How to build an AI agent that checks order status and processes refunds

How to build an AI agent that checks order status and processes refunds

WISMO. It stands for “Where is my order,” and it’s the question support teams answer the most (and enjoy the least). Fortunately, high-volume, low judgment tasks like this are perfect for handing off to an AI agent. 

An existing API already has the answer, too. The customer wants it fast, and they don’t care how they get it.

Refund requests are usually part of the same conversation. Your customer checks on a late order, finds out it’s lost (or buys a substitute), and now wants to cancel the order and/or get their money back. The AI agent already has the order pulled up, and the refund mechanism is just a tool call away.

However, one bad prompt or one clever fraudster, and your AI agent is issuing refunds it shouldn’t. Woof.

Below, we show you how to build an AI agent (the right way) that can look up orders, issue refunds, and escalate specific conversations to a real person.

How to build your order-checking, refund-processing AI agent

Your AI agent will need access to certain tools and APIs to find orders and process refunds. Plus, it needs guardrails that control when it can act on its own. Here’s where to start:

• Tools: Functions the model can call to look up orders, check eligibility, start refunds, and update tickets.

• Identity: Confirmation that the person asking for a refund owns the order.

• Policy: Your real return and refund rules, available to the agent at decision time.

• Escalation: A path to a human when the request exceeds what the agent should approve alone.

• Record updates: Your agent needs to record the actions taken and the outcome.

1. Define the tools

A tool is a function the model can choose to call, with a name, description, and schema describing its inputs. The model decides when to call it, and your code decides what it does.

Your refund agent will typically need four tools:

1. Get order status: Takes an order number and returns status, carrier, and tracking details from your order management system.

2. Check refund eligibility: Compares the order against your rules (like purchase date and item condition) and returns whether a refund is allowed.

3. Initiate refund: Calls your payment processor or order system to issue the refund and returns a confirmation ID.

4. Update ticket: Writes the outcome to your help desk so the record matches what happened.

Twilio Agent Connect lets you define custom tools with the @function_tool decorator in Python or the defineTool function in TypeScript. Agent Connect follows the standard function-calling pattern LLM providers use and formats your tool definitions for whichever model you run.

Now, in Python, a custom tool can take the conversation session as a parameter, which gives it access to the customer's profile data. Your order lookup can scope results to the customer on the line, so the agent only ever sees that customer's orders.

Twilio's tutorial on connecting a Twilio agent to external APIs walks through tool calling on a live voice call, with versions in Python, Node.js, and PHP.

2. Verify the customer

You’ll want more than just caller ID and a matching order number. Your customers might get their phone numbers spoofed, and order numbers show up on shipping labels, emails, and screenshots. Your agent should be processing refunds on a bit more verification information. 

Add a verification tool the agent must call before any refund. Twilio Verify sends a one-time passcode to the phone number or email on file, and the agent asks the customer to read it back. It’s two-factor-authentication before a refund. 

No code, then no refund. Fair enough. 

Your system prompt sets the rule, and your refund tool should enforce it in code too. If the initiate-refund function checks for a successful verification on the session before running, a prompt injection can't talk its way around it.

Now, if you’re just looking up an order (and not processing a refund), you can usually skip this step. Reading a shipping date is generally low(er) risk, while moving money isn't.

3. Ground refund decisions in your policy

This is where AI hallucinations can (and will) hurt you, unless you do your due diligence upfront. If you ask a language model about a return window without giving it your policy, it'll answer confidently with something plausible. But simply plausible is how you end up refunding a 90-day-old order under a 30-day policy.

Yeah, that’s not going to cut it. And that might even be the least of your worries with basically no-rules enforcement. 

Twilio Agent Connect includes a built-in knowledge search tool that queries Enterprise Knowledge bases. Just load your return policy, refund exceptions, and shipping terms into a knowledge base, and the agent can retrieve the real rule before answering (and your system prompt should tell it to).

Keep the hard rules in your code, too. The eligibility tool should enforce return windows and item restrictions directly, so the knowledge base explains the policy to the customer while your code decides it.

4. Set approval limits

Not every refund should be automated. When it’s low-stakes, low judgement, then sure. But some refunds need a manual touch:

• High-value orders

• Repeat requests

• Outside of policy

• Upset customer

These are the times when a human needs to step in. Decide these limits before launch. You might consider something like this for your initial setup (and tweak as you learn):

• Auto-approve: Verified customer, eligible order, refund under your set threshold.

