EleganceAI.ai Lab

Procurement Exception Triage Agent

An enterprise agentic AI pilot: Claude Haiku investigates delayed POs, blocked invoices, and vendor risk through governed MCP tools — with deterministic hooks, structured errors, and human-review escalation you can watch live.

Demo boundary: mock procurement data only. The agent cannot approve payments, release invoices, change purchase orders, or touch production systems — and the demo shows why that matters.

Guided scenarios — pick one, then run

Agent Answer

Run a triage request to see the final answer.

Executive Summary

Run a triage request to generate the executive summary.

Exception Value Reviewed

$0

Delayed POs

0

Blocked Invoices

0

Follow-ups Created

0

Escalations Created

0

Human Approval

Not required

Why this scenario matters — End-to-end demo

One message, five concerns. The agent decomposes the request, runs the loop across both MCP servers, hits the governance threshold, and synthesizes a single unified answer — the pattern behind every credible enterprise agent pilot.

  • Multiple tool_use turns before the final end_turn stop reason.
  • Read-only lookups route to procurement-data; actions route to procurement-workflow.
  • The $42,000 items trigger escalation while the follow-up for smaller items proceeds.

Claude Architect exam concept: Agentic loop and multi-concern decomposition: continue on stop_reason=tool_use, finish on end_turn, synthesize a unified response.

Agent Loop Timeline

Run a request to watch the loop unfold step by step: model turn → stop reason → tool request → hook decision → MCP route → final response.

Run Cost & Tokens

Model Turns

0

Input Tokens

0

Output Tokens

0

Est. Cost (Haiku)

$0 (mock)

A full multi-tool triage on Claude Haiku typically costs a fraction of a cent — the economics a CIO asks about before piloting.

Tool Result Data (raw)

Show the raw structured JSON the MCP tools returned (hidden by default — the agent answer above is the human-readable version)

MCP Tool Trace

No tools have run yet.

Audit Trail

Audit events will appear after a run. They record the user request, planner identity, hook decisions, MCP routing, and structured tool outcomes.