Microsoft Agent Academy · Human-AI Handoff Copilot · Demo Artifact

Agents should know when to hand off, what to preserve, and how humans can resume.

Human-AI Handoff Copilot packages a support workflow into a Copilot-style architecture: retrieve evidence, draft carefully, evaluate cost/risk, ask for approval, then preserve the handoff trail.

7handoff events
1human approval checkpoint
GraphMicrosoft product contract checked live
0final submission actions executed

Copilot Architecture Components

Conversational intakeSupport manager asks for help with a customer issue.
Policy retrievalApproved source material is fetched before drafting.
Drafting agentThe answer cites the retrieved policy evidence.
Approval cardHuman approves escalation-only model usage.
Microsoft Graph contractTeams, chats, Planner, drives, and sites are validated against the public Graph service root.

Handoff Flow

01

Intake

Support manager requested an agent that answers customer questions and escalates uncertain cases.

evt-0020
02

Planning

Agent created a workflow plan using retrieval, MCP tools, cost guardrails, and human escalation.

evt-0021
03

Evidence Collection

MCP retrieval returned the current refund policy and source URL for citation.

evt-0022
04

Execution

Agent drafted a customer answer grounded in the retrieved refund policy.

evt-0023
05

Risk Review

Cost guardrail warned that the high-quality model should be reserved for escalations.

evt-0024
06

Approval

Human approved the workflow with a rule that expensive model calls are escalation-only.

evt-0025
07

Handoff

Closed cloud-agent workflow case with source citations, cost guardrail, and human approval captured.

evt-0026

Human Approval Decision

evt-0025 · support-manager approved escalation-only model usage

Costly model calls are not automatic. They become a policy-controlled escalation path.