How to Give AI Access to Company Data: Start With the Sensitive Data

Matthias Heim · 2026-11-16

Find the sensitive data first, set one rule per data class, then open up the rest one workflow at a time. Four steps, and what the Swiss data protection law means for them.

To give AI access to company data safely, start with the sensitive data, not with a big cleanup. Find where personal and confidential data sits, set one rule per data class, separate it, and then open up the rest one workflow at a time.

Access is where the gains are. An AI that can read your files and use your tools can do the work itself. A chat window waits for someone to paste text in.

This article covers the four steps, why messy files are rarely the problem, and what the Swiss data protection law means for each step.

In short

In our projects, messy folders were rarely the blocker. Sensitive data mixed into everything was.

Work in four steps: find the sensitive data, set rules by data class, separate it, open up the rest.

Under the Swiss Federal Act on Data Protection, FADP (nDSG), check where data is processed and sign a processing agreement with every AI provider before you connect real data.

The Four Steps

In most of my AI projects, the first real discussion is not about models. It is about whether the AI may see the data. This order gets you past that discussion:

Find it. List where customer, personnel and other sensitive data actually sits, including the copies nobody planned.

Set rules by data class, not by tool. Public, internal and sensitive is enough to start. Each class gets one rule for which AI services may process it.

Separate it. Move sensitive data into folders, mailboxes or systems with their own access rights, for example an HR drive that only HR can open. The data owner in each department decides what counts as sensitive, and IT sets the access rules.

Open up the rest. Connect the AI to internal and public data, where most of the daily work lives, and start with one workflow.

Rules by data class also last longer than rules by tool. Tools change every few months. Whether a salary sheet is sensitive does not.

Messy Files Are Not the Problem

The usual objection is that the data isn't ready: folders full of PDFs, Excel copies of Excel copies, three versions of the same price list.

In our projects, current models have read those well, as long as they had access. A full cleanup before you start tends to delay the project, so we clean only what the first workflow needs. That is experience, not a rule: check it on a sample of your own files.

We looked at the spreadsheet side of this in our article on AI data readiness .

The Real Blocker: Sensitive Data Spread Through Everything

What actually stops companies is sensitive data spread through all of it. Customer details in a project folder. Salaries in a planning sheet.

Nobody knows exactly where it sits, so nobody dares to connect an AI tool to the shared drive. The AI could do the work, but nobody can approve the access.

That caution is justified. An AI agent, meaning software that works in your systems on its own, acts with the access it is given. If sensitive data sits in a folder it can read, it can read that data.

Why Access Matters More Than the Model

In most companies, every AI task starts with someone copying text into a chat window. A better model does not change that. It still waits for someone to paste.

Anthropic, which sells AI models, reports in its IPO filing with the US Securities and Exchange Commission that its own internal AI usage grew 250-fold between January and September 2026, valued at list prices. It is a vendor's figure about its own company, so read it as a direction, not as a benchmark for yours.

Companies that already have their data rules and connections in place can plug in each new model. Companies that automated a few things here and there start from scratch each time.

What Swiss Data Protection Law Means Here

The FADP has applied in its revised form since 1 September 2023. For AI access, three points matter most:

An AI provider that processes personal data for you acts as your processor. You need a contract that binds it to your instructions, and you remain responsible for the data.

Personal data may only be disclosed abroad to countries with adequate protection, or with safeguards such as standard contractual clauses. Where the model is hosted therefore matters.

If a new use of personal data is likely to pose a high risk to the people concerned, you need a data protection impact assessment before you start.

This is not legal advice, so involve your data protection officer. In practice, the data class decides the route: internal data can often go to a mainstream provider under a processing agreement, while sensitive data may need a model hosted in Switzerland or must stay in your own systems.

We describe these options on our Swiss sovereign AI page .

Frequently Asked Questions

Is our data ready for AI?
Often more than you think. In our projects, messy folders, PDFs and Excel copies have been readable for current models. The real question is whether sensitive data is mixed in, and whether you know where.
How do we give AI access to company data safely?
Find and separate sensitive data first, set rules by data class (public, internal, sensitive), check hosting and processing agreements under the FADP, then connect AI to the classes it may see, starting with one workflow.
Do we need to clean up our data before using AI?
Not everything. Clean what the first workflow needs and check a sample of your files. Separating sensitive data is the step you cannot skip.
Can we use AI with customer or employee data in Switzerland?
It can be done when these conditions are met: a processing agreement with the provider, hosting that meets the FADP rules for disclosure abroad, an impact assessment where the risk is high, and access limited to the data the task needs. Removing personal data before it reaches the AI, is often the simplest route.

How we work · Work · Insights