Customer support · Production

Minimal builds a multi-agent e-commerce customer support system with LangGraph and LangSmith

The problem

E-commerce businesses face repetitive and complex customer service workflows, including difficult T2 and T3 issues, that require consistent and accurate replies at scale while support teams lack intelligent tooling for autonomous resolution.

First attempt

A monolithic language model prompt conflated multiple tasks, leading to errors and expensive usage, which prompted Minimal to adopt a multi-agent architecture.

Workflow diagram · grounded in source
1
Incoming ticket triggers AI
trigger
“When a user turns the AI on, the system begins generating accurate, context-rich replies to incoming tickets”
2
Planner Agent decomposes query
ai_action
“Planner Agent: Breaks each incoming query into sub-problems (e.g., "Return Policy" vs. "Troubleshooting Front-End Issues"). Communicates with specialized research agents that perform retrieval and re-ranking of relevant documentation or …”
3
Research Agents retrieve knowledge
ai_action
“Research Agents: Handle each sub-problem by scouring the training center's knowledge base—like returns guidelines or shipping rules. Aggregate relevant information for the Planner Agent”
4
Tool-Calling Agent executes actions
ai_action
“Tool-Calling Agent: Receives the final "tool plan" from the Planner Agent. Executes decisive actions, such as refunding an order via Shopify or updating address records. Consolidates logs in one place for post-processing and chain-of-tho…”
5
Draft reply produced
output
“the system's final step is to produce a carefully reasoned draft reply to the customer—one that references the correct protocol, checks relevant data, and ensures compliance with the business's rules around refunds or returns”
6
Human escalation for edge cases
human_review
“escalating only 10% to human agents”
7
LangSmith iterative refinement
feedback_loop
“Whenever they found an error—say, a policy misunderstanding or a missing step—they created new tests in LangSmith's trace logs, added more few-shot examples, or further split a sub-problem”
Reported outcome

Minimal's AI agents deliver 80%+ efficiency gains across e-commerce stores while improving customer satisfaction, with the company projecting that 90% of support tickets will be handled autonomously and only 10% escalated to human agents.

Reported metrics
Efficiency gains80%+
Support tickets handled autonomously (projected)90%
Tickets escalated to human agents (projected)10%
Customer satisfactionimproving customer satisfaction
Reported stack
LangChainLangGraphLangSmithShopifyMonta Warehouse Management ServicesFirmhouseZendeskFrontGorgias
Source
https://blog.langchain.dev/how-minimal-built-a-multi-agent-customer-support-system-with-langgraph-langsmith/
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

Minimal's AI agents deliver 80%+ efficiency gains across e-commerce stores while improving customer satisfaction, with the company projecting that 90% of support tickets will be handled autonomously and only 10% escal…

What tools did this team use?

LangChain, LangGraph, LangSmith, Shopify, Monta Warehouse Management Services, Firmhouse, Zendesk, Front, Gorgias.

What results were reported?

Efficiency gains: 80%+; Support tickets handled autonomously (projected): 90%; Tickets escalated to human agents (projected): 10%; Customer satisfaction: improving customer satisfaction (source-reported, not independently verified).

What failed first in this deployment?

A monolithic language model prompt conflated multiple tasks, leading to errors and expensive usage, which prompted Minimal to adopt a multi-agent architecture.

How is this customer support AI workflow structured?

Incoming ticket triggers AI → Planner Agent decomposes query → Research Agents retrieve knowledge → Tool-Calling Agent executes actions → Draft reply produced → Human escalation for edge cases → LangSmith iterative refinement.

Related customer support cases