# Is AI Safe for a Business That Holds Customer Data?

*Question — 2026-10-01 — by Mahmoud Zalt*

Using AI when you hold customer data: what GDPR expects, who is responsible, what to tell clients, and the controls that keep you defensible.

**Short answer.** **Yes, if you treat the AI tool as a supplier who touches customer data, which is exactly what it is. **That means checking where the data is stored, whether it trains a model, who can read it, how deletion works, and whether your client contracts require you to disclose new suppliers. None of this is unique to AI. It is the same diligence you already do for your email provider.

When you hold other people's data, the question stops being about your own comfort. Your clients trusted you with names, addresses, health details, or financial records, and you signed something promising to look after them.

This article is a practical walkthrough, not legal advice, and if your sector has specific rules you should still speak to someone qualified. What it will do is give you the right questions, the right order, and an honest account of the parts that are genuinely unresolved.

Here is the relevant shape of **Sistava**. Your workspace is separated from every other customer at the database level, the infrastructure runs in Europe, your business data is never used to train a model, every action is logged, and you can delete data whenever you choose. Those five facts are what most of the diligence questions are really asking about.

## At a Glance

- **Europe** Where the infrastructure runs
- **No** Training on your customers' data
- **Per customer** How workspaces are separated
- **On request** Deletion, per item or whole workspace

## Does using an AI tool break GDPR?

No, using AI is not forbidden. GDPR does not ban a category of software, it puts conditions on how personal data is handled: you need a lawful reason, you should use no more data than the purpose needs, people keep their rights to access and deletion, and your suppliers must be held to the same standard by contract.

So an AI tool is treated like any other supplier who processes data on your behalf. The work is the same work you did when you chose a CRM or a helpdesk. What trips people up is not the AI part, it is that they skipped the supplier diligence because the signup was three clicks and felt like a toy.

## Who is responsible if something goes wrong, you or the vendor?

In most cases you are. You decide what data goes in and why, which makes you the one answerable to your clients and to a regulator. The vendor is the supplier acting on your instructions, responsible for keeping the data secure and doing only what you asked. Both duties are real, and only one of them is yours to explain in a client meeting.

That is a good thing once you see it clearly, because it means the controls are in your hands. You choose the scope of access, you decide which client data the AI Employee can reach, and you decide which actions need a human yes. Responsibility and control land in the same place, which is the only arrangement that ever works.

**Check your own contracts first.** Many client agreements require you to notify or get consent before adding a new supplier who touches their data. Read that clause before you connect anything, because it is the obligation people forget and the one clients notice.

## What should you tell your clients?

Tell them plainly which supplier you use, where the data sits, that it is not used to train models, and how deletion works. Four sentences in your privacy notice covers most situations, and telling them early is far easier than explaining after they discover it themselves.

The instinct to say nothing is understandable and usually wrong. Clients rarely object to a tool that is properly controlled. They object to finding out from somewhere else, which turns a routine supplier update into a trust problem you then have to repair.

Nadia runs a nine person marketing agency in Dublin, and two of her clients have contracts requiring European data residency plus written notice before any new supplier processes their data. Before connecting anything she checked the hosting region, confirmed in writing that content is not used for training, and sent both clients a short paragraph naming the supplier and what it would touch. One replied within an hour with a thumbs up, the other asked two questions and then agreed. Total elapsed time was three days, most of it waiting, and the whole matter is now settled rather than hanging over her.

## What if a client contract says data cannot be shared with third parties?

Read the clause closely, because most of them permit suppliers who process data on your behalf under a written agreement, and forbid something quite different: selling data or handing it to unrelated parties. Your email host is already a third party under a strict reading, and nobody treats that as a breach.

Where the clause really is absolute, you have two honest options. Ask the client for written permission, naming the supplier and the safeguards, or keep that client's data out of the AI Employee's scope entirely. The second is genuinely workable, because access is scoped per account and per folder rather than all or nothing.

Now the limits, said straight. No vendor can promise your regulator will agree with your judgement calls. Deleted data lingers in encrypted backups until those backups expire on their schedule, and any vendor claiming otherwise has not thought it through. An AI Employee can also be manipulated by instructions hidden in content it reads, which is why approval on anything leaving your company matters more here than anywhere else. And a rule you write in a policy document is not enforcement, so the enforcement has to live in the permission scope.

### Diligence before customer data goes near an AI tool

1. **1. Read your client contracts for supplier clauses** — Look for notification, consent, and confidentiality wording. Do this first, because it can change what you are allowed to connect.
2. **2. Confirm the hosting region in writing** — Ask which country the data sits in. Global is a marketing word, not an answer, and some client contracts require a specific region.
3. **3. Get the training answer in writing** — A plain confirmation that your content is not used to train models. Keep the reply in your compliance folder alongside the contract.
4. **4. Check the deletion path end to end** — How to delete one record, how to delete everything, and how long backups holding it survive. Vague answers are the warning sign.
5. **5. Scope access away from your most sensitive clients** — Connect the mailbox and folders the job needs. Leave restricted client material outside the AI Employee's reach entirely.
6. **6. Tell your clients before they ask** — A short paragraph naming the supplier, the region, and the safeguards. Notification is cheap, and discovery is expensive.

| Obligation you hold | What it means in practice | How it is met |
|---|---|---|
| Use only the data you need | Do not connect everything by default | Scope access per mailbox, folder, or project |
| Keep it secure | Encryption, separation, controlled staff access | Per customer separation and restricted, logged access |
| Honour access and deletion requests | You must be able to find and remove a person's data | Search your workspace, delete per item or in full |
| Control your suppliers | Written terms, known region, no surprise reuse | European hosting, no training on your content |
| Show what happened | Reconstruct events when a client asks | A full action log you can read and export |
| Prevent unauthorised disclosure | Nothing leaves without a human decision | Approval gates on sending, spending, and deleting |

## FAQ

### Do I need consent from customers before an AI reads their data?

Usually not a fresh consent, because your existing lawful basis for handling their data generally continues to apply when a supplier processes it on your behalf. What you often do need is to update your privacy notice and check whether client contracts require supplier notification. If your sector carries extra rules, get that confirmed by someone qualified.

### Is data stored in Europe?

Yes, the infrastructure runs in Europe on managed servers rather than a general public cloud storage service. If a client contract names a specific country, raise it before you sign up rather than after you have connected an account.

### How do I handle a deletion request for one person?

Find their records in your workspace and delete them, which removes them from live systems straight away. Encrypted backups containing them expire on their normal rotation, and that window is the number to quote if a client asks for precision.

### Can I keep one client completely out of the AI Employee's reach?

Yes. Access is granted per account and per folder, so you can leave a restricted client's material entirely disconnected. This is the cleanest answer when a confidentiality clause has no room in it.

### What happens if there is a data breach?

You would be notified so you can meet your own reporting obligations, which in Europe are tight and measured in hours rather than weeks. Ask any vendor how quickly they commit to telling you, because that number decides whether you can meet your deadline.

### Is an AI Employee riskier than hiring a contractor for the same work?

It is different rather than uniformly worse. A contractor cannot be scoped to a single folder with every action logged, which favours the AI Employee. A contractor also will not be influenced by text hidden inside a document, which favours the human. The approval gate is what closes that second gap.

**Tags:** ai-gdpr, customer-data-protection, ai-confidentiality, ai-compliance, ai-employee