A request to connect everything sounds like a request for power. In practice, useful private AI begins by choosing a much smaller boundary. What is the task? Which information is needed? Who should be able to use the result?
These questions are part of the architecture. They determine what the system can read, where it runs, and which decisions remain with its user.
Start with selected context
A computer contains much more than working knowledge. It holds credentials, personal correspondence, unfinished ideas, and data belonging to other people. Treating all of it as model context would remove distinctions that matter.
For the TKS pilot, the starting point is a workspace for selected files. Selection gives the user an opportunity to understand the scope before inference begins. It also makes the first useful result easier to evaluate: an answer can be checked against a defined set of source material.
Separate access from compute
A model running on powerful hardware is still a service with an access boundary. Moving a workload from a laptop to cloud compute does not answer who can call it or which documents should be copied.
Our initial cloud verification used a synthetic document. We treated model availability, GPU inference, private access, and file scope as separate things to check. A successful model response is evidence about that request; it is not evidence that every application or cloud account has been integrated.
Make the lifecycle visible
Private AI includes what happens after an answer appears. A file may need to be removed. A cloud machine may need to stop. An owner may need to revoke access. A useful system should make these actions as understandable as adding the first document.
The public company website has a different job. It explains the work and receives inquiries. Keeping that service separate from a private workspace makes the distinction visible to both visitors and the owner.
Privacy becomes practical when a boundary turns into a clear action.
Grow from evidence
There is room to grow: better retrieval, broader document support, more careful evaluation, and additional model providers. Each addition should begin with a reason. The question is whether it improves the user’s task while preserving an understandable system.
That is the value of a focused first version. It gives the next decision something concrete to stand on.