Private AI for business: what private should mean

"Private AI" is on a lot of websites now, and it rarely means the same thing twice. For one vendor it means a business plan that does not train on your chats. For another it means a model running on hardware you look after. For a third it is just a word on the pricing page. This guide sets out what private should mean for a small business, what it cannot mean, and the questions that tell the difference.

Private AI vs public AI

Public AI is the chatbot anyone can sign up for. Each person brings their own account, pastes in whatever they need, and the business has no say over the settings, the history or who else can see it. When people search for a "private ChatGPT", they usually mean the opposite: an AI the business controls, that their team can use on real work without client details ending up in personal accounts.

That is a fair thing to want. The trouble is that "private" is easy to claim and hard to check, so it helps to define it in terms you can test.

What private should mean: four tests

For a small business, private AI should pass four tests.

• Scoped access. The AI sees only the systems and records you agreed to, for the task in front of it. A question about one client should not pull in the whole client list. Access is set by rules, not by whatever someone happens to paste.
• Logging. Every time the AI looks at something, it is recorded, and someone can review the record later. If you cannot answer "who saw what, and when?", it is not private in any useful sense.
• People approve changes. Reading is one thing. Changing a record, sending an email or updating a customer file should wait for a person to approve it.
• Data you own and can export. Your business data, and what the system learns about your business, should be yours in the contract, with the right to take it all with you. You should own your data, whoever provides the software.

Notice what is not on the list: where the box sits. A server in a cupboard with no logs and one shared password fails these tests. A managed server with scoped access, logs and approvals can pass them.

What private does not mean

The capable AI models in use today are large, and most businesses use them as a service rather than running their own. Be wary of any vendor that implies your information is never seen by the model that answers your question. If the AI is going to answer a question about your data, the model has to see the part of the data the question needs.

So the honest version of private is not "the model never sees anything". It is "a controlled layer decides what the model may see, for which task, and keeps a record". That layer is where your rules live. It is also what lets you change models later without rewriting the rules.

Private is also not the same as compliant. A private setup can make your obligations easier to keep, but it does not certify anything, and it does not replace the agreements you need with each vendor.

Three ways businesses get private AI

A business plan of a public chatbot. This is the quickest step up from personal accounts. You get work sign-ins and admin settings. It sees what people paste in or connect, such as shared files, and any actions it takes follow the vendor's settings rather than approvals and logs you set up yourself. For one vendor's plans in detail, see Is ChatGPT safe for business?

Build your own. Some businesses ask how to build an internal AI system without sending company data to outside providers at every step. It can be done, but someone has to choose the model, host it, connect it to your systems, write the access rules, keep the logs and fix it when it breaks. For most small businesses that job lands on whoever is most curious, on top of their real work.

A managed private AI service. A team builds and runs the controlled layer for you, connected to your systems, and supports it. You trade some choice over the parts for not having to run them.

What private AI costs

People ask about private AI server cost and private ChatGPT cost, and there is no single figure. The cost depends on the server, the model usage, the number of systems connected, how much is built, and who supports it. For a managed service, expect three parts: an assessment, a build, and an ongoing fee for running it. How much managed AI services cost breaks down what to compare.

How we do it: EdgeKeeper

EdgeKeeper is our software on a dedicated Microsoft Azure server that we set up and manage for your business. It has no public internet address and restricted outbound access. It connects to the systems you choose, under permissions you agree.

• Each question runs read-only, in a temporary sandbox.
• System credentials stay on the server.
• A person separately approves a specific change before it runs.
• Access is recorded for review.
• We agree what information the system can use, where it is stored, and what may be shared with any outside service it uses.
• Your business data and knowledge are yours, with full export rights in the contract. The server and the EdgeKeeper software are ours.

The AI model that answers is our engine, and it may change over time. What makes the setup private is the layer around it, not the name of the model. Your team works in Switchboard, where it can make as many read-only AI agents as it wants. The security page sets out the controls and their limits.

Questions to ask any private AI vendor

• What can the AI see, and who decides?
• Is every access logged, and can we read the log?
• Can it change anything without a person approving it?
• Which outside services receive our data, and under what terms?
• Is our data used to train any model?
• Can we export all our data and what the system learned, and what happens if we leave?

Where to start

Private AI works best when it replaces something staff are already doing in personal accounts. Find that first with how to detect shadow AI, then set the rules with our AI policy template.