How AI Agents Handle Refunds and Returns

Refunds and returns are where AI customer support meets real money. The agent's choices have direct cost implications. Done right, AI handles 70%+ of refunds and returns within policy, escalates the edge cases cleanly, and saves your team hundreds of hours a month. Done wrong, you end up with surprise refund spikes and confused customers.

TL;DR

• Define a refund/return policy the agent can mechanically follow. No improvisation.
• Cap the agent's authority — refunds above $X always escalate.
• Always confirm before issuing a refund. "Just so I have it right, we're processing $35 back to your card?"
• Log every refund the agent issued; spot-check weekly.

The basic flow

For a refund or return request:

1. Identify the order. Caller mentions order ID, or agent looks up by name/email.
2. Check eligibility. Within return window? Item-eligible? Customer in good standing?
3. Determine action. Refund? Replacement? Store credit? Escalation?
4. Execute and confirm. Process the action; tell the caller; send confirmation.
5. Log. Store the decision and the reason for audit.

Each step needs the right tool calls and policies wired in.

Defining the agent's authority

This is the most important policy decision. What can the agent do without human involvement?

A reasonable starting point:

• Refund up to $50 for items in the standard return window.
• Refund up to $100 for shipping issues clearly the company's fault.
• Above those: escalate to human review.

Express in the prompt:

You can issue refunds up to $50 for in-window returns or
up to $100 for shipping issues. For any refund over those
limits, escalate via transfer_to_human and don't make any
commitments to the caller about whether the refund will be
approved.

And in the function:

function process_refund({ order_id, amount, reason }) {
  if (amount > 100) throw new Error("Amount exceeds agent authority");
  if (amount > 50 && reason !== "shipping_issue") {
    throw new Error("Amount exceeds standard refund limit");
  }
  // ...
}

Belt and suspenders. The function-level guard prevents the LLM from going rogue.

The "always confirm" pattern

Before any irreversible action (refund, charge, account change), the agent should confirm explicitly:

"Just to confirm, we're processing a $35 refund to your card
ending in 4521 for order 1976. Should I go ahead?"

Wait for "yes." Then execute.

This catches:

• Misheard amounts.
• Caller second-guessing.
• The caller asking about something different.

The cost: 5–10 seconds per refund. The benefit: avoided wrong refunds.

When to escalate vs handle

Escalate refunds when:

• Amount exceeds the agent's authority.
• Caller is disputing fraud.
• Caller is angry about prior interaction.
• Order is outside return window but caller has a compelling reason.
• Multiple items / partial returns get complicated.
• Caller asks for a manager.

Handle when:

• Item is within window, item is eligible, caller agrees to standard refund.
• Shipping issue is clear (lost, damaged, wrong item).
• Caller's request is reasonable and within policy.

The "policy lookup" function

Don't bake policy into the system prompt verbatim — it'll go stale. Instead:

function get_return_policy({ category, days_since_purchase }) {
  // Returns: { eligible: bool, refund_amount: number, restocking_fee: number, ... }
}

The agent calls this; the result determines what the agent says. Update the policy logic without redeploying the agent.

Patterns for common refund scenarios

"I never received my order."

• Look up order; check tracking.
• If marked delivered: ask if they checked common spots, neighbors, etc.
• If still missing: process as lost shipment per policy.
• If shipping carrier dispute: file claim and offer refund or replacement.

"The item arrived damaged."

• Ask for photos (escalate to human or agent-assist channel for image review).
• If clearly damaged: process refund or replacement.
• If borderline: escalate.

"I changed my mind."

• Within return window: process per standard policy.
• Outside window: escalate or politely decline per policy.

"My subscription auto-renewed and I want a refund."

• Check policy on auto-renewal refunds.
• Often these are case-by-case; consider escalating.

What kills the refund flow

Common failures:

Refund issued without confirmation. Caller meant to ask about a refund, agent issued one. Always confirm.

Wrong amount refunded. Caller said "for the shoes" referring to one of three items; agent refunded the entire order.

Refund issued outside policy. Agent rationalized that the caller deserved it. The function should prevent this.

Refund offered when not appropriate. Agent gave a refund as a courtesy when the customer hadn't asked for one. Annoying and costly.

No log of why. Six months later, no one remembers why this customer got a refund. Always log the reason.

Compliance

For some regions / industries, refund handling has compliance implications:

• EU: 14-day right of withdrawal regardless of policy.
• Some US states: specific rules for digital goods, subscriptions.
• Healthcare: HIPAA constraints on what can be discussed.

Build the agent's policies with legal review.

Measuring refund quality

Beyond standard support metrics, track:

• Refund rate by category. Is the agent refunding more than expected for any specific intent?
• Refund amount distribution. Spike in larger refunds → investigate.
• Refund-to-resolution ratio. Of calls about returns, what percentage resulted in refunds vs other outcomes?
• Customer satisfaction post-refund. Did the refund actually solve the customer's issue?

For more, see how to measure voice agent quality.

Related reading

• The Definitive Guide to AI Customer Support in 2026
• Building a Tier-1 AI Support Agent Step by Step
• Why "Human-in-the-Loop" Beats "Fully Autonomous" for Most Teams
• Designing AI Agents That Cancel Subscriptions Honestly
• Voice Agent Onboarding: A 30-Day Plan for Support Teams

FAQ

What's a reasonable refund authority for a tier-1 agent? Up to $50–$100 for most consumer products. Higher for premium products. Always escalate above the cap.

Should the agent handle disputes? Triage them; route to a human for resolution. AI shouldn't make judgment calls on disputes.

Can the agent issue partial refunds? Yes, with appropriate caps. Often easier than escalating.

What about coupons / store credit instead of refund? Agent can offer; should be a positive add-on, not a replacement when refund was promised.

How do I prevent agents being too generous? Function-level caps. Audit logs. Weekly review of refunds issued.