Forward Deployed EngineeringKnowledge base

Using AI for customer emails without sounding like a robot

Last updated

On this page
  1. Why it sounds like a robot
  2. A prompt that works
  3. Before and after
  4. Edit it every time
  5. Emails to write yourself
  6. Be careful what you paste
  7. Should you tell customers?

You know the email. "Thank you for reaching out! We truly value your business and are delighted to assist." Nobody talks like that. Customers can tell, and they trust you a little less for it.

AI tools such as ChatGPT and Claude can still save you real time on email. The trick is treating the output as a rough draft from a new assistant, not as something you send.

Why it sounds like a robot

The tool has never met your customers or heard you talk. Left alone it falls back on the safest, most generic business voice there is. That voice is full of phrases like "I hope this email finds you well," "please don't hesitate to reach out," and three exclamation points in a row.

You fix this by giving it better material. Three things help.

Show it your voice. Paste in two or three emails you've written that sounded like you. Say, "Write like this." Remove customer names and private details first.

Give it the facts. The more you hand over (the price, the date, what you can and can't do), the less it makes up. Vague prompts get vague, padded replies.

Tell it how long and how casual. "Four sentences. Friendly, like a text to a neighbor." Without that, it writes five paragraphs.

A prompt that works

Here's one to adapt.

You're helping me reply to customers for my taco shop. I write short,
warm, plain emails. No corporate phrases, no exclamation marks except
maybe one. Here are two emails I wrote that sound like me:

[paste email 1]
[paste email 2]

A customer wrote the message below. Draft a reply of about 4 sentences.
Facts to use: we do catering for 20 to 200 people, we need 3 days notice,
the deposit is half. Don't promise anything not listed here.

[paste the customer's message]

That last line matters. "Don't promise anything not listed" cuts down the habit of making up policies you don't have.

Before and after

Here's the kind of draft you get with a lazy prompt:

Thank you so much for reaching out! We are thrilled that you are considering us for your upcoming event. We would be delighted to assist you with your catering needs. Please don't hesitate to contact us with any further questions!

And here's what it looks like after you've given it facts and your voice, then trimmed it:

Hi Maria, yes, we can do 60 people on the 14th. We need three days notice and a half deposit to hold the date. Want me to send a menu and a quote today?

The second one is shorter, has an actual answer, and sounds like a person. That's the goal.

Edit it every time

Even a good draft needs a pass.

  1. Check every fact. Names, dates, prices, hours, addresses. The tool will get one wrong sooner or later, and it'll look very sure about it.
  2. Cut the first sentence. Usually it's throat-clearing.
  3. Read it out loud. If you'd feel silly saying it to the customer's face, change it.
  4. Delete any promise you didn't mean. Refunds, discounts, deadlines. The tool is generous with your money.
  5. Look for the sign-off. "Warm regards" from a taco shop is a tell. Use whatever you really say.

Emails to write yourself

Some messages shouldn't come from a draft at all.

  • A complaint from someone who's actually upset. They want to know a person read it.
  • Anything about money owed, a refund, or a dispute
  • Anything involving someone's health, a loss, or a hard situation
  • Anything with legal weight, such as contracts, cancellations, or damage claims

For those, write it yourself, even if it's clumsy. Clumsy and human beats smooth and generic. You can still ask AI to read your draft and point out anything that sounds harsh, but the words should be yours.

Be careful what you paste

Whatever you paste goes to a company's servers. Before you paste a customer's message, take out anything that doesn't need to be there: last names, addresses, phone numbers, account details. If you run a clinic or work with health information, don't paste patient details into a general AI tool at all. Health information has its own rules, and your tool's settings and agreement matter. Check what the tool does with what you type, which is usually in its privacy settings.

Should you tell customers?

You don't need to announce it, any more than you'd announce using spell check. You're still the one sending the email and standing behind it. What you shouldn't do is pretend a chatbot is a person. If customers are talking to an automated assistant, say so.

For a few more ways AI trips people up, see what AI gets wrong, and how to catch it.

Was this helpful?