In early 2025, a boardroom of an Italian SME in the Northeast. Four people: CEO, CFO, IT head, marketing. The agenda item is "define the company's AI strategy". On the table is a stack of printed newsletters, a consulting firm's 80-page report, two quotes from different vendors, and the weekend edition of Sole 24 Ore with a full-page headline on generative AI.
The meeting lasts three hours. They come out with a decision: "evaluate over the next six months". In June 2025 a plan is presented. In September 2025 it's postponed. In January 2026 it's resurrected. In May 2026 a budget is approved. In September 2026 — now — implementation will begin.
Seventeen months.
Yes, the company is a real case. No, we can't name it. But it's exactly the average Italian company — among the SMEs that did something on AI in 2024-2026.
This piece is a long-form review. Eighteen months of generative AI applied to Italian companies, distilled into what we've seen work, what turned out to be hidden cost, and where it's worth investing in the next 12 months. It's deliberately long. It's deliberately frank. It's deliberately written to be re-readable at Christmas — because many things said here might already be outdated, and those that won't be will be our shared compass for 2027.
Three phases, one trajectory
In 18 months, AI in business went through three clear phases.
Phase 1 — The promise (January 2024 — June 2025). Everyone wants to do AI. Nobody knows exactly what to do with it. Budgets inflate, quotes proliferate, "ChatGPT for managers" courses multiply. AI strategy is a statement of intent: "we want to be AI-first". Underneath, little concrete.
Phase 2 — The disappointment (July 2025 — December 2025). First pilots go badly or go less well than expected. Companies discover that integrating an LLM into their operations is more complex than the demo suggested. CMOs who pushed AI projects come under pressure from the CFO. Some initiatives get frozen, others reduced. "AI deception" becomes common talk.
Phase 3 — Discipline (January 2026 — today). Companies that survived phase 2 stopped asking "should we do AI?" and started asking "for which specific problem is AI the best solution?". Pilots become smaller, more targeted, more measurable. The vocabulary refines: generative AI, agentic AI, RAG, context engineering, multimodal are distinguished. Decisions are made with discipline.
The three phases are not over. They coexist. Every new company entering the AI market lives the three phases in compressed scale: from promise to discipline in a few months if lucky, in two years if less attentive. But the trajectory is clear, and it's that one.
What really worked (in Italian SMEs)
We give you the real cases, anonymized but recognizable as patterns. They are the AI applications where we saw verified ROI in Italian SMEs.
1. AI-assisted (not replaced) customer service. For companies with recurring weekly ticket volumes (services, e-commerce, software), an AI assistant that prepares human responses (doesn't send them directly) reduces average handling time by 40-60%. It works because it preserves human responsibility, but eliminates repetitive typing. Investment: medium. Time to return: 3-6 months.
2. Internal data research and analysis. Companies with lots of internal documentation (legal, consulting, manufacturers with technical manuals) found great value in RAG systems (see the hardword piece on RAG) that allow querying the archive in natural language. It works because it solves a real problem — "I have a specific question and the answer is in one of our 4,000 documents". Investment: medium. Time to return: 6-9 months.
3. Creative variant generation. Marketing teams managing many campaigns used AI to generate copy and creative ad variants, accelerating testing. Doesn't replace the creative. Amplifies them. Works very well for ambient marketing, performance marketing, social ads. Investment: low. Time to return: 1-3 months.
4. Accelerated software development (with rules). As written in the July 23 vibe coding piece, internal development teams got +50-80% release speed. With governance. Investment: medium-high (in training and processes). Time to return: 6 months.
5. Complex document audit. Professional firms and consultancies are using multimodal AI for first passes on balance sheets, contracts, technical sheets. Human review remains indispensable, but initial "decoding" time collapses. Works very well for accounting firms, mid-sized law firms, tax consultants. Investment: low. Time to return: 3 months.
What was hidden cost
Now the part less willingly told. Three things that, in many SMEs, turned out to be expensive without anyone noticing first.
1. Time spent "evaluating vendors". Italian companies above a certain size spent dozens of half-days of the CFO, IT head, marketing, in meetings with AI vendors in 2024-2025. The vast majority of these meetings produced no decisions. Submerged opportunity cost: easily 30-50 thousand euros of managerial time per average company. On a thousand Italian SMEs, we're talking about 30-50 million euros a year of managerial time burned on evaluations. Few notice because it's not in any cost center.
