Enterprise AI Adoption: The Number Everyone Cites, and the One Nobody Does 

47% of Italian workers already use AI tools at work, largely for individual, day-to-day tasks: drafting, summarizing, first-pass analysis. That figure, from the Artificial Intelligence Observatory at Politecnico di Milano's School of Management, gets cited constantly as proof that enterprise AI adoption has arrived. 

What gets cited less is the other half of the same finding. Structural change inside organizations hasn't kept pace. Most companies operate largely as they did before, and the more ambitious adoption models, the ones the Observatory calls "AI Factory," building AI into the core operating model of a business unit rather than bolting it onto individual tasks, remain the exception, not the norm. 

That's the real shape of enterprise AI adoption right now: high at the individual level, thin at the organizational one. 

Enterprises Adopting AI Solutions: A Divide by Size, Not by Technology 

The same research shows this isn't evenly distributed. Among enterprises adopting AI solutions, 71% of large Italian companies started at least one AI project in 2025, against just 8% of small and medium enterprises, a divide the Observatory frames less as a technology gap and more as a gap in managerial readiness and the maturity of a company's own information assets. 

Even among large enterprises already running AI projects, only one in five has meaningful AI pervasiveness across multiple business functions. Most deployments stay confined to a single team or use case. 

Adoption isn't just uneven across company size. It's shallow even where it exists. 

The Generative AI Enterprise Challenges Blocking the Shift to Organizational Adoption 

Two structural gaps explain why adoption stalls before it becomes transformation. Together, they account for most of the generative AI enterprise challenges organizations hit the moment they try to move past individual use. 

The first is governance. Just 9% of large Italian enterprises have a structured AI governance framework, clear ownership, initiatives actually checked against ethical principles and business goals. Another 54% are working toward a centralized model but haven't built it. On EU AI Act readiness, more than half have started AI literacy training, but only 15% have a structured compliance project underway. Without that structure, scaling an AI use case past a single team means scaling risk right alongside it, which is a strong incentive to keep things small. 

The second is fragmentation. Custom, bespoke AI projects still account for 77% of the Italian market, while Process Orchestration and Agentic AI solutions, the more standardized, scalable end of the spectrum, represent only 4%. A landscape built mostly out of one-off, custom deployments doesn't lend itself to the kind of repeatable, organization-wide rollout that turns adoption into transformation. Every custom build is its own island. 
 

"Agentic AI" Is the Word of the Year. Reliability Is Still the Blocker. 

Nicola Gatti, Director of the Observatory, named Agentic AI the word of 2025, not for its current economic weight but for what it unlocked conceptually. He's also direct about the current ceiling: today's systems still lack the native logical reasoning and self-correction needed to match their fluency, and until that gap closes, keeping an expert in the loop isn't a cautious add-on. It's what makes scaling past the individual level safe to do at all. 

Read against the adoption numbers, that's not a side note. It's the condition under which the 47% turns into something closer to the AI Factory model, rather than staying stuck at the micro-task level indefinitely. 

About the Research 

The data cited above comes from the Artificial Intelligence Observatory of the Politecnico di Milano School of Management, part of the wider Osservatori.net Digital Innovation research group active since 1999. The 2025/2026 edition combines a CAWI survey of 198 CIOs, Executive IT, Innovation Managers and R&D leads at large Italian enterprises, a separate survey of 500 SMEs, a three-country consumer survey (Italy, France, UK), and a market-sizing model built on direct interviews and balance-sheet analysis of over 1,000 mapped companies. Full findings are available at osservatori.net. 

For a deeper look at what a structured AI governance framework actually needs to cover, beyond the adoption numbers, see Contextual AI Governance: What Enterprise AI Needs Beyond Accuracy.  
 

A few terms this piece leans on, useful if you're skimming. 

 
Enterprise AI Adoption 
The use of AI tools and systems within a company, spanning individual, task-level use (drafting, summarizing, first-pass analysis) through to full organizational transformation, where AI is built into how a business unit actually operates. Most enterprise AI adoption today sits at the individual end of that spectrum. 

AI Factory 
A model of AI adoption, as defined by the Politecnico di Milano's AI Observatory, in which AI is embedded into the core operating model of a business unit rather than layered onto individual tasks. Still the exception among Italian enterprises, not the norm. 

Structured AI Governance Framework 
A governance setup with clear ownership, defined responsibilities, and initiatives that are actively checked against ethical principles and business goals, as opposed to a policy document that states intent without a mechanism to enforce or track it. 

Expert-in-the-Loop 
A model in which a domain expert is a structural part of how an AI system keeps its knowledge accurate over time, validating outputs, correcting gaps, and feeding those corrections back into a governed knowledge base, rather than acting as an external auditor reviewing the system after the fact. 

Governed Knowledge Base 
A knowledge base maintained with mechanisms for expert review, correction, and write-back, so that its content stays accurate and traceable as the AI systems drawing on it scale past individual use. 

Closing the Enterprise AI Adoption Gap Without Reinventing the Organization 

Here's the part usually missing from the adoption conversation: closing the gap between individual use and organizational transformation doesn't require reinventing how people work. 

Experts are already experts. What most organizations are missing isn't a redesigned workflow, or a mandate to rebuild every process around AI. It's a governed, natural-language environment where the validation experts already do, the correcting, the flagging, the "actually, that's not quite right", becomes part of the system instead of disappearing into a chat log nobody indexes. That's a smaller, more achievable version of transformation than an AI Factory overhaul, and it's available today. 

This is precisely the layer Syllotips is built for. An AI agent with no mechanism to route its gaps to the right expert, capture the correction, and write it back into a governed knowledge base isn't scaling adoption safely. It's scaling exposure. Syllotips' Expert-in-the-Loop model closes that distance: the expert isn't an external auditor checking the system after the fact, but a structural part of how the system keeps its knowledge accurate as usage grows past the first few micro-tasks. 

Where Enterprise AI Adoption Quietly Breaks: The Knowledge Base 

This is also where a lot of enterprise AI adoption efforts run into trouble without anyone noticing right away. When retrieval pulls the wrong source, the model still produces something polished and confident, with no visual cue that what it just said rests on shaky ground.  

Whether a system can flag its own uncertainty instead of guessing is usually what separates one experts keep trusting from one they quietly stop using. A disorganized, poorly maintained knowledge base doesn't just lower answer quality: it creates costs that stay invisible until months into a rollout, once documentation has drifted and adoption has already stalled at the team level nobody expected it to stop at. 

Scaling enterprise AI adoption past the individual level, in other words, isn't primarily an organizational redesign problem. It's a governability problem: integrating AI into the workflows experts already have, with answer correctness and an accurately maintained knowledge base built in from the start, rather than patched on after adoption has already plateaued.