Local AI is a deployment choice, not a complete privacy policy
A model can run on your computer while its surrounding application still uses external services. Map the whole workflow before sending sensitive data.
- Confirm where the selected model actually runs.
- Include tools, storage and backups in the data-flow review.
- Do not treat fewer refusals as evidence of stronger privacy.
Start with the destination, not the interface
The phrase local AI is most useful when it identifies where inference runs. It says much less about the surrounding application: a chat interface may offer both local and cloud models, connect to tools, or save a conversation. A familiar desktop window is not itself evidence of an entirely local workflow.
Ollama’s documentation distinguishes local operation from cloud capabilities and describes configuration choices. Its developer announcement for the Claude Desktop integration likewise offers both local and cloud model selection. That distinction is a reason to inspect the selected route, not to assume the same data path for every request.
Draw a simple data-flow map
Before sending a private document, list each stage: where the document is opened, where text is extracted, where the model runs, where tools send requests, and where outputs are stored. Record what you have verified and what is only a provider’s statement.
Then try a harmless example. Confirm the selected model and endpoint using the application’s available settings and documentation. Check whether browsing, retrieval or other integrations are enabled. This publication has not independently audited every application or service; the checklist is a method for identifying questions that a particular deployment must answer.
Consider storage as well as transmission
Even an offline inference process can leave material in files, conversation histories or backups on the machine. Decide who can access that computer, what should be retained and how a project is cleaned up. These decisions are separate from the model’s willingness to answer a request.
For a team, write down the permitted workflow rather than relying on an informal promise that everyone uses local AI. A useful record names the runtime, model revision, enabled tools and handling of documents. If a step changes, revisit the record. Control is valuable when it is observable and repeatable, not merely asserted by a label.
