Private AI that is private
by architecture.

What private AI means

Private AI is an architecture in which every component of the system — the model, the retrieval index, the embeddings, the orchestration and the audit log — runs inside infrastructure the enterprise controls. No component depends on an external service to function.

The distinction worth holding on to is between privacy by policy and privacy by architecture. Privacy by policy is a contract saying a vendor will not train on your data. Privacy by architecture is a system in which there is no outbound path for that data to take. Only the second survives a change of ownership, a change of terms, or a breach on the other side.

Private AI compared with public AI

Where inference happens

Public: a shared tenancy you do not control. Private: hardware inside your perimeter.

What leaves the building

Public: every prompt and attachment. Private: nothing.

Cost behaviour

Public: metered per token, so cost rises with adoption. Private: fixed infrastructure, zero per-token cost.

Failure mode

Public: a provider outage or a terms change is your outage. Private: your availability is your own.

Audit position

Public: you are attesting to someone else’s controls. Private: you are attesting to your own.

Sovereign AI solutions and data sovereignty

Sovereign AI is AI whose models, data and operation remain under the legal control of the organisation running it and of the country it answers to. Data sovereignty is the property underneath it: data stays subject to one jurisdiction’s law because it never leaves that jurisdiction. For Indian enterprises that means data residency satisfying RBI localization requirements and processing that fits the DPDP Act, demonstrated by where the system runs rather than by what a contract promises.

Sovereignty is an architectural property. If the design requires an outbound call to work, the system is not sovereign, however the paperwork reads.

How private AI is built

Most deployments run a mixture-of-experts configuration — several LLM and ML models working together behind one interface, each selected against what the agent actually has to do, rather than one general model asked to do everything.

Security is enforced in the system. Each agent runs with only the permissions its task needs, scoped to the session and limited in time; every action it takes is recorded; anything above the approval threshold waits for a human. Six principles govern this, and they are implemented in the runtime rather than written into a policy document.