We are moving to self-hosted AI

For the past two years, Herodesk’s AI features ran on OpenAI’s models. Translation, classification, and eventually the AI Agents that answer customer questions and take action in your webshop, all of it was processed by sending requests to OpenAI’s servers in the US and waiting for a reply.

That’s no longer how it works.

As of the end of May, every AI request in Herodesk runs on our own servers, using open-source models we host and manage ourselves in Denmark and Germany. No OpenAI. No third-party AI provider sitting between your customer data and your support inbox.

Your data stays exactly where you think it does

If you’ve ever had to fill out a vendor security questionnaire, you know how this goes: every sub-processor gets listed, and every sub-processor gets scrutinized. For a growing number of our customers, having a US AI provider anywhere in that list was a dealbreaker, even with anonymized data and a strict no-training policy on OpenAI’s side.

That’s gone now. Every AI request that touches an order, a customer message, or your product catalog is processed entirely on our own infrastructure, hosted in Denmark and Germany. Nothing leaves the EU. Nothing passes through a third party. If GDPR compliance is something your business takes seriously, this removes one more thing you need to worry about when you choose Herodesk.

Lower costs, without passing them on to you

AI usage in Herodesk has grown fast, especially since we launched autonomous AI Agents last October. More customers, more conversations, more complex requests, all billed per token by OpenAI. That cost curve was starting to look uncomfortable.

Running our own models on rented GPU infrastructure changes that math. Instead of a variable, ever-growing bill from a US provider, we control the cost directly. That’s not a detail we’re sharing to sound impressive. It’s the reason we can keep our AI Agents at a very competitive price point.

Nothing changed on your end

Switching AI providers usually means downtime, reconfiguration, or a rocky transition period. Here, it didn’t.

The tools we use to run our own models speak the same language as OpenAI’s API, so from Herodesk’s point of view, it was a straight swap. Your AI Agents kept working exactly as before, on the same channels, in the same languages, with the same response times. If you didn’t read this post, you wouldn’t have noticed anything changed at all.

A quick look under the hood

Without getting too deep into it: we rent NVIDIA GPUs from a data centre in Denmark, connected securely to our servers in Germany, and run open-source language models on them ourselves. Different models handle different jobs, lighter ones for simple tasks like translation, larger ones for the more complex reasoning our AI Agents need to do. It’s more work to run than simply calling an API, but it’s work we’re set up to do well, and it’s fully within our control.

Why this matters going forward

Herodesk has always run on infrastructure we control. Our servers live on Hetzner in Germany, not AWS, Azure, or Google Cloud, and that’s a deliberate choice, not a shortcut. Moving our AI stack to the same model just extends that same principle to the part of the product that’s growing fastest.

For you, that means a support platform where your customer data stays in the EU, your costs stay predictable, and the AI features you rely on keep getting better without a third party in the middle of it.