Sheet 10 · Controls

Security and Governance

Controls designed for industrial and export-sensitive environments.

Public

Approved for anyone. Marketing copy, published documentation.

Internal

Everyday business content. Approved cloud AI platforms with tenant controls.

Confidential

Commercially sensitive. Isolated processing, strict need-to-know.

Restricted

Export-controlled, defense-related, or customer-controlled. Never sent to public consumer AI services.

Data classification drives everything. Four classes — Public, Internal, Confidential, Restricted — determine which platforms may touch which data. Restricted engineering, defense-related, export-controlled, or customer-controlled information is never sent to public consumer AI services.

Controls in this plan:

Approved-platform matrix and least-privilege access
Human approval on all published knowledge and pricing content
Audit logs for administrative and AI activity
Model and prompt versioning, with citations required on assistant answers
Server-side-only AI calls — the browser never talks to the model directly
Rate limiting, upload allowlists, and malware scanning where available
Incident response, vendor review, and periodic evaluation

Honest failure modes: when the assistant lacks an approved source, it says so and escalates to a human instead of guessing.