Use cases › Customer service & contact center
LendingTree: how the agent works and how to build one like it
Summary: LendingTree built a multi-agent mortgage assistant using a supervisor agent that delegates to worker agents, built with LangGraph and MCP. The source reports over 97% of conversations are handled without escalation.
Source: https://aws.amazon.com/blogs/machine-learning/how-lendingtree-built-a-multi-agent-mortgage-assistant-on-amazon-bedrock/ source read via search extract
Why this use case works
The supervisor-and-workers design keeps each worker narrow (rates, eligibility, application status, education), which makes them easier to test and lets the supervisor focus on routing. MCP gives the workers a standard way to reach internal tools instead of one-off integrations. (Our analysis.)
Agent pattern: Support chat agent
A support chat agent: a retrieval layer over help-center and policy content, tool calls into the systems of record (billing, orders, accounts), and explicit escalation rules. The quality of the knowledge base matters more than the model choice.
How a smaller team could replicate it
Teams with several distinct question types can start with a router plus two or three specialist agents in LangGraph or a similar framework, each with its own tools and test set, and add workers only when a question category is large enough to justify one. (Our analysis, not a claim made by the source.)
Want a rough business case first? Try the AI agent ROI calculator, or ask us to scope it.
Risks and guardrails
Risks: a supervisor that routes wrongly sends confident but irrelevant answers, and financial-rate questions carry disclosure obligations. Guardrails: per-worker test sets, logged routing decisions, and approved compliance language for rate and eligibility answers. (Our analysis.)
What to measure
Escalation rate, routing accuracy (did the supervisor pick the right worker), answer correctness on a labeled test set, and conversion from conversation to application.
Related Customer service & contact center deployments
- Travelers: Travelers: a fully autonomous voice claims agent handles auto first-notice-of-loss. 85–90% of users complete their claim through it.
- Maccabi Healthcare: Maccabi Healthcare: voice/chat agents handle info requests, emergency guidance and appointment booking for 2M+ users.
- ENGIE: ENGIE: an Agentforce agent answers billing/EV questions and switches on smart charging through the SAP API.
- Commerzbank "Bene": Commerzbank "Bene": a chatbot that handles 2M+ chats and resolves 70% of them.
- Amdocs: Amdocs: a telco CX agent resolves complex issues end-to-end.
Part of the AgentStack Atlas catalog of 52 sourced agentic AI deployments. Spotted an error? Email agentstackatlas@mail.grokbot.com.