2. Overlapping "tool stacks". Many companies today pay simultaneously: ChatGPT Enterprise licenses, Claude Team licenses, Copilot licenses for Microsoft 365, and one or two vertical tools (for chatbots, for analysis, for coding). Aggregate cost: often 30-60 thousand euros a year for a 50-employee company. Effective usage: ridiculously low. Rationalizing the tool stack is one of the most profitable exercises of Q4 2026.
3. Late-managed compliance. As written in the June 25 piece on AI Act phase 2, many companies discovered after putting AI systems into production that they had to document, evaluate, and supervise. The cost of doing it after is triple compared to doing it before. Italian SMEs that structured AI governance practices already in 2025 are saving money and time in 2026.
The compass for 2027
Let's come to the prediction part. Four trends we see consolidating, and which you should budget for in the next 12 months.
Trend 1 — AI becomes "vertical", not "horizontal"
For the first two years of generative AI in business, the dominant tools were horizontal: ChatGPT, Claude, Gemini, Copilot. Tools that do many things mediocrely well.
In 2027 value shifts to vertical tools. AI systems built specifically for: accounting, legal management, manufacturing, healthcare, consulting, education. They work much better for the specific case — because they're designed around it — and in return they only work for that.
What to do: evaluate what you really need, and look for the vertical vendor that solves it. Stop thinking "ChatGPT is the universal base".
Trend 2 — AI agents start replacing workflows (not people)
Agentic AI — systems that execute chains of actions autonomously — is moving from promise to operability in the next 12 months. But the way it replaces people is different from what everyone thought.
It doesn't replace customer service. It replaces the customer service workflow — the three different tools the operator uses to handle a case. The operator remains, but their work environment simplifies radically.
What to do: identify the most friction-filled workflows in your company (those requiring passages between many tools, many hands, many waits). Agents will enter there. Not elsewhere.
Trend 3 — "AI-native" become a new segment
In 2027 we'll see companies emerge that are AI-native from birth: organizations built with the assumption that AI is a constitutive part of the process, not an addition. It's what we're trying to do with OpificioAI, Tuken, BP, BI, and Blogger. We don't do it for fashion — we do it because we understood, in these 18 months, that AI-native companies operate in a structurally different way. They have smaller teams, faster decisions, products that scale with almost zero marginal cost.
What to do: if you're planning a new business unit, evaluate building it AI-native from the start. On existing products it's difficult to retrofit. On new ones it's the right move.
Trend 4 — Trust becomes the real asset
In 2026 we saw a strange phenomenon: some technically worse AI vendors won contracts against technically better vendors. Why? Because they knew how to build trust. Data transparency, cost clarity, ready compliance documentation, open audit processes.
In 2027 this trend will amplify. Companies will increasingly buy AI from vendors who can demonstrate, not just declare, to be reliable. Trust becomes part of the product.
What to do: if you're an AI vendor, invest in transparency and documentation now. If you're a customer, choose vendors based on transparency, not just performance.
Five operational tips for the next quarter
To close, five concrete actions for Q4 2026. All doable in 30-90 days.
-
AI tool stack audit. How many tools do you pay for? How used are they? What can you consolidate? Typically: -30% annual costs, +20% clarity.
-
Audit of AI systems in production. How many are in "high risk" tier per the AI Act? For each, is there a compliance plan? Involve legal. It's done in half a day, avoids millions of problems.
-
Define the product metric, not the agent. For each AI system in the company, what's the business metric telling you if it's working? If you don't have it, build it. A clear metric beats a thousand engineering dashboards.
-
Invest in governance before the next tool. House rules (see the May 28 piece on the 4 recurring mistakes) are worth more than a new model, a new agent, a new license. Spend there.
-
Start a single vertical pilot, not a horizontal one. See trend 1. Identify the most specific problem in your company where AI can help, and put three months on it. More focus, more ROI, more learning.
The final note for those reading us
This is the last piece of the May-August 2026 editorial plan. Twenty-six articles, written two a week, without skipping any. A promise we made you in May when we reopened the blog after seven months of silence, and that we kept.
In September we'll be back with a new editorial block. Topics, cadence, tone — we'll define them in the coming weeks. If you have suggestions, write to us. This blog isn't a monologue, it's a conversation.
Happy Ferragosto. Good return. See you in September, fresh, with some more ideas than now. And maybe — with a little discipline and a good dose of curiosity — one more piece of AI that really works in our SMEs.
Do you feel halfway between the disappointment phase and the discipline phase? Under artificial intelligence for business and with OpificioAI we help Italian SMEs choose and put into production the AI that makes a difference. Let's talk in September, when you're really ready.