01

Keep commitment human

Prices, proposals, payments, availability promises, contracts and public claims should remain under explicit human authority unless a separately approved rule is genuinely deterministic.

02

Keep sensitive decisions human

Safety, complaints, legal issues, health information, employment decisions and other high-risk matters require appropriate people and professional judgement.

03

Design review as a workflow

Show the reviewer the source, draft, reason for escalation and permitted actions. Approval should be an accountable decision, not a vague instruction to “check the AI”.

04

Use corrections as evidence

Record what changed and why. Corrections should improve knowledge, rules or routing—not disappear into an unmeasured inbox.

Technical evaluator notes

Evaluation should confirm architecture, data handling, knowledge ownership, permissions, human authority, QA/UAT criteria, integration boundaries, monitoring, audit evidence, release gates and rollback for the specific workflow. The guide does not claim that external systems are connected.

Find where AI could helpBack to resources