Guardrails & consent

GrootDesk never invents an action. This is the safety model that lets support teams hand real work over with confidence.

Order verification

Before any order status is shared or any change is made, the customer's identity is verified against your store data. Unverified requests get a polite prompt, never private details.

Consent-gated actions

Returns, cancellations and refunds are only filed after the customer confirms on an explicit, on-screen card. Until they tap confirm, nothing changes on the order.

The confirmation card lists exactly what will happen — order, amount, and outcome — so consent is informed, and the record is kept.

Capability guardrails

The agent can only do what you allow. Anything outside its configured remit is refused clearly and escalated to a human — it never guesses or oversteps.

Evaluators

Automated evaluators score every answer before it's trusted:

  • Hallucination detection — checks the answer is supported by retrieved context
  • Language matching — reply language always matches the customer's
  • Conciseness scoring — keeps responses tight and on-point

Audit trail & escalation

Every message, decision and action is recorded and reviewable. When a human is genuinely needed, GrootDesk escalates into your helpdesk with the full context attached — no re-typing for the customer or the agent.

← Knowledge base & RAG Models & routing →