Every email, expected vs actual
Illustrative assumptions — not real company data. Defaults come from data/model.json; the same function runs in CI.
Assumptions
What moves the 3-year NPV most (±25%)
Trust is earned in steps
- 0 · Shadow
Runs on a copy of the mailbox; outcomes only in the log. Gate: agrees with agents on ≥ 90% of in-scope emails.
- 1 · Draft-only
Every reply is a draft with a Teams card. Gate: ≥ 95% sent unedited, no data-protection rejections, DPO sign-off.
- 2 · Graduated
Automatic for plain status and POD to contacts. Weekly review of 20 automatic replies.
- 3 · Expand
Phase 2: slot changes. Separate design and business case.
Going down is one setting, and automatic after a data-protection incident, a harmful reply in the weekly sample or a failed evaluation after a model update.
KPIs
| KPI | Baseline | Target week 12 |
|---|---|---|
| Median first response, status/POD | 2.9 h | < 15 min auto, < 2 h draft |
| French-language first response | 5.2 h | equal to NL/EN |
| In-scope emails answered automatically | 0% | 55% |
| Drafts sent unedited | — | ≥ 85% |
| Repeat emails | 12% | < 5% |
| Data-protection incidents | unknown | 0 |
What this demo is not
- The reading step is a deterministic stand-in with the same JSON contract as the Copilot Studio agent. A real model must pass its own evaluation first.
- The Copilot Studio agent and the Power Automate flow are a design and a build guide, not exports from a tenant.
- No real emails, customers, pilot or results. Every figure is an illustrative assumption.