• Escalate: Refunds over the threshold, policy exceptions, multiple recent refunds on the account, or any request the agent can't confidently resolve.

Agent Connect's built-in handoff tool can handle your escalations. When the model calls it, Agent Connect packages the conversation ID, memory store ID, and the reason for escalating, then routes the conversation through a Twilio Studio flow to a human agent in Twilio Flex. The customer stays in the same call or chat, and with Twilio Conversation Intelligence configured, the human agent opens on an AI-generated summary, so your customer doesn’t have to repeat the scenario or the order number.

5. Update the ticket and the customer record

The final tool call in every refund flow should write the outcome back: 

• What was requested

• What was approved

• Refund confirmation ID

• Whether a human was involved

The customer's history matters, too. Twilio Conversation Memory extracts observations from each conversation and adds them to the customer's profile, reconciling new facts against existing ones. The next time that customer calls, the agent already knows about last week's refund. If it's their fourth refund request this month, you’ve got a problem, and that context informs the escalation decision.

One agent across voice, messaging, and chat

Your customers don’t have one universally agreed upon channel for communications. Some will want to reach out to you by phone call, while others might prefer text, WhatsApp, or web chat. Building a separate agent for each channel just isn’t efficient (or practical).

Fortunately, with Twilio, you can write the flow once and scale it across voice, SMS, WhatsApp, RCS, and chat.

Agent Connect routes every channel through a single message handler, so the same tools, verification step, and escalation rules apply everywhere. On phone calls, ConversationRelay handles speech-to-text, text-to-speech, and interruptions, with Twilio's internal benchmarks putting median latency under 0.5 second and the 95th percentile under 0.725 second. 

Your customers care about that speed, especially when they’re already irritated at a late order.

Pre-launch checklist

Run through this list before your AI agent handles a real refund:

• Can a prompt talk its way into a refund? Test with requests that skip verification, cite fake policies, or claim urgency. Your tools should refuse in code, whatever the model says. Give it your best take, and think of the most clever, ridiculous scenarios—you want your agent ready to defend the policy no matter what. 

• Does card data stay out of the conversation? Refunds should run through your payment processor by order or transaction ID. Card numbers should never pass through the agent.

• What happens when an API fails? Your order system will time out eventually. It happens. The agent should tell the customer clearly and offer a human, without guessing at an answer.

• Is every refund logged? Each one should land in your help desk with a confirmation ID.

• Are your thresholds right? Start conservative, review escalated cases weekly, and raise auto-approval limits once you trust the agent's judgment.

Start with order status before refunds

Order status is the safe first deployment. It's read-only, high-volume, and easy to measure. Once you know that’s working, you can double-check verification, policy grounding, and escalation are in place and tested.

Build the guardrails before you need them. Seriously, don’t wait until there’s an issue. Think ahead.

An agent that refunds the right orders and escalates the rest saves your team hours every week. An agent without them creates an inevitable fraud problem. It’s just a matter of when (not if).

Ready to build your own? Take a look at the top conversational AI platforms behind these powerful agents.

Frequently asked questions

What are the best services for building an AI agent that can look up order status and initiate refunds via API?

Twilio Agent Connect lets you define custom tools that call your order and payment APIs, with built-in tools for knowledge search and human escalation. It runs on your own infrastructure, so tools reach your systems with your own credentials.

What platforms let an AI support agent call APIs and update tickets?

Twilio Agent Connect provides a function-based tool system where your agent calls your APIs at runtime, including help desk APIs for updating tickets. Custom tools are defined in Python or TypeScript and work across voice and messaging.

How do you stop an AI agent from issuing fraudulent refunds?

Twilio Verify can confirm the customer controls the phone or email on file with a one-time passcode before any refund runs. Pair it with eligibility rules enforced in code and approval limits that escalate high-value or repeat requests to a human.

Can the same AI agent handle order questions over phone and SMS?

Twilio Agent Connect routes Voice, SMS, WhatsApp, RCS, and Chat through a single message handler, so one agent with one set of tools covers every channel. ConversationRelay handles speech for phone calls.

What happens when an AI refund agent can't resolve a request?

Twilio Agent Connect’s built-in handoff tool transfers the conversation to a human in Flex with the conversation ID and escalation reason. Add Conversation Intelligence and the agent also gets an AI-generated summary. The customer stays on the same call or chat.