What is private AI?
Private AI usually means a model only your business uses, in a cloud setup reserved for you or on hardware in your building, so you control where your data goes. The term describes the setup. It says nothing about how well it's secured.
The National Cyber Security Centre (NCSC) treats security as work at every stage, from design to day-to-day running, wherever a system is hosted.
Does private AI mean the model runs in your building?
Not always. It can be a cloud setup a provider runs for you alone, or hardware you own on site. Only the second keeps everything in your building, and it's the most work: someone has to update it, watch it and fix it.
When is private AI worth it for a business?
When a rule you can point to says the work can't go to a shared outside service, or it has to keep working with no internet connection. If the reason is a general worry about safety, check first whether the business version of software you already have meets your rules.
For example, Microsoft says Microsoft 365 Copilot doesn't use your prompts or files to train its models, and only shows people what they can already open. Check whether that meets your rules before you pay for a build.
What does private AI cost?
Three parts. Usage is running the model, per request or for the hardware. Setup is connecting it to your documents and testing the results. Running it never stops: updates, monitoring and support. Ask any supplier to price all three.
There's no reliable published figure for a business your size, so price your own case, including who does the running and for how many hours a month.
Private AI, on-premise or the software you already have?
The check compares four routes in order. Each later route has to show why the earlier ones won't do, so you only pay for private AI when you need it.
| Route | When it fits | What it has to prove |
|---|---|---|
| Software you already have | Microsoft 365, Google Workspace or another system you pay for has an AI feature. | It does the job within your rules. |
| An app built on an outside AI service | Your software leaves a gap, and an approved outside service is allowed. | The provider's terms, access rules and running costs hold up. |
| A dedicated cloud setup | Your rules allow a hosting provider, but not a service shared with other customers. | Why it's needed, and who keeps it running. |
| Hardware in your building | It has to work with no internet connection. | The model is good enough, and someone can look after the machines. |
What should you test first?
The business version of software you already pay for, against the rules you'd set a private build. If it falls short, write down exactly where, because that gap is what a private setup would have to fix.