Private AI is a data-flow decision
VidSatAI helps Canadian businesses and professionals evaluate private knowledge assistants, local or self-hosted models and managed AI services. A product label or a private document collection cannot establish privacy on its own.
A useful assessment follows information through the complete system: document extraction, storage, search indexes, model requests, logs, integrations and actions.
Compare the architecture options
Managed AI services
An external provider processes model requests. Assess the actual service terms, account settings, retention, access and supporting components for your intended use.
Local or self-hosted inference
Model processing can run within an environment you control. This still leaves operational responsibilities: infrastructure security, updates, access boundaries, performance, logs and any external connectors.
Retrieval-based knowledge assistants
Retrieval can supply relevant passages from approved documents to a model. Document permissions, source references, missing-information behaviour and access-boundary tests remain important.
Source: “Section 4: retain working documents for the period specified in the approved retention schedule.”
Question: “How long should I keep working documents?”
“Section 4 refers to the approved retention schedule. The supplied document does not give a duration, so I cannot provide one.”
A useful assistant identifies its source and acknowledges what it does not know. This is a static illustration, not a live AI result or client deployment.
What an assessment can cover
- Data minimisation and the categories of information involved.
- Identity, document permissions and tool access.
- Retention, logs and external processing.
- Source support, failure cases and evaluation criteria.
- Human approval, escalation and ongoing ownership.
Training use and external processing are different
“Not used to train a model” does not mean “never sent to a provider.” Assess both the handling of requests and any secondary use separately. Security and privacy depend on the actual configuration and operation.
Independent guidance
Canada’s privacy commissioners provide privacy principles for generative AI. This resource is independent guidance, not an endorsement of VidSatAI. An architecture assessment does not replace an organisation’s professional or legal obligations.