SECURITY & PRIVACY

Working independently. Staying accountable.

An assistant that does things needs more safeguards than one that only writes text. Here’s how we handle that.

Concretely and verifiably.

Approvals by risk

Every action NEVYLO can perform is described in a registry: with inputs and outputs, risk class, required permission, cost estimate and reversibility. Low risk runs automatically depending on your package, medium shows a preview, high always needs your approval, with What, Why, Where, Data, Cost, Undo and Who.

No model output is executed unchecked. Whatever the AI proposes is checked against your permissions, budgets and approval rules before anything happens.

Audit log and undo

Every step of every task is logged: trigger, time, affected systems, data used, result. Anything technically reversible has an undo button; anything irreversible is never executed without approval. A kill switch stops running tasks immediately.

Strict tenant isolation

Every business is isolated at the database level: every access goes through a layer that enforces ownership; even programmatically, a query cannot accidentally end up in someone else’s data. This isolation is protected by automated tests.

Credentials and integrations

Connections to Google (soon), Meta, Microsoft and others run via OAuth. We never store your password. Access tokens are stored encrypted (AES-256-GCM), refreshed regularly and can be revoked at any time. Without a connection, an integration visibly runs in sandbox mode; nothing is shown as connected that isn’t.

No secrets in the client, no invented credentials, no fetching of URLs without protection against internal networks.

AI with a sense of proportion

AI providers receive only the data needed for the specific job. Personal data is minimised. Your data is not used to train third-party models. The model router picks a tier per task (economy to premium) and measures cost per business, with budget limits.

Privacy and operations

Data minimisation, purpose limitation, export and deletion options are built into the architecture; invoice data is subject to statutory retention periods. Access is role-based (owner, admin, staff, agency, support with justified, logged impersonation). Passwords are hashed with argon2id, sessions are stored as hashes only.

Transport encryption for all connections, encryption of data at rest at the infrastructure level, EU data residency as the target architecture, separate environments, error monitoring without personal content.

This page describes the platform’s security architecture. It is not legal or tax advice; the binding rules are in the privacy policy and terms of service (currently marked as drafts).

BUILT IN VIENNA
SERVERS IN THE EU
7 LANGUAGES
GDPR COMPLIANT
START FOR FREE

Connected with:Anthropic

Your business. One chat. One system.

Test it for free, no payment details.

Set up your business profile, connect your systems, write your first command.

Start for free Security & privacy