What’s private AI?

Private AI is an AI assistant that runs inside an environment your business controls, works with your own company information, and keeps that information out of anybody else’s hands. The model, the data and the processing all sit somewhere you can point to. Nothing your people type is sent off to a public AI service to be answered.

The phrase gets used loosely. Some vendors call a business subscription to a public tool “private”. Others mean a model running on servers in your own building. Both use the same word for very different arrangements, so it’s worth knowing what to ask.

What makes AI private

Four things decide it. If any one of them is missing, your information is going somewhere you didn’t choose.

• Where the model runs. In a private setup the AI model runs inside your environment, or your provider’s private environment. A public tool sends each question to the AI company’s own servers, usually overseas.
• Where your information is kept. Documents, conversation history and anything the assistant learns about your business are stored somewhere you know and can name.
• Who can see what. A finance manager and a site supervisor can use the same assistant and see different information, the same way they do in your ERP.
• What your information is used for. It answers your questions and makes your own assistant better. It never goes into training ChatGPT or any other public AI.

How it compares with a business subscription

A business or enterprise account on a public tool is a real step up from staff using free personal accounts. The provider usually commits not to train on your data, and an administrator gets some control. For general work, that’s often enough.

The difference shows up with sensitive work. On a business subscription, your contract, your pricing file or your board paper still leaves your environment, gets processed on the provider’s infrastructure and is held there for a time under their terms. A private setup keeps all of that inside the environment you chose. The question is whether your customers, your board and your insurer are comfortable with that information travelling at all.

A business subscription limits what happens to your data. A private environment limits where it goes in the first place.

Why it’s more useful, as well as safer

Most conversations about private AI start with security. The bigger day-to-day benefit is usefulness. A public tool has never seen your P&L, your contracts or your cost model, so it answers in general terms. A private assistant can be given that information on purpose, with permissions, and then answer the question you actually asked.

Ask a public tool where your margin went and you’ll get a sensible list of things manufacturers usually check. Ask a private assistant that can read your P&L, your customer master and your freight costs, and it can name the customers and the reason.

What to ask a vendor

• Where exactly does the model run, and who operates it?
• Is my question sent to any outside AI service to be answered, even for a moment?
• Where is conversation history stored, and who decides how long it’s kept?
• Can different people in my business see different information?
• Is anything we put in used to train a model other than our own?
• Can we have our data deleted, and how would we know it was?

These are the questions Australia’s cyber security agency asks too. The Australian Signals Directorate’s guidance on engaging with AI tells organisations using a third-party AI system to find out whether their inputs will be used to retrain its model, and how their data will be managed if the agreement ends.

If the answers are vague, the arrangement is probably less private than the brochure suggests.

The deletion question matters more than it looks. The OAIC warns that once personal information goes into a generative AI product, it may be “potentially impossible to remove the information from the system”.

Is it only for large companies?

It used to be. Running your own AI meant buying hardware and hiring specialists. A mid-sized business can now have a private assistant running on hosted infrastructure in Australia, set up and supported by a provider, without building a data science team. The sensible way to start is small: connect the few sources that answer your real questions, give it to a handful of people, and widen it once it has earned its place.