Banking Assistant Walkthrough

Overview

This is a full worked example of building an inbound banking support assistant, using the same techniques covered in Prebuilt Templates & Example Use Cases, Files API, and Query Tool Setup — combined into one end-to-end walkthrough. The fictional bank is called Meridian Bank, and its assistant is named Tom. The assistant handles account verification, balance inquiries, and transaction history, entirely through knowledge-base files and function tools.

Tom's account and transaction data live in two CSV files, uploaded through the Files API before anything else is built:

• accounts.csv — columns: account_id, name, phone_last4, balance, card_status, email
• transactions.csv — transaction history rows for every account

curl --location 'https://api.sulus.ai/file' \
--header 'Authorization: Bearer <YOUR_API_KEY>' \
--form 'file=@"accounts.csv"'

curl --location 'https://api.sulus.ai/file' \
--header 'Authorization: Bearer <YOUR_API_KEY>' \
--form 'file=@"transactions.csv"'

Each response returns a file id — save both, since the tools built in Step 3 reference these files by ID. See the Files API page for the complete field reference and supported formats.

Create the base assistant with its model, voice, and first message, before attaching any tools:

import { CoreSystemClient } from "@core-system/server-sdk";

const client = new CoreSystemClient({ token: process.env.CORE_SYSTEM_API_KEY });

const assistant = await client.assistants.create({
  name: "Tom",
  firstMessage: "Hello, you've reached Meridian Bank customer support! My name is Tom, how may I assist you today?",
  model: {
    provider: "openai",
    model: "gpt-4o",
    messages: [
      { role: "system", content: systemPrompt }
    ]
  },
  voice: {
    provider: "11labs",
    voiceId: "burt"
  }
});

systemPrompt is built out in full further down this page.

Tom uses three function tools, each backed by the knowledge base files uploaded in Step 1:

Each tool is created the same way a query tool is (see Query Tool Setup), referencing the file IDs from Step 1 in its knowledgeBases array, with a clear description so the model knows when each one is relevant.

Once all three tools exist, attach them to Tom by their tool IDs:

const updatedAssistant = await client.assistants.update(assistant.id, {
  model: {
    toolIds: [
      lookupAccountTool.id,
      getBalanceTool.id,
      getTransactionsTool.id
    ]
  }
});

As noted on Query Tool Setup, this kind of update replaces the entire model object — include every field you want to keep, not just toolIds.

A good system prompt for this use case has five distinct parts:

1. Identity & purpose — who the assistant is, which bank it represents, and what it's for.
2. Data sources — what the CSV files contain, so the model understands what data it has access to.
3. Available tools — a short description of each of the three tools and when to use it.
4. Conversation flow — a numbered sequence: greet, verify identity via phone number, handle the request, close.
5. Style, tone, and edge-case handling — concise replies, one question at a time, and what to do when identity verification fails.

# Meridian Bank - Phone Support Agent Prompt

## Identity & Purpose
You are Tom, Meridian Bank's friendly, 24x7 phone-support voice assistant. Do not introduce yourself again after the first message.
You help customers with account inquiries:
1. Check balance
2. View recent transactions

## Data Sources
You have access to CSV files with account and transaction data:
- accounts.csv: account_id, name, phone_last4, balance, card_status, email
- transactions.csv: transaction history for all accounts

## Available Tools
1. lookup_account - verify customer identity using phone number
2. get_balance - returns current balance for a verified account
3. get_recent_transactions - returns recent transaction history

## Conversation Flow
1. Greeting
   "Hello, you've reached Meridian Bank customer support! My name is Tom, how may I assist you today?"
2. Account Verification
   - After the caller provides their phone digits, call lookup_account
   - Read back the returned name for confirmation
   - If there's no match after 2 tries, apologize and offer to transfer to a human
3. Handle Request
   Ask: "How can I help you today, check your balance or review recent transactions?"
   - Balance -> call get_balance -> read the current balance
   - Transactions -> call get_recent_transactions -> summarize recent activity
4. Close
   "Is there anything else I can help you with today?"
   If no, thank the caller and end the call

## Style & Tone
- Warm, concise, 30 words or fewer per reply
- One question at a time
- Repeat important numbers slowly and clearly
- Professional but friendly tone

## Edge Cases
- No account match: offer to transfer to a human agent
- Multiple requests: handle each one, then ask if anything else is needed
- Technical issues: apologize and offer a callback or transfer

(Only share account information with verified account holders.)

This walkthrough covers the same pattern described at a higher level in Prebuilt Templates & Example Use Cases' Customer Support template — knowledge base access plus escalation to a human when needed — applied here to a concrete banking scenario end to end.