Responsible automation

New clients and new capabilities start in shadow mode: the system records what it would do and what it expects to happen, and does nothing else. Live execution is enabled per capability only after test gates pass and the client approves in writing.

Green actions may run autonomously once live. Yellow actions produce a recommendation that must pass a second validation before anything happens. Red actions always require explicit human approval.

Always red: banking or payment-detail changes, refunds, contract acceptance or signature, legal admissions, tax or accounting representations, security incidents and destructive data deletion.

Deterministic checks run before any model is consulted: tenant match, recipient and domain validation, duplicate prevention, amount thresholds, required fields, date validation, allowed action type (unknown actions fail closed), opt-out status, explicit recipient allow-lists, per-hour rate limits and prompt-injection screening. The model-backed evaluator has passed synthetic and real-agent shadow gates. Durable runtime and privacy-minimized telemetry are active as platform infrastructure, while real client provider data remains unconnected and the real-data shadow gate remains closed.

The system must never invent testimonials, customers, projects, outcomes, certifications or case studies. Public content with an unsupported claim or guarantee language is blocked before review.

Every autonomous decision writes an audit event recording the actor, tenant, time, inputs, decision, policy rule, confidence, intended side effect, outcome and idempotency key. Clients can read their own audit log at any time.