Meridian Health: L1 agent that must not invent clinical advice

You're on a call with Priya, VP CX at Meridian Health, a 14-hospital regional system. They run ~38k chat/email tickets per month on Zendesk, plus a separate voice IVR that still dumps to a 90-person BPO. First-response SLA is 15 minutes for chat; CSAT target is 4.4/5. Average handle time is 11 minutes. Top intents they think they know: appointment reschedule (~22%), billing explanation (~18%), portal password resets (~12%), referral status (~9%), and a long tail of clinical-adjacent questions ('is this medication covered,' 'what does this lab mean'). Priya wants an AI agent live on web chat in 10 weeks that 'deflects 40% of L1' so she can cut BPO spend before the next board meeting. Marketing already announced 'AI Care Concierge' on the patient portal homepage. The CMO is nervous but not in the room. Ops claims macros and FAQs are 'mostly clean'; Legal says anything clinical must escalate. IT will only allow read-only FHIR access to Epic for demographics/appointments in v1, and the Zendesk sandbox has three years of tickets with PHI still in free-text notes. Her success metric is 'containment rate.' She does not have a labeled intent taxonomy, only 40 sticky-note categories from a workshop. Unionized hospital front-desk staff are already circulating a petition that AI will 'give medical advice.' Budget is approved for licenses and one internal engineer at 50% for the pilot. She asks: how would you scope the pilot, what would you build, and when can we go live?

This is the round that fails the most FDE candidates. Start the clock, decompose it like you're on the call, then get graded by the coach against the hidden rubric top labs use.