Skip to content
All work
Multi-agent · Ops2024

Multi-Agent Ops Swarm for E-commerce Back Office

A coordinated swarm of specialised agents that triages refunds, chargebacks, delivery escalations and wholesale enquiries — with humans only on edge cases.

SUPMULTI-AGENT · 68% AUTO-RESOLVED
Multi-agent · Ops
Multi-Agent Ops Swarm for E-commerce Back Office
68%
Auto-resolved tickets
$220k
Annual ops savings
+7 pts
CSAT delta (BFCM)
9 / 9
Workflows automated
The problem

A 40-person ops team was drowning in 14,000 tickets a week; CSAT was slipping and holiday season was six weeks away.

Previous automations were rule-based spaghetti. Each new SKU added ten new edge cases.

Leadership wanted measurable $ saved per month and a clear audit trail before trusting anything agentic.

The approach
  1. Mapped 40 workflows, collapsed them into 9 archetypes, and built a specialist agent per archetype on Temporal.

  2. Introduced a supervisor agent with hard budgets: tokens, tool calls and reversible actions per ticket.

  3. Every write action requires a second-model critic plus a reversible receipt stored in Postgres.

  4. Built a shadow-mode rollout — agents answered for 30 days with a human gate — then flipped thresholds one workflow at a time.

Architecture
Rendering architecture diagram…
Outcome
  • 68% of tickets fully auto-resolved at steady state.

  • $220k/yr in fully-loaded ops savings at current volume.

  • CSAT up 7 points during Black Friday week, the first time in three years.

  • Audit trail passed a Big Four controls review on first pass.

What I owned end-to-end
  • Agent topology and Temporal workflows
  • Reversible action pattern and critic model
  • Shadow-mode rollout plan and success metrics
  • Exec-facing reporting: $ saved, CSAT, escalation rate
Next up

AI Outbound Engine with Voice + Email Orchestration