June 30, 2026·18 min read·

Massimo Mentrasti, Founder & CEO

Founder & CEO

Massimo Mentrasti

With three decades of experience working for international corporations, of which last two playing managerial and entrepreneurial roles, he has gained an unparalleled experience selling and delivering large transformation program across different industries and functional areas (e-commerce for retail, digital marketing for automotive, CRM and applications for banking, ground control system for aerospace industry). As Accenture partner during last 15 years, he has built from scratch and led the growth of practice focused on highly innovative offering as digital transformation, analytics, digital marketing and automation.

massimo@aevoluta.com

Artificial intelligence is the most transformative technology of the last twenty years. It writes complex code, analyses thousands of documents in seconds, sustains multi-step reasoning, and produces in half an hour market analyses that would have taken an expert consultant a week. For sheer pace, the capability leap of the last three years has no precedent in any other technology.

But we live inside a deafening hype. Gartner estimates that by 2026 global AI spending will reach $2.59 trillion, up 47% year on year.¹ In the last quarter tracked, the term "AI" appeared in more than 65% of S&P 500 earnings calls, peaking at 94% in the technology sector.²

Everyone is talking about agents, models, and artificial intelligence, and the effect is an epoch-making wave of Fear of Missing Out.

It is no coincidence that those feeding the hype most insistently are those with the greatest financial interest in doing so. On 1 June 2026, Anthropic confidentially filed for IPO after a funding round that valued it at $965 billion³; a week later, OpenAI did the same, valued at $852 billion.⁴ The closer the listing day, the greater the incentive for these companies' leaders to keep the narrative about AI's unlimited potential at a high pitch. Not by chance has even the term 'p(doom)' — the probability that AI causes a catastrophe for humanity — entered common parlance: it was once an insider joke, today it is a conference question.

Meanwhile, outside boardrooms, almost everyone is already using AI in their own way. 44% of American adults have used ChatGPT at least once (34% a year ago), and about a quarter do so every day.⁵ At work the phenomenon is equally widespread: 78% of those who use AI at the office bring personal tools, often without telling anyone.⁶

But this informal, individual use — what analysts call shadow AI — almost never translates into organisational capability. Companies in practice remain often stuck at the pilot project, and the hype increasingly contrasts with concrete results. The reason, as we will see, is the same as every digital transformation of the last thirty years.

The same model, the same reasoning capabilities, are today available to everyone: the market leader and the last entrant, the multinational and the micro-enterprise. And if the technology is identical for all — if the market analysis reaching the board of a trillion-dollar company resembles the one that lands on the desk of a small business — then the difference must be sought elsewhere: in people, in proprietary data, and in the operating model that holds it all together. The question of technological sovereignty remains in the background for now, but could further complicate the picture.

The fable is seductive: take an existing process, add an agent, and it will become faster, cheaper, more scalable. It works because it stays within the corporate comfort zone: buy the technology, ask IT to implement it, measure the benefits — without tackling complex organisational change. But in the vast majority of cases, that fable is false.

Gartner predicts that more than 40% of agentic AI projects will be cancelled by 2027, due to out-of-control costs or unclear business value.⁷ McKinsey, for its part, finds that only 1% of companies consider themselves "mature" in AI adoption, while 92% intend to increase investment over the next three years.⁸

There is a paradox behind these numbers. ChatGPT, Claude, and Gemini were designed to be used by one person, on a browser, for one task — and that is exactly how most employees already use them, regardless of what the company has approved. This is shadow AI: informal, individual, ungoverned use of AI tools, outside any process or control. The result is disorienting: the company as a whole records very low adoption, while desk by desk, AI is already everywhere. But that use never becomes collective capability: each person reinvents the same prompt, no lesson reaches a colleague, and the most sensitive data ends up — one conversation at a time — on servers no one has validated.

Shadow AI is not proof that a company is adopting AI. It is often proof that it is not.

In Italy the gap is already visible: in 2025, 71% of large enterprises launched at least one AI project, against just 8% of small and medium-sized enterprises.¹⁶ This is not a question of access to technology, but of who, inside the company, knows what to do with it.

A recent signal arrives precisely as I write: according to the Financial Times, Amazon, Walmart, Uber, Cisco, and Meta are introducing token spending caps and stricter rules on when an agent should actually be used. Uber set a limit of $1,500 per employee per month on each AI coding tool, after burning through its entire 2026 AI budget in four months.¹⁵

This is not a rethink on AI — it is counter-evidence for the fable above. Investment is flowing into AI, in most cases, without impact on economic results. The reason is simple: a sub-optimal process, once automated, stays sub-optimal — only faster and harder to correct because now "the machine does it". The same applies to KPIs: measuring a call centre on calls closed, when an agent closes them quickly without solving them, produces a worse service, only more rapidly. And to governance: if no one decides what AI can do without supervision, the technology vendor will decide, by default — an implicit, often unknowing delegation.

Organisations that achieve concrete results do the opposite: they train their people and redesign the operating model (organisation, processes and governance, talent) before building agents.

