaller realities, data exist, but they are fragmented, sporadic, ungoverned, and, what is more, there is a lack of technology and in-house expertise to make a process 'trainable'.

For smaller companies, it is not just a budget issue, it is an issue of priorities, organisation and risk awareness.

If MIT insists on the difficulty of the transition between pilot and production, Italian analyses help to understand which obstacles weigh heaviest on real companies. The first is economic and organisational: estimating the costs of a GenAI solution is not trivial. The Politecnico emphasises how 'consumption-based' pricing models make it difficult to predict the overall cost ex-ante. It is not surprising, therefore, that more than one in two companies consider managing GenAI costs to be complex.

The second aspect is regulatory and reputational: the path to compliance is 'still long'. More than half of the companies active on AI say they are not fully clear on the regulatory framework of the AI Act, and less than a third have taken concrete steps on application ethics.

The third - and often decisive - obstacle is the governance of everyday use: AI enters anyway, even when the company is not ready. "Shadow AI' is the key word: people using unapproved tools to work faster. Large Italian companies are showing a reaction: in more than four out of ten companies, guidelines and rules of use have been published, and in 17% of cases the use of unapproved tools has been banned.

In essence, adoption without defined governance can become a risk multiplier, but excessive caution can freeze innovation. What is needed, therefore, is a middle way: clear rules, approved tools, widespread training, selected and measurable use cases.

In the 'evolving factory', AI is increasingly perceived as a competitive lever. The Mecspe Observatory photographs a manufacturing sector that declares a good level of digital maturity (almost six out of ten companies), but which identifies skills as the bottleneck. Among the digital skills priorities reported by companies: robotics and intelligent automation, then programming and development of AI and machine learning, then Big Data analysis and management.

In short, there is no shortage of technology, no shortage of vendors, even incentives (when they exist) drive investment. But without people capable of designing and making change work, AI risks remaining in proof-of-concept limbo.

At the strategic level, the first distinction is having an operational objective and metrics to measure it. The Observatory notes that only 8% of companies have established metrics to measure the impact of digital: without objectives, AI remains 'cool tech'. In any case, it requires an integrated approach that challenges flows, activities and responsibilities, in essence the entire business organisation, as MIT also points out. If not, the technology runs the risk of going off the rails: successful projects in Italy arise precisely from choices of process re engineering and application integration.

But the real crux is data, as already pointed out. Almost all SMEs and many large companies stumble here: scattered, uncleaned, non-versioned data: data strategy, catalogues and industrial pipelines are prerequisites to scale, along with ethics/AI Act rules and policies against shadow AI.

On these fronts, the large ones have critical mass and can invest in data platforms, governance and process engineering: that is why the majority have active projects and a quarter widespread initiatives. But they lag behind the European average: bridging the gap means standardising adoption paths and industrialising value use cases.

For SMEs, the path is more complex, held back by resources and data: the most pragmatic way is to start from repetitive processes, use ready-to-use tools and measure micro-benefits immediately. The 5.0 incentives have helped, but now the enabler is training and the ecosystem that accompanies from idea to production. From this point of view, something is moving, starting with the entrepreneurial system, which is well aware that this time missing the train means risking remaining on the sidelines of the global economy.