Control is part of the product
Agents that act need limits you can see, change and prove. These are the controls in the October 2026 release.
Risk policy on every action
The same policy engine in every service, with shared test cases. Strategy agents are read-only; legal agents never write to a counterparty without a lawyer.
Human oversight
One approval inbox for messages, offers, workflow steps and agent actions, routed to the people of the owning function.
High-risk uses contained
HR agents are held disabled by a database rule, and the HR pack is assist-only.
Disclosure and consent
An AI disclosure on e-mail, PEC and voice. Recipient opt-in and opt-out on WhatsApp, SMS and voice that imports cannot override.
Versioned specifications
History, differences, dry runs and rollback for every agent.
Explanations
What triggered an action, what the agent did, why, and which rule allowed it.
Security and privacy, stated plainly
In place
- Tenant isolation by row-level security on every tenant table.
- Operayde sign-in, with multi-factor authentication required for changes.
- Encrypted credentials and a per-tenant daily AI budget with a hard stop.
- Erasure across business records, with suppression so erased contacts are not re-imported.
- Prompt and response logs deleted after 30 days by default.
Not yet
- Agents runs on a single host without high availability.
- No independent audit or penetration test of Agents yet.
- Per-service tokens, a content security policy and customer-managed keys are being rolled out.
The evidence pack and draft data processing terms are shared under NDA.
Pronto a gestire l'IA come infrastruttura che controlli?
Dedica 20 minuti a un nostro deployment engineer. Mappiamo i tuoi workload sul confine di controllo, confermiamo l'offerta e mettiamo per iscritto il piano di deployment.