Why "sovereign"?
Every AI answer you receive is produced by a computer somewhere. The only question that matters is who controls that computer — and the data flowing through it. This module is the "how" behind our Philosophy page: what it takes to run AI where you hold the keys.
1. The Four Things You Own
"Sovereign AI" means no outside party holds a copy of your working context. In practice, it comes down to owning four things:
The Model
The "brain" that generates answers. Sovereign means you choose which model runs, you can swap it anytime, and no vendor can retire it out from under you.
The Data
Your documents, prompts, memory, and history. Sovereign means they live on storage you control — and never train someone else's product.
The Infrastructure
The computer the AI runs on. Sovereign means you decide where it sits — your office, your rack, or a private cloud you contract for.
The Policy
Retention, access, and logging rules. Sovereign means you write the policy instead of accepting a provider's terms of service.
2. The Control Spectrum
Not every task needs the same level of control. There are three places your AI can run — each with an honest trade-off:
Public AI Services
Least controlThe fastest start and the lowest cost. Your prompts are processed on someone else's computers under their retention rules. Fine for: public marketing copy, generic research, anything you'd be comfortable posting on a billboard.
Managed Private AI
BalancedA dedicated, isolated deployment operated for you. Your data stays in your tenant with your retention policy; the operating labor is absorbed by specialists. Fine for: organizations that want sovereignty without hiring infrastructure staff.
Self-Hosted AI
Most controlThe model runs on hardware you own. Prompts never leave your network, and you audit everything. Fine for: regulated work, client-adjacent data, and anyone whose compliance posture says "it doesn't leave the building."
3. Where Your Prompts Actually Go
When you hit enter, your prompt travels to a computer that processes it in memory — that part is true of every provider and cannot be avoided. What can be controlled is what happens afterwards:
Retention is a window, not a yes/no. Providers differ wildly: some keep prompts for minutes of monitoring, some for up to 30 days, some indefinitely by default.
Zero Data Retention (ZDR) exists — but ask for it. Some providers only stop storing your data after you enable a setting or sign an agreement. Never assume it's the default.
Treat retention as "unknown until verified." Before sending anything sensitive to a hosted AI, the question to ask is: after you process my prompt, what do you keep, and for how long?
The dividing line: does the task contain client identifiers, legal matters, medical or financial details, or anything covered by a contract? If yes — that prompt belongs on sovereign infrastructure.
4. Running It Yourself
What it actually takes to host your own AI.
Right-size the hardware
A single modest server handles private chat and document Q&A for a 10–50 person organization. You do not need a GPU cluster or a dedicated ML engineer.
Choose an open model
Open-weight model families (Llama, Mistral, Qwen and others) run on your own hardware with tools like Ollama or llama.cpp. If a better model appears next quarter, you swap it — no renegotiation.
Keep the documents home
Pair local inference with storage you already own (your Nextcloud, your file server). The AI reads your documents where they live — nothing is uploaded to a third party for processing.
Write the policy down
Three rules to start: retain nothing by default, log every action to storage you audit, and grant access by role. A policy you wrote beats a policy you inherited.
Honest trade-off: self-hosting means you own patching, capacity, and backups. That labor is real. If nobody on your team wants it, Managed Private AI exists precisely for you — you still own the data and the policy; someone else racks the servers.
5. The Decision Checklist
Before sending work to a public AI service, run it through five questions:
Is the content already public, or non-sensitive? Yes → a public AI service is fine.
Does it contain client identifiers or confidential material? Yes → sovereign.
Is there a compliance obligation — legal, medical, financial? Yes → sovereign.
Would you be comfortable if a stranger at a vendor read this prompt? No → sovereign.
Could you explain, in an audit, exactly where this data went and why? No → sovereign.
When in doubt, run it where you own the keys.
Knowledge Check
Three quick questions to test yourself. Nothing is submitted; this is just for you.
1. Which four things do you own in a sovereign AI setup?
2. What does "zero data retention" mean?
3. A task includes client account numbers. Where should it run?