What's included
Fleet-wide answers
Ask once and the assistant answers across your whole book — security posture, policy drift, alert activity, admin changes — streamed as it works. Answers are advisory: reading the fleet never changes it.
Executive summaries on reports
Reports can open with an AI-written executive summary: the plain-English paragraph a customer actually reads, generated from the same data the report charts.
Drafted fixes, never direct action
When the assistant spots something fixable, it drafts a proposal from a fixed registry of known actions — it cannot invent an operation, and every draft is re-validated server-side, fail-closed. The model proposes; only an operator executes.
Dry-run first, then a confirmation ladder
Every proposed action goes through the same dry-run preview as a manual bulk change, and confirmation scales with risk — routine actions ask you to confirm the affected count, unusual ones get a full interstitial. Content an action carries, like a signature, is typed by the operator — never generated into place.
Audit-stamped, tamper-evident
Anything executed from an AI draft is marked ai-assisted in the audit log, with the prompt that produced it recorded alongside. The log itself is hash-chained, so the record of what the AI touched cannot be quietly edited afterwards.
Redaction before every call
Mailbox and file contents are stripped before anything is sent to the model — it reasons over settings, posture, and metadata, not your customers' email. The guard sits on every AI code path and is tested to fail if raw content ever reaches a payload.
Off until you say otherwise
The AI master switch ships off — nothing model-facing runs until an operator enables it. Postures are then set per action and per engineer, and they only ever tighten: a permissive default can be restricted anywhere down the chain, never loosened.
Gemini policy oversight
Separately from CrossTenant's own assistant, the AI & Gemini page shows each customer's Google-side Gemini and AI settings across the fleet — who has what enabled, tenant by tenant. That is Google's AI, surfaced; the governed assistant above is CrossTenant's.
AI features are optional and stay off until an operator enables them. The assistant is built on Claude, Anthropic's language model — the redaction, confirmation, and audit rules above apply to every call regardless.
AI, governed
Intelligence without handing over the keys
Turn the assistant on when you are ready — it ships off, and every action it drafts still goes through you.