Before you buy an AI tool: five questions for the vendor

In 2025, 16.4% of Italian enterprises with at least 10 employees used at least one artificial intelligence technology, double the 8.2% of 2024, according to Istat's "Imprese e Ict" release of December 15, 2025. Further down the same release sits a figure its headline leaves out: among enterprises that say they use AI, 33.4% name none of the business purposes the questionnaire offers, from marketing to administration, up from 15.5% the year before. Of those, 83.3% have between 10 and 49 employees, and Istat describes the adoption as "ancora poco strutturata", still loosely structured.

Against the EU average the gap is about three and a half points: Eurostat puts AI use at 19.95% of EU enterprises with at least 10 employees in 2025, and Italy trails the EU average in every size class, by 1.9 to 2.8 points. For anyone running a small company, the third that has AI in house matters more: perhaps a subscription an employee turned on, with no process to attach it to.

The AI nobody bought is already at work

The executive summary of Verizon's 2026 Data Breach Investigations Report, from May 2026 on data from October 2024 to November 2025, finds that 45% of employees are regular users of AI on corporate devices, authorized or not, up from 15% the year before. Verizon's press release sums it up as shadow AI use "tripled to 45%", but in the report that 45% covers all regular use. Shadow AI proper, meaning tools the company has not approved, Verizon measures with a different number: 67% of users access AI services with non-corporate accounts on their corporate devices, a share that fell slightly from the previous year.

In the data from software that flags confidential data about to leave the company, DLP (data loss prevention) systems, shadow AI became the third most common non-malicious insider action in 2025, a fourfold increase in share in a single year. Among the data sent to external models, the most common type was source code, followed by images and other structured data.

What it costs is measured by IBM's Cost of a Data Breach Report 2025, conducted by the Ponemon Institute on 600 organizations worldwide that suffered a data breach between March 2024 and February 2025. One in five reported a breach due to shadow AI, and where unauthorized use was widespread the breach cost 670,000 dollars more on average. Of the breached organizations, 63% either have no AI governance policy or are still writing one.

When access goes through a personal account, the company keeps no trace of who asked a model what, with which data and with what outcome. The trace comes from routing every request to the models through an AI gateway, a single control point that logs them, manages access keys and applies filters, and from planning a human review (human-in-the-loop) wherever the output reaches a customer or touches the books.

How long does an AI app subscription last?

RevenueCat's State of Subscription Apps 2026, published in March 2026, covers more than 115,000 apps that run their subscriptions on its platform, for over 16 billion dollars in revenue and mostly with 2025 data. These are mostly consumer apps on iOS, Android and the web, and they measure a move you know well: trying a new tool and deciding whether to keep it.

Apps with AI as their core feature start out ahead. They turn 20% more downloads into paying customers and earn 41% more per customer in the first year. Then retention, the share of subscribers who stay, gives way: RevenueCat finds these apps lose subscribers 30% faster. After 12 months, 21.1% of annual subscriptions are still active, against 30.7% for apps without AI. The report specifies that it counts subscriptions one by one, and not the people who pay for them.

The same report, with Appfigures data, counts the new subscription apps launched each month: about 2,000 in January 2022, more than 14,700 in January 2026, over seven times as many. For a buyer, the tool you are evaluating is one of thousands, and the cost left out of the quote is adoption: training people, changing the process, connecting the data.

For business software the closest source is a Bessemer Venture Partners playbook from May 13, 2024, qualitative and built on case studies from inside and outside the fund's portfolio, which acknowledges there is not yet sufficient longitudinal data to set B2B AI retention benchmarks. Its observation: AI apps for companies retain users better when they "meet users where they are", inside the platforms they use every day, and usage tends to drop whenever users face unnecessary friction, such as another sign-in or one more window to open. The example closest to an SME is DataSnipper, which works inside Excel as a plug-in. A point solution, they write, can sometimes struggle to become embedded in existing workflows.

What is already arriving in the software you pay for?

Gartner forecasts that by the end of 2026, 40% of enterprise applications will integrate task-specific AI agents, components that carry out a task on their own, up from less than 5% in 2025. It is a forecast from August 2025, which in one passage of the same release becomes "up to 40%". If it holds even in part, the first job before evaluating a new app is an inventory of the AI features already arriving in the management software, the email and the office suite you pay for, and of who switched them on.

On the AI startups founded in the last three years I hold a blunt view: in my judgment almost all of them are wrappers around ChatGPT or Claude, an interface and a few prompts on top of someone else's model, sold as a product. For a buyer the problem is twofold. You pay a markup on a capability you often already have, or that the model maker and your software vendor are building in on their own. And the tool's survival depends on a decision its seller does not make: one release from the lab upstream can make it redundant. The exceptions exist, and they are vendors that bring proprietary data, an industry process or deep integration, something the model alone does not do. The burden of proving it sits with the vendor.

Five questions before you sign

One question comes before the others, and Istat has already asked it: what is it for. A third of the enterprises using AI named none of the purposes offered. If you have an answer, these are the five questions I put to every vendor of an AI tool.

1. Where does it live? Inside the management software, the email, Excel or the office suite your staff already open every morning, or in a new window with a new login? In the second case, ask what it adds to the model it is built on: if it is a thin AI wrapper, just an interface over someone else's model, the same feature may reach you inside software you already pay for.
2. Who sees the logs? Logs are the records of every request and every answer, with the name of whoever made them. Ask whether they exist, whether you can export them, who reads them on the vendor's side and how long they are kept.
3. Does the data leave, and does it train models? Ask where it is processed, how long it stays and whether the contract rules out in writing its use to train the vendor's models or anyone else's. Ask for the clause, not a confirmation by email.
4. How is it switched off, and who can do it? You need a named person who can revoke access within minutes, and a written procedure to get your data back if you change vendors. What that exit costs is the measure of vendor lock-in, and it is worth knowing before you sign.
5. Who answers for the output? If the tool writes to a customer, drafts a quote or touches the accounts, a person in the company reviews the result and answers for it. If the vendor names itself, ask to see that in the contract.

If two or three answers stay vague, look at the decision from the other side, the build vs buy side. In the cases I see, the need is often better covered by a small integration into the system the company already has, built by someone who works on its real data and answers for the result, the way a Forward Deployed Engineer does.

In short

Before buying a new AI tool, take inventory: the AI features arriving in the software you already pay for, and the personal accounts your employees use today. Then take the five questions to the vendor, and sign only if you can say what the tool is for and who in the company answers for it.