EdithBack to Security
AI guardrails

AI assists; controls decide.

The AI layer should explain and draft, but it should not post, file, close, reverse, or expose secrets.

Allowed help

AI should make work faster without owning the system of record.

Explain variance
Draft reminders
Suggest matches
Summarize documents
Map imports
Extract bill details

Disallowed authority

Material actions stay deterministic and human-approved.

No auto-posting
No auto-filing
No auto-close
No secret access
No broad tenant reads
No silent external action

Traceability

Answers should show context and sources.

Context chips
Used sources
Prompt version
Tool log
Approval drawer
Audit event