Reply to a difficult customer complaint without promising what you can't give

You're writing a first response today that does four jobs: shows Joanne someone has actually read her email, apologises for what went wrong, tells her what's already fixed and what happens next, and routes the refund and compensation request to the person who decides it.

You are not deciding the refund, arguing about the details or winning the exchange. You're also not trying to prevent the review by saying whatever it takes. A reply that promises too much creates a second complaint when it isn't honoured.

Useful, with care. AI is good at two parts of this job: pulling the checkable claims out of an emotional email, and giving you a calm first draft when you're rattled. It's weak at the part that matters most, knowing what Fernway can and can't offer, and its default complaint tone is corporate and hollow.

Use it to sort the facts and draft. Write the promises yourself.

The complaint as written: ■ HIGH RISK for a general AI tool.

It contains the customer's name, job title, business address and mobile number, and a health detail about a third person (a colleague's bad back). Health information is special category data under UK data protection law, which means it needs extra protection. That colleague hasn't agreed to anything and doesn't need to be in an AI prompt at all.

After removing personal details: ▲ CAUTION

The order details and the story of what went wrong are still Fernway's customer records. Use a tool your organisation has approved for customer correspondence, signed in with your work account. If the only tool available is a personal account on a consumer AI service, write this reply yourself. Your organisation's policy and approved tools still take priority.

Give the AI what it needs to draft, and nothing it doesn't. It needs the substance of the complaint, the relevant facts from the log and the limits of your authority. It doesn't need Joanne's phone number, the office address or anything about her colleague's health.

Removing the name isn't the same as making this anonymous: a four-year customer who opened an office on 14 September is identifiable to anyone inside Fernway. That's why the approved-tool rule still applies after redaction.

The drafting principles are the same in any tool. Where you do it changes how much you need to remove.

Start by asking the AI to separate what can be checked from what can't. This is quick, and it stops the draft being built on the angriest version of events.

Here is a customer complaint and our order log. Make a table with three columns: what the customer says, what our log shows, and whether it's confirmed, partly confirmed, not in our records, or a request for a decision. Don't judge the customer. Don't add anything that isn't in either document.

[paste redacted complaint and log]

Then draft. The weak instruction produces the reply in step 7.

This is the kind of answer the weak instruction can produce. Read it as Joanne would.

Other things to check in any draft:

• Arguing with the details. Some drafts "correct" the customer: "only 3 of 27 items were damaged". True, and it reads as dismissive. Say exactly what's being fixed instead.
• Repeating the health detail. Don't mention the colleague's back in the reply. Acknowledge the disruption to her team.
• Inventing a fix. Check any date or action in the draft is in the log. A draft that says "we have arranged a fitting on 30 September" has made that up.
• Ignoring the driver. A conduct complaint can't be checked from the log, but it can't be ignored either.

Before you correct the draft, do the one thing the AI can't: ask Priya when she can respond. She says she'll call Joanne by Friday 25 September. Now that date can go in the reply.

Rewrite this reply. Remove the refund, the compensation offer and "full responsibility". Instead, say the refund and compensation request has gone to Priya Shah, Head of Operations, who makes those decisions and will contact the customer by Friday 25 September. Apologise specifically for: delivery on 17 September instead of 11 September; the delay email going to their accounts inbox instead of to the customer; the two chipped desk tops and broken chair. Say the replacement chair arrived on 18 September and the desk tops are due on 29 September. Say the driver concern has been passed to the depot supervisor. Remove every stock phrase. Sign it from Leah Bennett.

Then edit by hand. Read it aloud. If a sentence sounds like something you wouldn't say on the phone, rewrite it in your own words.

Probably, but less than it seems. The facts table was the real time-saver: it turned an angry email into a list of what to apologise for and what to pass on, and it's easier to think clearly from a table than from capital letters.

The draft saved less. Every sentence that mattered (the promises, the dates, Priya's deadline) had to come from you and your checks. If you write complaint replies often, you'll be faster from your own notes than from fixing an AI's template.