How to build a Banking Helpdesk

Workflow Overview

This flow is designed to help financial advisors and private bankers quickly answer questions about clients and bank products by combining CRM data and internal documentation into a single, professional response.

Node-by-Node Breakdown

1. Question (Input Node)

• Purpose:
  This is where the user (advisor or banker) enters their question.
  Example: “What’s the latest with client C001’s mortgage application?” or “What are the eligibility criteria for the new savings account?”

2. CRM Retriever (LLM Node)

• Purpose:
  Uses an advanced language model (Anthropic Claude Sonnet 4) to search Salesforce CRM data and summarize relevant client information.

• How it works:

  • Receives the user’s question.

  • Searches for client details, recent activities, account status, and notes.

  • If information is missing, it politely informs the user and suggests next steps.

  • Strictly follows privacy and compliance guidelines.

3. Product and Services Retriever (LLM Node)

• Purpose:
  Uses the same advanced language model to answer questions about the bank’s products, services, eligibility, and procedures using internal documentation (like Confluence or product sheets).

• How it works:

  • Receives the user’s question.

  • Searches for product features, eligibility, procedures, and rates.

  • If information is missing, it suggests where to find more details.

  • Also follows compliance and privacy rules.

4. Formatter (LLM Node)

• Purpose:
  Uses OpenAI’s GPT-4 to combine the outputs from the CRM Retriever and Product and Services Retriever into a single, clear, and actionable answer.

• How it works:

  • Takes the client information and product/service information.

  • Formats them into a professional, concise response.

  • If both types of information are available, presents them in separate sections.

  • If only one is available, notes what is missing.

5. Answer (Output Node)

• Purpose:
  Displays the final, formatted answer to the user.

• What the user sees:
  A clear, professional summary that combines client details and product/service information, ready for use in client meetings or communications.

How Data Flows

1. User enters a question in the Question node.

2. The question is sent to both the CRM Retriever and Product and Services Retriever nodes.

3. Each LLM node searches its respective knowledge base (CRM or internal docs) and generates a summary.

4. Both summaries are sent to the Formatter node, which merges and polishes the information.

5. The Answer node presents the final, user-ready response.

Key Features

• Dual Retrieval: Answers both client-specific and product-specific questions in one go.

• Compliance: Built-in privacy and compliance guardrails.

• Professional Formatting: Ensures responses are clear, concise, and ready to use.

• Fallbacks: If information is missing, the system gracefully informs the user and suggests next steps.

In summary:
This workflow streamlines the process of answering complex client and product questions by automatically searching both CRM and internal documentation, then merging the results into a single, polished answer for the advisor or banker.