Personal ChatGPT is not your company’s memory

01 Two numbers that don’t add up

Put two recent surveys side by side and you get a picture that should worry anyone running a company. On the company side, Istat reports that in 2025 16.4% of Italian firms with at least ten employees used at least one AI technology. The figure doubled in a year, from 8.2%, and tripled in two. It is still a minority: 83.6% of firms adopted nothing at all, and the European average is 19.95%.

On the worker side, the Politecnico di Milano’s Artificial Intelligence Observatory finds that 47% of workers use AI tools at work, and that only 19% of those users work exclusively with tools their company provides. Eight in ten, the observatory writes, use tools the company does not provide, at least part of the time: a personal account, a free plan, a subscription paid out of their own pocket.

The two numbers describe the same offices. Nearly half the workforce has AI in its daily routine, while the overwhelming majority of employers have not adopted any. There is no contradiction: the technology entered through the front door of individual curiosity, not through a procurement process. It is not a uniquely Italian story either. Microsoft and LinkedIn measured the same thing globally back in 2024: 78% of AI users bring their own tools to work, a share that climbs to 80% in small and medium companies. In Italy, by Microsoft’s count, it was 73%.

Source: Istat, Imprese e ICT, Anno 2025 (published 15 December 2025).

The gap between large firms and SMEs is the part worth pausing on. Above a certain size, adoption becomes a project with a budget and an owner. Below it, adoption is whatever each person decided on their own last Tuesday. Istat sees it in its own data: a third of the firms that say they use AI (33.4%) cannot name a single business area it serves, and 83.3% of them have fewer than 50 employees.

The most recent figure confirms the shape of it. Among companies with at least 20 employees, the Bank of Italy counts 32% using AI tools in early 2026, experimental use included; only 5% have built them extensively into their processes. The full numbers, by size and sector, are in our research: How many Italian companies use AI in 2026.

02 The work gets done. It just doesn’t stay

None of this means people are doing anything wrong. Quite the opposite: someone in the technical office spends an afternoon getting an assistant to summarize twelve years of assembly instructions, and by evening they have something genuinely useful. The problem is what happens to that afternoon’s work.

It stays in a personal account. The prompts that took three attempts to get right, the uploaded documents, the corrections made along the way, the ten conversations that taught the tool how this company actually talks about its own products: all of it sits in a thread only one person can open. The colleague two desks away, facing the same question next month, starts from an empty box.

So the pattern repeats, once per person. Each individual gets faster. The organization does not: it has simply added a new place where knowledge accumulates without being shared, on top of the shared folders nobody opens and the PDFs nobody reads. Information was already hard to find: in a Gartner survey of digital workers in the United States, the United Kingdom, India and China, 47% said they struggle to find the information their job requires, across an average of eleven applications. Our study on the cost of knowledge that can’t be found puts a price on it for Italy. The last thing anyone needed was a twelfth silo, one per employee.

03 Context debt: who pays it, and how

Every system that shares no memory with the others hands the organization a bill: reconstructing the whole picture, from scratch, every single time. We have called this context debt, and a personal AI account is a particularly expensive form of it, because it looks like the opposite of a problem.

The interest is paid in small instalments, which is why nobody books it:

• Re-uploading. The same price list, the same technical spec, the same contract template gets uploaded again by every person who needs it, into every account.
• Answers with no traceable source. An answer arrives without saying which document it came from and from which page. Either you trust it, or you go and check by hand, and checking costs more than not asking.
• Divergent versions. Two people ask the same question of two different accounts holding two different vintages of the same document, and get two different answers. Both look equally confident.
• Invisible permissions. A document that should never have left the finance folder gets uploaded to a personal account so it can be summarized. Nobody decided this. Nobody logged it.

None of these items ever appears as a line in a budget. They surface as the feeling, common in any growing company, that everyone is working harder while things move no faster.

04 What happens the day the account is closed

There is a moment when context debt comes due all at once, and it is the same moment Italian family firms have been bracing for anyway: someone leaves.

It used to be that when a plant manager or a senior estimator retired, the company lost what was in their head. That was already the harder half of the problem. Now it also loses what was in their account: the assistant they had patiently taught, the documents they had gathered in one place, the phrasing that finally produced the right answer. Two years of quiet organization, gone with a password.

