Your AI. In the right place.

What you type in, the documents you share, the corrections you make and how you use it: the know-how of how your business works. In his essay “The Reverse Information Paradox” (as reported by The New Stack), Satya Nadella puts it as paying for intelligence twice, “once with money, and again with … the proprietary knowledge you must reveal to make that intelligence useful.”

The more of your work flows through someone else's model, the more of your know-how sits with them, and the more it costs you to leave. Prices, terms, retention and model changes are theirs to set. You end up renting back a body of knowledge that started as yours.

Running your steady, repeatable and sensitive work on an AI model you hold a copy of, rather than one you can only rent. It looks things up in your own documents, can be trained on your terms and formats, and runs on hardware you own or in an Australian private cloud you control. A single entry point decides what, if anything, still goes to a rented model.

Mostly, no. Documents that change, such as procedures, client files and price lists, are looked up: the model finds the pages it needs when someone asks, answers from them and shows which ones it used, so an update counts as soon as it's saved. Training is for what stays the same, such as your terms, your report formats and the way your team writes. It adds a small layer to the model, and the Ownership Pilot shows whether it's worth doing.

For some work, yes, and we'll tell you when it is. Some big cloud vendors will train a model on your data inside your own account with them, which fits when you're committed to their cloud. Many of the best-known rented models can't be trained by the customer at all, and where they can, the trained model stays on the vendor's cloud, under their terms. A model you own runs on hardware you choose, is tested on your own examples, and is yours to keep: no vendor can put the price up or take it away.

No. Most businesses keep their productivity suite and cloud. Our Private Model Service can send work to your existing cloud account, or to an Australian data centre, where that's right, and to a model you own where owning pays.

Not always, and not never. It pays where the work is steady and repeatable, especially once AI agents multiply how much AI you use. We'll show the payback on your own numbers, and when renting is cheaper we say so.

Not on everything. On many everyday business jobs it can do the work well enough, but you only know by testing it on your own examples. That's what the Ownership Pilot measures before anything is bought.

For most everyday jobs, yes, once the hardware is sized for how many people use it at once. Memory decides which models fit, and how quickly the machine reads its memory decides how fast answers come back. A slow answer costs your team time, so the Ownership Pilot measures speed on your own work before anything is bought.

Many of the strongest models you can own today were made in China; others are American or European, and more American ones are arriving. “Allied” here means models from the US, UK, Europe and other close partners. For every model we show or install, we record where it comes from, its licence, and a digital fingerprint proving it hasn't been changed. You can accept any origin, prefer allied models, or use allied models only. If the rules on a model change, the record makes it a swap, not a rebuild.

You do: your data, the records of what the AI did, the test sets used to check it, the trained model, and the library of your documents it looks things up in, plus the register that lists them, agreed in writing for each engagement. Nadella describes the same list as what should “accumulate and improve together” inside a firm's own boundary (The New Stack).

It depends on the vendor's terms, which is why the Inference Audit checks them. Nadella argues it should be every firm's right; until it is, read the contract.