Nobody sets out to leak customer data. Somebody in the office has a long, awkward email from an unhappy client, or a spreadsheet of enquiries to tidy up, and a free AI tool will do it in ten seconds. They paste it in, get a lovely answer back, and carry on with their day. The data has now left your building, and you had no say in it.
That's the AI risk most small businesses miss: ordinary people being helpful on a personal account you don't control.
How common is it?
Irish small firms are mixed on AI, but plenty are using it. Research commissioned by Google from Amárach, reported by RTÉ in March, surveyed 400 Irish SMEs. One in five did not use AI at all, rising to a third of micro-businesses, and the main barriers were fear of mistakes, lack of skills and cost.
What happens inside the firms that do use it is the worrying part. Security firm LayerX studied how people use generative AI at work in large organisations, as covered by The Register. That's enterprise data, not Irish small business data, so treat it as a pattern rather than a prediction. But the pattern is hard to ignore:
That last figure is the one that matters. A personal account means no company settings, no admin and no record of what was shared.
What the Data Protection Commission expects
The DPC has published guidance on AI and data protection aimed at organisations using these products, not just building them. In plain English, it asks you to do four things:
- Know where the data goes. Understand what personal data the tool uses, how, and where it ends up
- Understand the vendor. If you use a supplier's AI tool, you need to know what protections it has and what it does with what you give it
- Be able to honour people's rights. Access, correction and deletion requests still apply to data that went through an AI product
- Keep a human in the loop. Relying on AI output without review can lead to harmful decisions about real people
None of that is exotic. It's the same GDPR thinking you apply to any supplier, extended to a tool your staff may have signed up for themselves in five minutes.
Banning it versus managing it
"Nobody uses AI here"
People use it anyway, on their phones and personal logins. You get all the risk and none of the visibility.
"Use this one, never put that in"
One approved tool, a short list of things that never go in, and staff who know why. Visible, simple and enforceable.
A policy you can write this week
Ask, don't accuse
Find out which tools people already use and for what. You will learn what the job actually needs.
Write the never-paste list
Customer names with contact details, health or financial information, passwords, contracts and anything under an NDA.
Choose one approved tool
Read its data terms before you decide. If you can't find out what happens to your inputs, it's not the one.
Brief the team
Ten minutes, real examples, and a rule that a person checks anything AI-written before it goes to a customer.
When it's worth running your own
Some jobs, such as summarising client emails, drafting quotes from your own files or answering common questions, involve exactly the data you would rather not send to a third party. For those, an AI agent that runs on a machine you own keeps the data in-house, and you decide what it can see. It isn't the right answer for everyone, and a sensible policy comes first. But it's a real option now, not just something big companies can afford.
Summary
The risk is ordinary staff being helpful on accounts you can't see. Find out what is already happening, set a short never-paste list, pick one approved tool, and have a person review what goes out. That covers most of what the DPC is asking of you.
