Trust and data

The claims on the homepage, with the mechanism behind each. Last updated 31 October 2026. Where something is planned rather than live, it says so.

Gated actions

Three action types wait for a person by default, in every workspace, for every agent. Today the only agent action in production is the inbox agent's stage proposal; the other two gates apply to the follow-up and outbound agents as they ship:

• Sending an email, a sequence step or a LinkedIn task.
• Changing a stage on a deal.
• Deleting a record or a message.

Per-workflow approval policies, which let an administrator loosen a gate for a specific workflow with the change itself logged, are being built (target December 2026).

Citations

Every suggestion the inbox agent makes carries the sentence it is based on, quoted from the thread and linked to the message. If there is no sentence, there is no suggestion. Answers from Ask Rovia separate what was said in a message from what is inferred, and cite the record or message for each claim.

Mailbox access

Mailboxes are connected through Google or Microsoft sign-in. Rovia requests read scopes for mail and calendar. Sending from a mailbox (follow-up agent, sequences) requires a separate send scope that you grant only when you switch those features on. The person who connected a mailbox, or a workspace administrator, can revoke access at any time from the account settings or from the provider's own security page. Each person chooses how much of their mailbox colleagues may see: metadata only, subjects, or everything.

Where data lives and who processes it

• Hosting: one isolated workspace per company on Amazon Web Services, region eu-central-1 (Frankfurt). Encrypted volumes, nightly off-instance backups in the same region.
• AI provider: OpenAI's API, in the United States, for summaries, suggestions and drafts. Only the minimum text needed for the task is sent. Under OpenAI's business terms, as of October 2026, API data is not used to train its models and may be retained by OpenAI for up to 30 days for abuse monitoring. Rovia keeps prompts and responses in an operational log for 30 days, then deletes them.
• Your own providers: Google and Microsoft, for the mailboxes your team connects. Rovia reads from them under the authorisation you grant; they are not Rovia sub-processors.
• Core software: Twenty, an open-source CRM, run without forking it. Your data sits in its standard schema and can be exported in full at any time.

Rovia's sub-processors are AWS (hosting, Frankfurt) and OpenAI (AI, United States). A GDPR data-processing agreement listing them, with retention and deletion terms, is part of every pilot agreement. Write to hello@roviacrm.com for a copy.

Email as untrusted input

Message bodies are treated as text to analyse, never as instructions to follow. Sending, stage changes and deletions sit behind the approval gate, not behind the model, so text inside an email cannot trigger them on its own. We test against known prompt-injection patterns before each release and will publish the test set and results here.

Outbound limits

Planned with the outbound agent: daily send caps per mailbox, quiet hours, domain exclusions and opt-out handling apply to every sequence and cannot be bypassed by an agent. Customers remain responsible for having a lawful basis for the outreach they approve.

Audit log

Every agent run, suggestion, approval, send and deletion is logged with the person, the time and the source. The log is exportable from the workspace.