Stop Sending AI Replies Until You Do This 2-Step Check

If you’re using AI to draft customer support replies, you’re probably feeling two things at once:

• Relief (because you’re saving time).
• Low-grade dread (because you know one bad reply can undo weeks of goodwill).

That dread is justified. In one evaluation summarized by Statista (based on reporting from the European Broadcasting Union and the BBC), 48% of chatbot responses contained accuracy issues, and 17% were significant errors (May–June 2025). (statista.com)

So here’s the rule: stop sending AI replies until you do this 2-step check. It takes under a minute once it’s a habit—and it’s the difference between “AI helps me” and “AI embarrassed me.”

The quick, neutral takeaway (before opinions)

• AI drafts are useful, but they can be confidently wrong and oddly cold.
• Customers are sensitive to support quality and trust.
• A lightweight review process beats “send and pray,” especially for indie teams without layers of QA.

Gartner’s research also signals customer skepticism: 64% of customers would prefer companies didn’t use AI for customer service. (gartner.com) You don’t fix that with better prompting. You fix it by catching errors and sounding human before you hit send.

The 2-step check (the whole post, distilled)

Step 1: The Truth Check (Is this correct for this customer?)

Your goal: ensure the reply won’t ship misinformation, wrong policy, or made-up steps.

Run this mental checklist:

• Claims: Any numbers, dates, limits, pricing, or “we don’t support X” statements—verify against your docs, code, changelog, or admin panel.
• Policy: Refunds, trials, cancellations, data retention, GDPR/DSAR language—confirm it matches your actual policy today.
• Product reality: Does the feature exist? Is it on their plan? Is it released on their platform (iOS vs web)? Is it behind a flag?
• User context: Did the AI assume things the customer didn’t say (device, plan, intent, urgency)?
• Next step: Are the instructions actually executable (correct menu names, settings paths, URLs, screenshots request, logs to capture)?

Fast trick: circle every sentence that would be harmful if wrong. Those are the only lines you must verify.

#### Common “AI draft” failure modes to watch for

• Invented UI labels (“Click Advanced Sync Mode…”) that don’t exist.
• Overconfident root causes (“This is definitely due to Apple’s iCloud…”).
• Policy improvisation (“We can offer a full refund after 90 days…”).
• Support theater: lots of words, no concrete path to resolution.

If you only do one thing, do this: delete anything you can’t personally vouch for. Replace it with a question or a verifiable step.

Step 2: The Trust Check (Does this sound like you—and does it reduce friction?)

Truth is necessary, but not sufficient. Support is also relationship management.

Zendesk’s CX Trends 2026 press release highlights how much customers value responsiveness and accurate resolution—86% of consumers say responsiveness and accurate resolution highly influence their purchase decisions. (zendesk.com)

So the second step is: read the draft once like you’re the customer. Then fix these four things:

• Ownership: Does it sound like you’re taking responsibility to help, or deflecting?
• Empathy (lightweight): One sentence that signals you “get it” without being cringe.
• Clarity: Short sentences, minimal jargon, obvious next step.
• Momentum: The customer should know exactly what happens next (what you need from them, what you’ll do, and when).

A simple template you can apply in seconds:

• Acknowledge: “Got it—thanks for the details.”
• Confirm (what you believe is happening): “It sounds like the export completes, but the file is empty.”
• Do (one concrete step): “Try X and tell me what you see.”
• If not (branch): “If that doesn’t work, send Y and I’ll check Z.”

This is also where “voice” matters. If your brand is direct and no-nonsense, say the thing plainly. If your brand is warmer, soften it—but keep it tight.

A real-world scenario (indie dev edition)

Customer message:

“I got charged twice. I’m on the monthly plan. Please fix ASAP.”

A typical AI draft might do something dangerous:

• It apologizes.
• It claims it’s a known Stripe issue.
• It promises a refund timeline that doesn’t match your process.

Run Step 1 (Truth Check):

• Verify in Stripe: is it actually two successful charges, or one charge + one authorization, or a proration, or multiple subscriptions?
• Confirm the customer’s email matches the Stripe customer.
• Confirm your refund policy and what you can do immediately.

Run Step 2 (Trust Check):

• Remove speculation (“Stripe sometimes…”).
• Add ownership + precise next step.

A safer, faster final reply often becomes shorter:

• “Thanks—double charges are scary. I’m checking this now.”
• “I see two successful charges for the same plan.”
• “I’m refunding one charge today; you’ll see it back in 5–10 business days depending on your bank.”
• “To prevent it happening again, I’ve canceled the duplicate subscription.”

(If any of those facts aren’t verified, you phrase them as questions or conditional steps.)

Pros and cons of AI-drafted support replies (with the check in place)

Pros

• Faster first drafts, especially for repetitive questions (reset password, cancel, invoice, bugs).
• More consistent structure (if you enforce it).
• Easier to maintain a calm tone when you’re tired.

Cons

• Accuracy risk (hallucinations, wrong UI paths, made-up policies). (statista.com)
• Customer trust risk: many customers don’t want AI in support, especially if it blocks humans. (gartner.com)
• “Polite but useless” replies that feel like stalling.

The 2-step check keeps the pros and cuts the worst cons.

Make the 2-step check almost automatic (process, not willpower)

You don’t need an enterprise workflow. You need a tiny ritual.

A 30-second “send gate” you can use anywhere

Before you send, ask:

1) What in this reply could be wrong? (Verify or delete.) 2) What would make me trust this reply if I were the customer? (Add ownership + next step.)

That’s it.

Where SupportMe fits (without turning this into a pitch)

If you’re using an AI assistant for support (including tools like SupportMe), the ideal setup is human-in-the-loop by default—drafts are generated, but nothing sends without your approval. That design choice matters because it forces the 2-step check to exist in the workflow, not in your head.

SupportMe’s angle—drafting in your writing style and learning from your edits via diff analysis—also helps with Step 2 over time: fewer “robot sentences,” more of your natural voice, less rewriting. (You still do Step 1 every time you’re stating facts or policy.)

The one quote worth remembering

Gartner put it plainly:

“Sixty percent of customer service and support leaders are under pressure to adopt AI in their function.” (gartner.com)

Pressure to adopt AI is real. So is the cost of sending sloppy answers. The way out is not rejecting AI—it’s reviewing AI like a grown-up.

Conclusion (short, practical, no drama)

AI can draft your replies. It can’t take responsibility for them.

Do the 2-step check every time:

• Truth Check: verify anything that could hurt you if wrong.
• Trust Check: make it sound human, clear, and action-oriented.

You’ll still move fast—just without the “what did I just send?” anxiety.