AI and your company's data: what leaves the building and what doesn't.
"If I add AI, do my data end up in America?" It's a good question, and more and more people ask it. The short answer: it depends how you build it, and you get to choose.
When a model reads an invoice, that invoice has to reach wherever the model runs. For the most capable models, "wherever" is a big provider's server, usually outside Romania. That isn't automatically a problem. But it isn't a detail to wave away either, especially when the invoice carries your customers' data.
Three options
A model over an API, from a big provider. The most capable and the easiest to get running. The business contracts of these providers usually state that the data isn't used to train models and is deleted after a short time. I check that in writing before I use one, and where possible I choose processing inside the European Union.
A model running on your side. Open models have become good enough for reading invoices or sorting emails. They run on your own server or on one rented in the EU, and the data goes nowhere. It costs a little more to set up and is a little less capable. For sensitive data, it's my preferred option.
No AI. If the data is tidy, we don't have this conversation at all. A rule sends nothing anywhere.
What I do whichever option we pick
- I send the model only what it needs. To sort an email, it doesn't need the customer's whole history.
- I strip or mask what isn't necessary: personal ID numbers, bank account numbers, phone numbers, if the task doesn't call for them.
- I write into your company's privacy policy who processes what. GDPR requires it, and your customers deserve to know.
- I keep a log: what was sent, when, to whom. If we switch providers in a year, we know exactly what went where.
Five questions for anyone selling you "AI"
- Where does the model run, physically?
- Are my data used for training? Show me where it says so.
- How long are they kept?
- What exactly do you send: the whole document or only the fields you need?
- Can I move, later, to a model that runs on my side?
If you don't get clear answers to all five, it isn't ready for your data. You shouldn't accept less from me either.
Not sure which case your business is? That's exactly what we find out in a free 30-minute audit. I'll tell you honestly, even if the answer is "you don't need AI".
Read next
AI or plain automation? When to use which.
The question I get most this year isn't "can this be automated?" but "shouldn't we use AI for this?". Most of the time the answer is no. The rule I decide by, with a table and three examples.
ReadWhere AI goes wrong, and how to build around it.
A language model fails differently from a program. It doesn't stall, it doesn't throw an error: it answers nicely, confidently, and sometimes wrongly. The four safety nets I put in every time.
ReadWhat AI actually costs inside an automation.
The good news: the AI part is almost free. The less good news: the AI part isn't what's expensive. Where the money goes, and how I work it out for a client.
ReadTen tasks: five without AI, five with.
A short list from real projects, so you can see the difference at a glance. And the pattern that repeats across all ten.
Read