For us, GDPR and the EU AI Act come first. Not because clients ask, but because we want every workflow to be compliant from day one – and still simple to run for our clients and for us.
So we’re testing European models on ourselves first.
What we use right now:
Mistral Small answers customers in our Telegram assistant
Mistral Medium reviews every reply after it’s sent
Training opt-out switched on in the Mistral console, minimal personal data in team alerts, successful executions not stored in n8n
One thing we learned:
Small once told a customer that the price “includes” two months of support – in fact that’s a time-limited bonus. Medium rated the answer as fine. We caught it ourselves and added a small Code node that checks price answers before they go out. A second model helps, but it doesn’t replace fixed rules.
Next steps:
We’ll give the Mistral setup some more time, then test IONOS AI Model Hub (hosted in Germany) and see whether combining both makes sense.
Your experience would really help here:
- Which EU or self-hosted models do you run with n8n, and what made you choose them?
- How well do they handle German in real conversations?
- Has anyone combined two EU providers, for example one for answers and one for checks?
Lessons from what didn’t work are just as welcome. We’ll share our results here as we go.
1 Like
That price example is the classic one tbh, and it’s not only prices. Anything the bot could “promise” (price, what’s included, deadlines, refunds, availability) is where a second model reviewing doesn’t help much, it reads the answer the same friendly way the first one wrote it.
What I’d do: the model doesn’t write that stuff at all. It pulls the exact wording from a source you keep anyway (price list, offer page, or a small Sheet copied from it), so no extra admin, you only touch it when your prices change. If a customer asks something that isn’t covered there, the bot doesn’t guess, it says it’ll check and pings a human. Those questions are basically your list of what’s missing in the source, so you add it once and it’s covered next time.
Then your Code node only has to check one thing: is there a number, date or “includes” in the reply that isn’t in the source → hold it, human decides.
For German I’d test with 20-30 real customer messages from your history, not made-up ones. You see the weak spots way faster.
Curious what IONOS gives you, keep us posted 
1 Like
Thank you, this was really helpful. We built it today and tested it right away.
What we have now:
every amount, percentage, time period and date in the answer is compared with our price list
if something is not there, the answer is not sent. The customer gets “let me check with our team”, and we get the draft
the bot keeps answering other questions while the team checks
What the tests showed:
Numbers were not enough, exactly as you said. The bot wrote “maintenance is included” with no number at all. So now sentences with “included”, “free” or “separate” are only allowed for topics that are in our price list.
Our first version paused the whole chat after a held answer. A customer asked the next question and got no reply. Now only the held question waits.
For a discount question the bot wrote “we don’t offer discounts” by itself. That’s a business decision, not the bot’s. Now discount questions always go to the team, and the customer gets one fixed sentence.
So in the end, more rules and less freedom for the model, but only where money or promises are involved.
For German, we’re collecting real messages now and will test once we have a good set.
Still curious: how do you check “includes” on your side? Exact text from the source, or the model names the line it used?
Glad it worked out, and the discount point is a good catch. That really is a business decision, not something the bot should answer.
I’d do both: the model names the line it used, and the Code node then checks that the claim actually matches that exact line. Exact text alone breaks as soon as the model rephrases, and a line reference alone only tells you where it looked, not whether it read it right. Together they’re cheap and easy to log.