The Wrong Refund Policy Email — and What It Actually Cost

The customer's email was polite but firm.

They'd asked for a refund outside your standard window. And somewhere in your inbox, buried in a thread from three weeks ago, your AI had drafted a reply that said you'd honor it.

You don't remember sending that. But it's there, in writing, with your name on it.

The cost isn't always immediate. Sometimes it shows up weeks later, when a customer quotes your own email back at you.

How it happens

You're handling a support email. The customer is asking for a refund on a subscription they forgot to cancel. You're busy. You open the AI draft, it looks reasonable, you hit send.

Your policy was updated six months ago, after a billing review: refunds are issued within 14 days of purchase, not 30. Generic AI doesn't know that. It writes from what sounds plausible for a company like yours — and plausible is not the same as accurate.

The customer now has an email that reads as a commitment. You have two options: honor it and eat the cost, or walk it back and damage the relationship. Neither is free.

The stale doc problem

Most companies update their policies more often than they think. Refund windows change. Cancellation terms get tightened. Trial conditions shift after a pricing review. What was true 12 months ago isn't always true today.

Generic AI has no way to know which version of your policy is current. It doesn't read your help center. It doesn't check your terms of service. It generates text that sounds like what a company in your category would say — which might be last year's version, or a competitor's version, or a version that never existed at all.

It shows up across every policy type:

• Cancellation terms — "You can cancel anytime and receive a prorated refund" when your policy no longer includes proration
• Trial extensions — "We'd be happy to extend your trial" when that offer was discontinued
• Upgrade pricing — "Your current rate will be grandfathered" when no such guarantee exists
• SLA commitments — "We respond to all tickets within 4 hours" when your SLA was revised upward

In each case, the AI is confident. The draft sounds professional. And the customer reads it as a binding statement from your company.

What grounded AI does differently

The fix isn't "write better prompts" or "double-check every draft." Both of those put the burden back on the person who was trying to save time in the first place.

The fix is grounding — connecting the AI to your actual, current policy documents before it writes.

Here's the difference in practice. A customer asks for a refund outside your standard window:

Generic AI draft:

"Thanks for reaching out. We understand these things happen — we'd be happy to process a refund for you as a one-time courtesy. Please allow 5-7 business days for it to appear on your statement."

Reads fine. Commits you to something you didn't intend.

KB-grounded draft:

"Thanks for getting in touch. Our refund policy covers purchases within 14 days — as your subscription renewed on March 3rd, it falls outside that window. I'm not able to process a refund in this case, but I'm happy to help you cancel before the next billing cycle to avoid future charges." Source: refund-policy.md, updated January 2026

The grounded draft pulled from your actual policy doc, cited the source, and gave you something you can send without second-guessing. The citation means you can verify the claim in five seconds before hitting send — and if the policy has changed again since the doc was last updated, you'll catch it.

The review step is only useful if the draft is verifiable

Most AI email tools ask you to review before sending. That's good advice. But review is only meaningful if you have something to check against.

If a draft says "our refund window is 30 days" with no source attached, review means opening your help center, finding the right article, and confirming the number. That's 3-4 minutes per email. At 20 support emails a day, that's over an hour of verification work — which is most of what the AI was supposed to save you.

If the same draft says "our refund window is 14 days" and cites refund-policy.md, updated January 2026, review means clicking the link and confirming the doc still says what the draft says. That's 15 seconds.

Source citations don't just make drafts more accurate. They make the review step fast enough that people actually do it — which is the only real protection against the wrong answer going out.

The cost is rarely just the refund

The direct cost is usually small. The indirect cost is harder to measure.

You've set a precedent. The customer knows that if they push back, you'll honor it — and if they tell others, you've implicitly communicated a more generous policy than you have. Internally, the cleanup takes time: someone has to find the thread, understand what happened, and draft a careful response either way. That's founder time, CS lead time, or manager time.

What to do about it

Three things that reduce the stale-doc risk in AI email drafting:

1. Connect AI to your live policy docs. If your refund policy lives in Notion, Google Drive, or your help center, connect it. A grounded system retrieves the current version before drafting — not a cached or assumed one.

2. Require source citations on factual claims. Any draft that states a policy, a price, or a term should cite the document it came from. If it can't cite a source, it shouldn't be stating the fact.

3. Keep your canonical docs current. Grounding is only as good as what you connect it to. A well-maintained, version-dated policy doc is the foundation everything else depends on.

Inbox SuperPilot connects to your help center, Google Drive, Notion, and other sources to create drafts for your review with inline citations on every factual claim. When a customer asks about your refund policy, the draft cites the exact document and update date. You verify in seconds, not minutes.

Try it free inside Gmail — no card required.

Further Reading & References

From the Inbox SuperPilot Blog

• What We Learned Building a Citation Engine for Email AI
• Gmail Gemini vs. Knowledge-Grounded AI: An Honest Comparison
• Why Generic AI Fails in Customer Support Email Workflows
• 5 Email Mistakes AI Catches That Humans Miss

External References

• Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks — Lewis et al., 2020. The foundational paper on why grounding generation in retrieved documents improves factual accuracy.
• Enhancing LLM Factual Accuracy with RAG to Counter Hallucinations — Li et al., 2024. Case study on domain-specific RAG in private knowledge bases — directly relevant to policy and pricing docs.
• Air Canada Held Liable for Its Chatbot's Hallucinated Refund Policy — TechHQ, 2024. A Canadian tribunal ruled Air Canada responsible for incorrect policy information generated by its chatbot — establishing that companies are liable for what their AI tells customers, regardless of accuracy.