Where it runs
Four ways to deploy an AI agent.
From a managed EU endpoint to your own hardware — the same agent, four ways to deploy it. Pick what your data policy needs today, and move between tiers later without a rebuild. The deployment tier is an architectural decision, and it is worth making before the use case, not after.
The four tiers
-
T1
EU region
EU Managed
Our EU-resident, zero-retention endpoint (Azure Sweden Central). Fastest to ship — nothing retained, never used for training.
-
T2
Your account
Your cloud
Deployed inside your own AWS, Azure or Google tenant — your account, your identity, your controls. The data never leaves it.
-
T3
Your VPC
Private
A dedicated, network-isolated model endpoint in your own VPC — for regulated data that can’t sit on shared inference.
-
T4
Your hardware
On-premise
Open-weight models on your own hardware. No external call is ever made — nothing leaves the building.
How to choose
The question is not “how secure do we want to be” — everyone answers that the same way. The useful question is what does your data actually forbid?
- Ordinary business data, speed matters → T1. EU-resident, zero-retention, live soonest.
- Your security policy requires everything inside your own accounts → T2.
- Regulated data that cannot share inference infrastructure → T3.
- An absolute confidentiality duty — § 203 StGB, classified material, or a policy that permits no external call → T4.
Most companies over-specify at the start and under-use what they paid for. Starting at T1 and tightening once the value is proven is usually cheaper and faster than the reverse — unless a duty makes T4 the only lawful option, in which case that is the starting point.
Moving between tiers
The agent is built so the inference backend is a configuration decision, not an architectural one. That is a deliberate design constraint carried from four production products, and it is what makes “start managed, tighten later” a real option rather than a sales line.
What changes between tiers is where the model runs and who holds the keys. What does not change is the agent, the workflow, the evaluation set, or the compliance documentation.
True at every tier
- Zero retention. Your data is not stored by the inference layer and is never used for training.
- You own the infrastructure and the source. No tier depends on continued access to us.
- The compliance artifacts come with it — AI-Act classification and documentation, and the GDPR paperwork, regardless of where the model sits.
- The evaluation set travels. Changing tiers changes the model; the way you prove it still works stays the same.
Questions
Does on-premise mean worse quality?
It means a different trade-off. Open-weight models have narrowed the gap and are often more than adequate for a scoped business workflow, but the strongest frontier models cannot be self-hosted. Where the confidentiality duty is absolute that trade is usually right — and it is a decision to make with a measured evaluation set, not by reputation.
What hardware does on-premise need?
It depends on the model and the throughput, and it is scoped in the assessment rather than guessed at here. It is frequently less than people expect for a single well-scoped workflow.
Who operates it once it runs?
Either your team, with the documentation to do it, or us under S3 or an AI Delivery Lead engagement. It is a choice, not a dependency.
Can different workloads sit on different tiers?
Yes, and that is often the sensible answer — privileged data on T4, everything else on T1, one agent, one evaluation set, two backends.
Which tier does your data require?
Tell me what data is involved and what your policy forbids, and the tier answers itself — usually in one conversation.