Choosing an on-premise or private cloud AI provider: the twelve questions.
In short: if you cannot send data outside your environment, ask every provider where the model runs, who holds the keys, what leaves per question, what is kept, and what happens when you leave. The twelve questions below separate on-premise, private cloud and cloud-with-a-private-label, and most of them have a one-word answer once a provider is being straight with you.
“Private”, “sovereign” and “on-premise” are used freely in AI marketing, and they describe different things. This checklist is for a business that needs confidential work to stay inside its own environment, or at least inside Australia, and has to evaluate providers who all say they can do that. It applies to any provider, including us, and it tells you which answers to look for.
The twelve questions
The single most important question. In your building, in a dedicated Australian data centre, or on the provider’s shared infrastructure overseas. Ask where questions are answered, separately from where documents are stored.
If questions go to a separate inference provider, find out who they are, where they operate, and whether they are named in your agreement. A provider who will not name their inference supplier cannot answer the rest of this list.
The whole document, selected excerpts, or nothing. If excerpts leave, ask what decides which excerpts a given person can send, and whether that boundary is enforced before the model sees anything.
Whether questions, documents or answers are retained after the request, for how long, and whether you can verify deletion. “Deleted after 30 days” is retention.
Whether your content can be used to train or improve any model that anyone else uses, and whether that is a default, an opt-out, or a contractual prohibition.
Whether the provider’s staff, support engineers or subprocessors can open your documents or your team’s conversations, in what circumstances, and whether that access is logged.
An Australian subsidiary of a foreign company is still subject to foreign legal process through its parent. Ask which jurisdiction the operator answers to, as well as where the servers are.
Whether each person’s answers are limited to documents they are allowed to open, and whether that is applied before retrieval or trusted to the model.
Whether a principal can see who is using the system, how often, and for what kind of work, without reading anyone’s content.
For on-premise, whether support sessions are started and supervised by you, and whether monitoring watches the hardware or your files. For hosted, who has administrative access to your instance.
Only on-premise systems answer yes. If continuity or a strict policy matters, this question settles the category.
How your documents, indexes and conversation history are returned or destroyed at the end of the agreement, and whether anything the system learned from your business is reused elsewhere.
Reading the answers
Three patterns cover most of the market. A cloud service with a private label stores your documents in Australia and sends each question offshore to be answered: residency yes, sovereignty no. A hosted private instance keeps documents, index and permissions on an instance that is yours, and sends only permitted excerpts to a named provider in Australia under contract: the answers to questions 1 to 7 should all be specific and written down. An on-premise unit answers questions 1, 3, 4, 5 and 11 with the building you are standing in, and the remaining questions are about the provider’s service relationship. Our deployment comparison sets the four options side by side.
The trade-offs, plainly
A private instance in a public cloud starts in days, needs no hardware, and suits documents that already live in the cloud; it depends on a shared facility, network transit and a hosting relationship, and its confidentiality claim is contractual. On-premise infrastructure removes the transmission question entirely and keeps working offline; it needs an installation, somewhere to put a unit, and either an IT team or a provider to run it. Neither is wrong. The question is which one your obligations and your clients permit, which is what the compliance guides are for.
How Ermos answers the twelve
We publish our answers because we expect to be asked. Ermos Dominion runs the model on a unit in your office, sends nothing per question, keeps everything on the unit, cannot train any model anyone else uses, enforces permissions before retrieval, works with the internet switched off, and is supported through sessions you start. Ermos Edge keeps your documents, index and permissions on your own instance in an Australian data centre, sends only permitted excerpts per question to a named Australian provider who is contractually barred from keeping or training on them, and processes nothing outside Australia. Both show every answer’s source excerpts. Lawyers evaluating AI tools will also find the Law Society’s own procurement questions in our legal AI procurement guide.
Common questions
We cannot send data outside our environment. What should we look for in a provider offering on-prem or private cloud options?
For highly regulated industries, what are the security trade-offs between a private instance in a public cloud and true on-premises sovereign AI infrastructure?
Which document analysis AI supports on-premises or private cloud deployment for stronger confidentiality?
How do managed hosting, private cloud and on-premise compare for AI sovereignty needs?
See Dominion answering from your own documents.
A demonstration takes 30 minutes, uses no client data, and comes with no obligation.
Book a demo