The cost of compute has collapsed. On 27 January 2025, Chinese lab DeepSeek published R1, a model trained on an estimated budget of around $6 million — a fraction of what American labs claim to spend — with performance competitive with the best Western models. In a single trading session, Nvidia lost $589 billion in market capitalisation: the largest single-day value destruction in financial market history.¹¹

The cost per token has collapsed even faster. According to the Stanford HAI AI Index, achieving performance equivalent to GPT-3.5 fell from $20 to 7 cents per million tokens in eighteen months: more than 280 times cheaper.¹⁰

Consumption has exploded. Google reports processing more than 3.2 quadrillion tokens per month — more than 300 times the volume of two years ago.¹³

The market has multiplied. From $196.6 billion in 2023 to a forecast of $1.8 trillion by 2030.¹⁴

The market did not need years to notice that the computational advantage was narrowing: it took a weekend. No industry has ever eroded its initial advantage so quickly. What a quarter ago was a rare and costly differentiator is now within reach of anyone with a credit card.

This leads to an uncomfortable conclusion: if intelligence becomes like electricity or broadband, it is a real advantage only for those inside the company who have the people capable of using it better and faster than competitors.

Imagine two analysts using the same model to prepare the same client proposal. The first asks the model to write the document, skims it once for a summary check, and sends it. The rest of the time they dedicate to something else, hopefully related to their professional duties. In little time they have produced something that "looks" acceptable.

The second uses the same tool for a first draft in fifteen minutes, then invests the remaining hours verifying every number, asking the model to dismantle its own argument, checking what competitors are doing, and adding the information only they possess: that the client, three weeks earlier, on the phone, confided the real concern behind that brief.

Two proposals, the same model. The first is indistinguishable from what any competitor with the same access would have produced. The second is not, and the difference lies not in the model, but in the person who used it.

It is for this reason that an organisation lacking critical capabilities, when confronted with powerful tools, does not become more productive: it becomes more rapidly dependent on automation. The contribution of whoever settles for the first answer tends to zero, because a first answer an agent produces better, faster, and at near-zero cost. A capable organisation, by contrast, does not just become more productive: it starts to think and decide at a speed that would until recently have seemed out of reach. And in a world where everyone's agents tend to do the same things in the same way, it is precisely that speed of thought that remains the only truly scarce thing: judgement, curiosity, the tacit knowledge of the client that no model has ever read anywhere. The competitive factor has stopped being the technology: it has shifted inside the people who decide what to do with it.

A language model, however powerful, is trained on general world knowledge: it can write a standard contract, but knows nothing about ten years of complaints recorded on a single plant, the reasons a certain client truly gave up on renewal, or how your warehouse behaves the Friday before a long weekend.

This knowledge already exists, almost always, inside the company: in the CRM, in support tickets, in production logs, in commercial emails, in years of quotes. But almost never in a form an agent can use. McKinsey, in a report from May 2026, calls it the "privileged data" moat: an advantage born not from the model you buy, but from the data that feeds it — because they are cumulative and no competitor can buy or replicate them.⁹

Two companies can buy the same model, from the same vendor, at the same price, and get radically different results: not because the agent "understands" one better than the other, but because one has spent years structuring, cleaning, and connecting its data, and the other has not. What remains different, company by company, is what those models know about your company in particular. And that knowledge, unlike the model, is not for sale anywhere.

Organisation. If no one has explicit responsibility for guiding AI adoption, that responsibility does not remain vacant: it fragments into hundreds of individual choices — one per employee. Clear ownership is needed, close to the business and not just IT, capable of turning into shared practice what today remains the private initiative of those who have already started using these tools on their own. That is exactly the void in which shadow AI proliferates.

Processes & governance. You do not automate a single step of the existing process: you redesign it end to end, starting from the result you want, not from the activity you want to speed up. Rules must be defined — what an agent can do alone, what requires human oversight — and KPIs redesigned: old metrics reward faster production of the wrong output.

Talent. Continuous training is not a course to take once: it is permanent infrastructure, on a par with a network or an ERP. Those who cannot critically interrogate an AI system will have their limitations amplified; those who use it well will have their strengths amplified. Of the three dimensions, this is the most important: a poorly designed process or an incomplete rule can be corrected in a few weeks, but critical competence takes years to build. That is why the companies that will win this phase will not be those with the most sophisticated agent, but those with the most capable people to use it.

First the operating model as a whole: organisation, processes and governance, talent. Only at the end, once it is clear what needs to happen and who is responsible, does it make sense to design the technical architecture of the agents: how many, with what autonomy, on what data, with what supervision.

Those who invert this order are not getting the implementation details wrong. They are choosing, without knowing it, to automate their organisational flaws faster — and will discover this, with some surprise, in the numbers of the following quarter.

This is precisely the work of Aevoluta. With our proprietary Aevolution™ framework, we guide companies through AI transformation in a structured and risk-controlled way: first the design of the AI-enabled operating model (organisation, processes and governance, talent, and proprietary data valorisation), then the implementation of the agents that support it. Because GDPR and AI Act compliance cannot be a last-minute check, we integrate it from the outset: compliance by design, at every step.

The first digital era rewarded those who owned the information. The second rewarded those who owned the platforms. The era of abundant intelligence will not reward those who spend most on tokens and least on people. The winners will be the AI-fluent companies: those whose people know how to use AI with mastery and critical spirit, valorise their own data, and boldly rethink their operating model.

Applied intelligence to evolve human.