S SARAH LOCKETT

03 / AI case study / Shared-services operations

Designed · not yet built

The Shared-Inbox Triage Agent

A team mailbox where every message is classified, its commitments extracted, and a reply drafted for approval. The inbox becomes a work queue.

TYPEWorkflow automation
PLANNED BUILDPlanned · ~2½ weeks
STATUSDesigned · not yet built

PROVEN EXPERIENCE → NEW BUILD

Rooted in inboxes I have watched drown

I owned dealer-facing communication at Nimbus USA: statements, chasers, and coordination across time zones. A workflow-discovery engagement I led also found a single operator managing as many as 50–80 inbox reminders a day. This case study is the system I kept wishing existed.

THE BRIEF

An AI layer over a shared operations mailbox: classification, commitment and deadline extraction, duplicate-thread merging, SLA-risk flagging, and approval-gated reply drafts that turn an inbox into an accountable queue.

The problem

Shared mailboxes are where commitments go to hide. Requests arrive unstructured, deadlines live in prose, three people answer the same thread or nobody does, and the only tracking system is memory and flags.

The system

Agents classify each message, extract who-owes-what-by-when into structured records, merge duplicate threads, flag SLA risk before it breaches, and draft replies that a human approves or edits. Nothing sends on its own. The mailbox stays the front door; behind it sits an actual queue with owners and deadlines.

  1. Shared mailbox
  2. Classification
  3. Commitment extraction
  4. Dedupe + threading
  5. SLA watch
  6. Approval-gated drafts
Copilot StudioPower AutomateSynthetic mailboxDataverseApproval flows

DESIGNED OUTCOMES

What “working” will mean.

  • Commitments extracted from prose into a tracked queue
  • Duplicate threads merged before double-handling
  • SLA risk visible before the breach
  • Replies drafted, humans deciding

MEASUREMENT PLAN

How it gets judged.

  • Classification accuracy
  • Commitment extraction accuracy
  • Duplicate-merge precision
  • SLA saves vs. misses
  • Draft acceptance rate

THE BUILD PLAN

How this gets built
when it gets built.

Days 1–3

Corpus

Build a synthetic mailbox with realistic traffic: requests, chasers, duplicates, buried deadlines, mixed tone.

Days 4–8

Triage

Classification, extraction to Dataverse records, and duplicate-thread merging.

Days 9–12

Guardrails

SLA-risk flags, approval-gated reply drafting, and the never-auto-send controls.

Days 13–17

Evaluate

Extraction accuracy against seeded truth, queue views, and the written case study.

Transparency note. This is a self-directed AI case study built on synthetic data. Results will be reported as evaluation findings against seeded ground truth, not client outcomes. When the build is complete, this page gains: working triage flow, commitment queue, synthetic mailbox corpus, extraction evaluation set, operating guardrails note, written case study.

NEXT CASE STUDY / 04The Month-End Close Copilot
→