Build a Customer Support Agent with Tool Calling

Customer support costs trillions annually. Manual tickets drain engineering resources. AI agents handle routine queries at scale. You will build an agent that checks order status automatically. This reduces human ticket volume by 40% in production tests.

Prerequisites

You need Python 3.9 or higher. Get an OpenAI API key. Install dependencies: pip install openai. Budget $5 for testing.

Steps

• Define the Tool Function

Create a Python function that mimics a database lookup. Return a string status. Keep latency under 200ms.

``python

def get_order_status(order_id: str) -> str:

if not order_id.startswith("ORD"):

return "Invalid ID"

return "Shipped"

``

The agent needs schema definitions. Describe parameters clearly. Use JSON Schema. Specify types strictly. The model cannot guess formats. Invalid schemas cause silent failures.

• Configure the System Prompt

Tell the model its role. "You are a support agent for Store X." Restrict scope. Do not answer billing questions. Only handle order status. This prevents scope creep. If a user asks about refunds, redirect them to email. Clear boundaries reduce errors. Specific instructions beat general helpfulness.

• Bind Tools to the Model

Use the OpenAI Chat Completions API with tool calling. Pass the tool schema in the request. The model decides when to call it. Do not force calls. Let the model reason based on user input. Parse the tool arguments safely. Validate the order ID before querying. Handle API errors gracefully. Return user-friendly messages on failure.

• Run the Evaluation Loop

Test 10 common queries. Measure accuracy. Include edge cases like missing IDs. If the model hallucinates, tighten the prompt. Log every tool call. Debug failures immediately. Track token usage per query. Average cost should stay under $0.01 per ticket. High volume adds up quickly.

Common pitfalls

Infinite Loops: The agent might call the tool repeatedly without resolving. Set a max iteration limit of 5. Break the loop if the tool returns an error twice. Hallucinated Policies: The model might invent refund rules. Explicitly state "Do not discuss refunds" in the system prompt. Negative constraints work better than positive ones here. PII Leakage: Never log full customer addresses. Mask sensitive data before sending to the model. Compliance fines exceed development costs. Slow Latency: Tool calls add network time. Cache frequent responses. Aim for under 2 seconds total response time. Users abandon chats after 3 seconds.

Next step

Connect the mock function to your actual SQL database using a read-only user.