The sovereign workspace checklist: 8 questions to ask any vendor

If you are evaluating AI workspaces right now, you will meet the word “sovereign” on nearly every vendor’s website — including this one. The trouble is that the word is free to print, while the architecture behind it takes years to build. So the same label now covers everything from genuine customer control to an ordinary cloud product with a regional flag on it. We covered what sovereignty should mean in the definition piece; this is the practical companion — how to test a vendor’s claim before you sign.

The test is eight questions, designed to be asked in writing, because written answers are hard to blur — and blurring is exactly what a hollow “sovereign” depends on. For each question you get four things: why it matters, the answer you should want, the evasive answer that tells you everything — below, we call it the dodge — and, since fairness demands it, CloudTop’s own answer stated plainly — CloudTop being the AI workspace you own, whose blog this is. Ask us the same eight, and hold us to them.

The old way: the RFP that asks everything except what matters

Standard vendor questionnaires are long where it’s cheap — certifications, region lists, uptime — and silent where it’s decisive: key custody, traffic paths, memory portability, exit mechanics. Vendors optimized for the questionnaire generations ago. The eight questions below are short precisely so they cannot be answered with an appendix.

The eight questions

1. Who holds the model keys?

Why: the model is the AI engine itself; the key is your account credential with its maker. Whoever holds the keys owns the relationship and the breach surface. Want: “You do — keys live in the user’s settings, on the user’s device.” Dodge: “Keys are securely encrypted in our vault” — that is custody with adjectives. CloudTop: the key sits in your own settings, on your own device; we never hold it.

2. Does our traffic transit your infrastructure?

Why: a proxy can log, retain, and be compelled; a path you don’t transit can’t. Want: “Requests go from the user’s device to the model provider directly.” Dodge: “We use enterprise-grade encryption in transit” — answering security when you asked topology. CloudTop: we never proxy the traffic; requests go from your device to the model you picked.

3. Do you bill for inference?

Why: an inference margin is a structural incentive to sit in your data path forever. Want: “No — you pay model providers directly, at their prices.” Dodge: “Our bundled pricing simplifies procurement.” CloudTop: we never bill for inference.

4. Can we choose and swap models freely — per task, on our own keys?

Why: the frontier moves quarterly; a fixed model menu means your capability is the vendor’s roadmap. Want: model choice belongs to the user, changeable at any time, with the keys in your custody. Dodge: “We support all leading models” (integrated ≠ yours to route). CloudTop: bring your own model — swap per task with your own keys.

5. Is our AI memory a portable asset?

Why: per-vendor memory is the deepest lock-in in the stack — it compounds daily. Want: memory as inspectable files in your workspace, readable by any connected model. Dodge: “Your assistant learns your organization over time” — that is the lock-in, described warmly. CloudTop: team memory as shared workspace folders is the design we’re shipping toward with the public release — stated in the future tense on purpose, in one memory, any model.

6. Where can the workspace deploy — and who operates it there?

Why: a regional flag on the vendor’s cloud changes the disk’s address, not your dependency. Want: your cloud, on-premise, or a provider you choose — your call. Dodge: “Our sovereign tier runs in-region.” CloudTop: deploy anywhere; no forced destination.

7. Is there one audit trail for human and AI actions?

Why: if AI actions live in a separate log (or none), every incident becomes a correlation project. Want: one event stream; every AI reply traceable to the tools it used. Dodge: “Full logging is available on the enterprise plan.” CloudTop: one stream; every AI action carries its visible tool trace.

8. Walk me through leaving you.

Why: the exit answer prices every negotiation you’ll ever have. Want: a procedure — what moves, in what form, what keeps working. Dodge: a partnership philosophy. CloudTop: the workspace is the unit that moves — apps, data structure, agents, permissions — because the environment runs where you point it.

One closing note of honesty: no vendor, us included, should get credit for answers in a blog post. Get them in writing, attached to the contract. Sovereignty that isn’t contractual is a mood.