The trust questions
"Is Microsoft training on our data?" No — tenant content doesn't train foundation models (data-boundary concept, with the terms citation to keep). The better question is the next one.
"Then why did Copilot show Dave the salary file?" Because Dave could always open it — Copilot collapsed the discovery cost of your sharing debt (oversharing concept). This is the answer to rehearse, because it redirects the incident from 'AI breach' to the fixable truth.
"Can we see what people ask it?" Yes — interactions are compliance records: auditable, searchable, retainable (security module). Decide the review posture deliberately and TELL people; secret surveillance discovered later costs more than the tool.
The money questions
"Why is it $30 when Chat is free?" Work grounding — the paid seat reads your tenant's world; included Chat reads the web (products map). Buy seats where grounded work has value; use Chat + credit-billed agents everywhere else.
"Are we getting value?" Measured, not vibed: usage dashboards for raw adoption, but the real ROI evidence is task-level (meeting-recap hours, draft cycles) from the pilot cohort's own numbers (adoption concept, planned). Unused seats are the loudest Copilot fact in any renewal — recycle them monthly.
"What do agents cost?" Depends on species (taxonomy): seat-hosted declarative vs credit-metered custom engines vs Agent 365 governance overlay. The consumption report is the truth; set budget alerts before the first maker success story scales.
The disappointment questions
"It's down again?" AI reliability became office-availability (the June 2026 outages made the point industry-wide). Treat Copilot like any dependency: service-health watched, degraded-mode guidance written ('drafting by hand for the afternoon' is a fine playbook entry), status comms templated.
"Why are its answers mediocre?" Retrieval ceiling (foundations): sprawl in, confident mediocrity out — governance IS the quality programme. Plus prompt literacy: the gap between 'summarise this' and a role-context-format prompt is most of the perceived quality gap (adoption concept).
"Why can't it do X in Excel?" Per-app capability gaps are real and moving — keep the expectations table current (apps module) rather than relitigating each disappointment.
What to watch (proofs)
- Your own tenant's question stream: helpdesk tickets tagged Copilot, monthly — this page's local edition writes itself and pre-empts the next wave.
- Seat utilisation: assigned vs active from the usage reports — the renewal number.
- Incident comms readiness: the degraded-mode card exists and was used in the last outage (resilience discipline from the Teams curriculum, applied).
- The salary-file drill: quarterly red-team prompt for known-sensitive content — finding nothing IS the oversharing programme's proof.