The first question I ask a leader who wants to get into AI is not which tool. It is which data, and where it goes.
Because an AI plugged into your content means your contracts, your prices, your customers, your methods pass through somewhere. Knowing where is the basics. And many never ask the question, until the day it comes back at them in full force.
When a whole country unplugs an AI tool
In March 2023, the Italian data protection authority, the Garante, simply blocked ChatGPT on Italian territory, until guarantees were provided on the handling of personal data. A G7 country unplugging the most popular AI tool in the world is not a detail. It is the signal that the data question is not a lawyer’s whim, but a real risk, capable of stopping a use you thought was settled.
For a company, the lesson is clear. What can make a national authority back away can also, at your scale, expose you to a complaint, a leak, or simply the loss of trust from a customer who discovers where their information ended up.
The three points to lock down upfront
Three decisions deserve to be taken before the first line of code, not after.
The first point is where the data is processed. A consumer tool may, depending on its terms, keep and reuse what you send it. For sensitive information, that is not acceptable. You need to know how to read those terms, or choose a solution that keeps your data with you. The difference between a public tool and a private AI plays out exactly there.
The second point is who has access to what. An AI that answers from your documents must not reveal the payslips to the first salesperson who asks, nor the confidential strategy to an intern. The access rights that exist in your company must be found again in the tool. Yet this is often forgotten: you plug an AI into everything, and you discover it answers everyone about everything. A well built AI respects the partitions you have already set.
The third point is traceability. In case of doubt, you must be able to know what was asked, by whom, and on which sources the answer was built. Without that trace, you steer blind, and you will be able neither to correct an error nor to answer an awkward question.
This is not technical, it is organisational
Here is what should reassure a leader who does not feel technical. None of these three decisions requires understanding how a model works. They are choices of organisation and common sense. Where do my sensitive items of information live? Who can see them? How do I keep a trace? A leader can answer that better than any engineer, because they are the one who knows the real value of each piece of information.
The honest nuance is that these decisions are taken before, not after. Once the data is circulating, catching it back costs dearly, and sometimes it is too late, information that has left does not come back. This is exactly the logic I follow when I design a private AI for a company: you start by mapping where the information lives and who can see it. The tool comes after. In the right order, AI becomes a safe asset. In disorder, it is a leak that ignores itself, until a Garante, a customer or a competitor reminds you it existed.