Worse, the loss is invisible on the way out. Nobody hands over a personal AI account during a handover. There is no folder to point at, no shared drive to migrate. The company does not even know what it had, which means it cannot notice that it is gone until someone asks the question that used to have an answer.

05 It is not a tools problem, and the data says so

The instinctive fix is to buy licenses: give everyone the company-sanctioned version of what they were already using privately, and consider the matter closed. Large Italian companies are already doing it: according to the Politecnico di Milano observatory, 84% have bought licenses for at least one generative AI tool, and 24% ban the tools they do not provide. It helps with data protection. It does very little for memory, because the memory problem was never about which model was answering.

Microsoft’s 2026 Work Trend Index, run on twenty thousand knowledge workers who use AI across ten markets including Italy, quantifies the point: organizational factors, culture, manager support and talent practices, account for 67% of the reported impact of AI, against 32% for individual factors such as mindset and behavior. It is a statistical association, not proof of cause, but the environment weighs roughly twice as much as the person. The same study identifies a group it calls blocked agency, 10% of users who are individually capable and organizationally stuck: people who know exactly what to do with these tools, inside companies that have not built the conditions for it to matter.

There is a further clue hiding in the Istat figures. Among Italian firms that have adopted AI, the single most common use is not writing marketing copy or generating images. It is extracting knowledge and information from text documents: 70.8%, well ahead of generative AI for text, voice and images at 59.1%. In this use Italy is level with Europe: 11.6% of Italian companies against 11.8% of European ones (Eurostat, 2025). Companies that made a deliberate choice are, overwhelmingly, using AI to get at what they already know. That is a knowledge problem being solved with a knowledge tool, and it is precisely the job a personal account cannot do on behalf of an organization.

Tools alone don’t move the numbers either. Among companies with at least 20 employees that use AI, the Bank of Italy finds that 70% have seen no effect on labor productivity so far; among those using it extensively, the share falls to 59%.

06 What a shared memory has to do

The alternative to twelve private assistants is not one bigger private assistant. It is a layer the company owns, sitting under whatever assistants people use. Four requirements, in order, and each one is what a personal account structurally cannot provide:

1. Sources stay where they are. The knowledge base connects to the folders the company already uses and follows them as they change. Nothing is migrated, nothing has to be tidied up first, and nothing lives in a copy that quietly goes stale.
2. Answers cite their source. Every answer says which document and which page it came from, and says so when the answer is not in the documents rather than inventing one. Without this, verification costs more than the answer saves.
3. The company decides who sees what. Whoever cannot open a document does not see its content inside an answer either. Visibility is decided once, by the company, not implicitly by whoever uploaded what into which account.
4. The memory outlives the person. Corrections, curated materials and configured assistants belong to the organization. When someone leaves, what they built stays, and their successor starts from it instead of from an empty box.

Note that none of the four is a feature of a language model. They are properties of an organization that decided to keep its own memory, and that is the reason the gap between individual and organizational productivity closes or does not. The 19% of Italian users who work exclusively inside company tools are not necessarily better equipped than the others. They are working inside companies that made a choice.

The choice is not urgent because AI is fashionable. It is urgent because the alternative accumulates quietly: every month without a shared memory is another month of work that gets done, gets used once, and does not stay. When the time comes to choose, our guide lists the twelve questions to ask a vendor.

Frequently asked questions

Sources

1. Istat, Imprese e ICT, Anno 2025, press release and tables, 15 December 2025.
2. Eurostat, Use of artificial intelligence in enterprises, Statistics Explained, and dataset isoc_eb_ai, 2025 data.
3. Bank of Italy, Relazione annuale sul 2025, box on artificial intelligence, 29 May 2026.
4. Osservatorio Artificial Intelligence, Politecnico di Milano, Intelligenza Artificiale in Italia: il mercato cresce del 50%, February 2026.
5. Microsoft and LinkedIn, 2024 Work Trend Index Annual Report, 8 May 2024 (31,000 knowledge workers in 31 markets), and the Microsoft Italy press release of 9 May 2024.
6. Microsoft, 2026 Work Trend Index Annual Report, 5 May 2026 (20,000 knowledge workers who use AI in 10 markets, Italy included).
7. Gartner, Survey on digital workers, press release, 10 May 2023 (4,861 employees of organizations with 100 or more employees in the US, UK, India and China).