The right question is not whether AI can answer everything

A useful AI assistant is not one that confidently responds to every request. It is one that recognises the limits of its information and authority. When a request is routine, factual and low-risk, automation may save time. When the request involves judgement, uncertainty or significant consequences, a person should take responsibility.

The escalation threshold defines the boundary between autonomous handling, preparation for human review and actions that require an authorised person from the outset. These boundaries should be written into the workflow and enforced by the platform, not left to a prompt alone.

Three practical zones

Green-zone work is repeatable and low-risk: sharing verified opening hours, collecting a callback request or creating an internal task. The system still needs a trusted source and a record of what it did.

Yellow-zone work can be prepared by AI but needs human judgement: a custom quote, a complaint, a disputed invoice or an unusual request. The handoff should include the customer’s context, facts collected, uncertainty and a suggested next step.

Red-zone work should not be executed autonomously: legal commitments, medical advice, safety decisions, financial transfers, deletion of important records or promises that bind the business. The system should explain the boundary and route the request appropriately.

  • Define allowed actions and prohibited actions explicitly.
  • Require verified source data for factual answers.
  • Use risk and consequence, not model confidence alone, to set escalation.
  • Make human override and emergency stop available.
  • Record the authority, execution and verification evidence for meaningful actions.

Confidence is not authority

A model may be confident and still be wrong, or may correctly understand a request but lack permission to act. Confidence scoring can inform a review, but it must not grant authority. The system should check whether the information is grounded, whether the action is permitted and whether the intended result can be independently verified.

UnitAxon’s design principle is people-led, AI-assisted operations. Routine tasks may eventually run within standing mandates, while consequential or out-of-policy decisions remain subject to explicit approval and audit.