Three Copilot agents, built end to end
One method, three levels of ambition: a helper that answers from a single document, a policy agent for the whole organisation, and an assistant that drafts replies and flags what must be escalated.
Returns & Warranty Helper
Decides covered or not, from one source, in a three-line verdict.
HR Policy Agent
Answers from several policy sources, cites the section, and redirects what it should not answer.
CS Email Assistant
Drafts the reply in the house template, flags escalations and leaves unknowns in brackets.
Agents draft. People decide.
Version one drafts. It does not send. Escalation limits live in one shared place, so a threshold changes once and every agent updates.
Scored on five checks before it ships
Bars show the launch bar, not a result.
Test twenty cases, not two
- Core cases
- Edge cases
- Missing information
- Must refuse
- Escalation triggers
- Adversarial
Write the expected answer for each one first. Otherwise you grade on vibes and pass everything.




