Knowledge · AI Capabilities

    AI Confirmation Requirements

    Which AI actions need explicit confirmation before running — destructive, financial, external-send, bulk and override — and how a confirmation gate should work.

    When should AI require confirmation before acting?

    Confirmation is required whenever an action is destructive, financial, sent outside the company, applied in bulk, or overrides a system rule. The AI returns a preview of exactly what will happen and executes only after a second, explicit approval tied to that specific preview.

    Key takeaways

    • Classify actions by consequence, then attach confirmation to the class.
    • A preview must state the real records and the real recipients.
    • Approval should be bound to the previewed action, not a general yes.
    • Confirmation fatigue is a design failure — gate consequence, not everything.
    • Both the preview and the approval belong in the audit trail.

    The five classes that always need a gate

    • Destructive: deleting, voiding, cancelling or archiving records.
    • Financial: charging, refunding, issuing credit or changing a price that has been agreed.
    • External send: any message, contract or document that leaves the company.
    • Bulk: anything applied to many records at once, where a mistake multiplies.
    • Override: booking past capacity, skipping a required stage, bypassing a rule.

    What a real confirmation looks like

    A confirmation that says 'proceed?' is decoration. A usable one shows the specific customer, the specific amount, the exact recipients, and what changes the moment approval is given.

    It should also be bound to that exact preview. If the underlying data changed between the preview and the approval, the gate must be re-issued rather than executed against stale facts.

    Why over-gating is its own risk

    If every read and every note requires approval, people stop reading the prompts and start clicking through them. The gate then protects nothing while still costing time.

    Reserve confirmation for consequence. Low-risk reads and internal notes should run immediately, and the audit trail carries the accountability instead.

    Confirmation is not the same as autonomy

    A confirmed action is still bounded: the user could have done it manually, the permission check still applies, and the entitlement limits still apply. Approval does not widen scope — it authorizes one specific instance of something already permitted.

    Where URBLD fits

    In URBLD, destructive, financial, override, bulk and external-send actions return a preview and a signed confirmation token first, and execute only on a second call carrying that token. Dispatch notifications and contract sends are separate confirmed actions, never implied side effects of another step.

    FAQ

    Frequently Asked Questions

    Straight answers about how URBLD runs the business end-to-end.

    More in AI Capabilities

    How to tell a chatbot from an assistant from operational AI.

    Browse AI Capabilities
    Share this page