All case studies
Case study 06AI SystemsAutomation

CRM Workflow Agents & MCP Connectors

LangGraph agents that research, draft, and update CRM records through permission-scoped MCP connectors.

Role
Senior engineer · agents and integrations
Core stack
LangGraphLangChainFastMCPRAGPython
Scoped
Each connector limited to the objects its workflow needs
Idempotent
CRM writes safe to retry without duplicates
Documented
Hand-offs with runbooks and evaluation datasets

Overview

Agents for client sales and service teams that gather context from documents and CRM history, draft updates and follow-ups, and write back only through scoped connectors, with a person approving anything that changes a record or contacts a customer.

Client teams wanted to automate repetitive CRM work without giving an AI unrestricted write access. Workflows had to be safe under retries, recoverable mid-run, and easy for the client's own engineers to own after hand-off.

Product walkthrough

The flow, screen by screen

Account copilot
01 / 02

Account copilot. The agent summarizes the account from CRM history and documents, then proposes field updates and a follow-up task that wait for approval.

ArchitectureCorePlatformAsync / control
regressioncontextproposed writesapprovedEvaluation datasetsper workflowDocument retrievalRAG · client filesCRM & inboxtriggers · requestsWorkflow agentLangGraph · checkpointsMCP connectorsscoped · idempotentCRM APIaccounts · tasksApproval gaterecord changes · sendsEmaildrafts · sends
A LangGraph workflow agent pulls context from documents, acts on the CRM only through scoped MCP connectors, and routes record changes and emails through an approval gate.

Approach

  1. 01Developed LangChain and LangGraph agents for document retrieval and CRM workflows such as account research, follow-ups, and record updates.
  2. 02Implemented MCP connectors with scoped permissions, so each workflow can read or write only the objects it needs.
  3. 03Put approval gates on record changes and outbound email, and made every write idempotent so retries never create duplicates.
  4. 04Checkpointed each run so a workflow can wait for approval or recover from a failure and continue where it stopped.
  5. 05Built evaluation datasets per workflow, tested failure recovery, and delivered documented hand-offs for client teams.

Key decisions

Least-privilege connectors

Scopes are declared per tool, so an agent that drafts follow-ups cannot change deal amounts or owners.

Automate reads, approve writes

Research and drafting run on their own; changes to records and messages to customers wait for a person.

Hand-off is part of the work

Runbooks, evaluation datasets, and recovery tests ship with each workflow so the client's engineers can run it.

Outcomes

  • Client teams kept control of every record change and outbound message.
  • Failure recovery was tested before each workflow was handed over.
  • LangGraph
  • LangChain
  • FastMCP
  • RAG
  • Python
  • FastAPI
  • PostgreSQL

Working on something similar?

I work with clients worldwide, across any time zone. Tell me what you're building.

Send a message Email meMore case studies