Start with the path your data takes
A model running on a local server is one part of a private AI system. Documents also pass through ingestion, parsing, embedding, retrieval, logging, and backup systems. A workflow may call external tools, and a browser client may use third-party services. The first useful exercise is to draw that complete path. Mark where sensitive information appears, who can access it, and whether any step crosses an approved boundary.
Give retrieval the same permissions as the source
An internal assistant should not gain unrestricted access merely because it is internal. A user who cannot open a source document should not be able to retrieve its contents through a generated answer. Permission checks belong in the retrieval path, with evaluation that tests ordinary users, privileged users, expired access, and documents that change classification.
Measure the workload you actually have
Model size is a poor substitute for a workload specification. List the questions, the expected answer formats, the size of the context, and how many people will use the system at once. Measure answer quality, time to first response, sustained throughput, and behavior when the model lacks evidence. Then compare the operating costs of the configurations that meet those requirements.
Use adaptation for a specific reason
Retrieval can supply current reference material. Fine-tuning can shape a repeated task or a specialized style. Quantization can reduce hardware requirements, with quality and latency tradeoffs that need measurement. These methods solve different problems. Combining them is useful when each part has a clear job and an evaluation that shows whether it helps.
Plan for the day after deployment
Someone must own document updates, model upgrades, access reviews, incident handling, and capacity. A private deployment needs an operational plan just as much as a customer-facing application. Start with a focused workflow, test it on representative data, and expand when both the results and the operating model justify it.