Private deployment

In one paragraph

Private deployment means running an AI model inside an environment your organisation controls, rather than sending questions to somebody else's service. It takes one of three shapes: on servers you own, inside a cloud account held in your own name, or fully disconnected from any outside network at all. Whichever shape it takes, the questions people ask and the documents the model reads stay inside that environment, instead of travelling to infrastructure run by someone else.

Why it matters

Where a model runs decides who can see what goes into it. For many organisations, that is not a preference but an obligation: data protection law, rules that apply in regulated sectors, and contracts with your own customers often say, in one form or another, that certain information has to stay within a defined boundary. Running the model inside your own environment makes that true because of how the system is built, not because of a policy someone has to remember to follow. It also turns a bill from an outside service, which can be hard to predict as use grows, into capacity you measure and plan yourself.

How it works

Private deployment can take three shapes:

• On your own servers, inside your own building or data centre.
• Inside a cloud account held in your own name, where you control the boundary even though the hardware belongs to someone else.
• Fully disconnected, with no outbound connection at all, for situations where that is what your policy requires.

In every shape, the model is connected to your existing logins and access rules, so the same permissions that already govern who can see what continue to apply to every question asked. Because the whole exchange, question in, answer out, happens inside your boundary, residency is a property of how the system is built rather than a claim about where a vendor says your data will be kept. Monitoring is put in place from the start, so how the system is used and how it performs can be watched from the first real question onward, not added as an afterthought once something has already gone wrong.

What it looks like in practice

A customer contract says a set of documents cannot leave the organisation's own network. Running the model on servers inside that network means the assistant can read those documents and answer questions about them without any of that material crossing outside the boundary the contract requires. The same login system controls who can ask the assistant anything, exactly as it already controls who can open the documents themselves.

How this connects to our work

Private deployment is what we build under private AI setup. It is closely tied to how data protection law applies to AI, since where a model runs is often what decides your obligations.