A good customer service reply does four things: it shows you read what the customer actually wrote, it says plainly what went wrong, it says what happens next and when, and it sounds like one person talking to another. AI can draft that quickly, but only if you give it the facts and it gives you back a message, not a template.
The fastest way to get there is to write two or three lines of notes (what happened, what you will do, by when) and ask the AI to turn them into the reply. Then read it once before it goes, because the AI does not know your policy unless you tell it.
Why so many support replies sound like a robot
People can tell a canned reply within a line. The usual signs:
- An opening that could go to anyone. "Thank you for reaching out to us regarding your concern."
- Apology without content. "We apologise for any inconvenience this may have caused." It says sorry for nothing in particular.
- No specifics. No order number, no date, no name of the thing that broke.
- Passive voice hiding the actor. "Your order has been delayed." By whom, and why?
- A closing formula. "Please do not hesitate to contact us should you have any further questions."
AI makes this worse when it is asked for "a professional customer service reply", because that phrase points it straight at the template. Ask instead for the message a helpful person on your team would send.
A structure that works
Most support replies fit five short parts. Not every reply needs all five, and none of them needs more than a sentence or two.
- Name what happened. Use their words and their details: the order, the date, the thing that went wrong.
- Own it, if it is yours. "That's on us" is stronger than "we apologise for any inconvenience".
- Say what you have done or will do. A refund issued, a replacement sent, a fix in progress.
- Give a time. "By Friday", "within three working days", "I'll write again tomorrow". A vague "soon" makes people chase.
- End like a person. Your name, and one plain line if it fits. No formula.
Example replies
The facts in these are made up for the example. Change them to yours.
A complaint about a damaged item
The customer wrote that their mug arrived in pieces and they are annoyed because it was a gift.
Hi Sarah,
I'm sorry the mug arrived broken, especially when it was meant as a present. That's our packing, not the courier's luck, and I've passed your photos to the team who pack the orders.
A new one went out this morning by next-day delivery, so it should be with you on Thursday. You don't need to send the broken one back.
Tom
A refund you can give
Hi James,
You're right, you were charged twice for the April subscription. I've refunded the second payment today. It usually shows on your statement within five working days, depending on your bank.
Sorry for the hassle of having to spot it yourself.
Priya
A refund you cannot give
Saying no is one of the hardest replies to write. Be clear about the no, give the real reason once, and offer what you can.
Hi Mark,
I've looked at your order. It was bought 64 days ago, and our returns window is 30 days, so I can't refund it.
What I can do is repair it free of charge if the fault is covered by the one-year warranty. If you reply with a short description of what's wrong, I'll book it in.
Anna
A delay
Hi Lena,
Your order hasn't shipped yet, and I should have told you before you had to ask. One of the parts came late from our supplier. It arrives Tuesday, and your order will go out Wednesday. I'll send the tracking link the moment it does.
Dan
Notice what these leave out: no "we value your feedback", no "rest assured", no second apology at the end, no list of options for the customer to choose between. Each one says what happened, what is being done and when.
Using AI to write them
AI is good at this job when you do the thinking and let it do the wording.
- Paste the customer's message. The whole thing, so it can pick up their details and their tone.
- Add your notes. What happened, what you will do, the date. For example: "damaged in packing, replacement sent today next-day, no need to return".
- Say who is writing. "Write it as Tom from the support team, friendly and short."
- Ask for the reply only. If your tool adds a preamble or three versions, you will spend longer cutting than writing.
- Read it before it goes. Check every fact against your notes and your policy.
The rule that matters most: never let the AI decide the facts. It does not know your refund policy, your stock or your delivery times. A model can be wrong and still sound certain, and in support that means a promise your company has to keep. Give it the decision and let it write the words.
Where Hemmingway fits
Hemmingway-1 was trained for exactly this kind of writing: short messages to real people, including the hard ones. In CommunicationBench, our own blind test of everyday writing, it came first of the models we tested. On the hardest requests, the messages people keep rewriting, a judge asked which reply sounded like a person wrote it picked Hemmingway-1's 72% of the time, against 9% for GPT-6 Astra. These are our own benchmarks, and we say so on the model card.
It also gives you the message and nothing else. In the same tests, several large models buried the reply inside commentary, options or notes in more than nine replies out of ten. For a support team, that is the difference between a draft you can send and a draft you have to edit.
There are two ways to use it.
- The app. Paste the customer's message and your notes into Hemmingway on the web, Mac, Windows or Android, and it writes the reply. The Inbox in the apps can also read your mail and draft replies for you to check and send. How that works.
- The API. If you run your own help desk software, Hemmingway-1 is available through an OpenAI-compatible API as
hemmingway-27b. The docs are here.
Try HemmingwayDownload the app
What to check before you send
- Every number, date and amount matches your records.
- It does not promise anything your policy does not allow.
- The customer's name and details are right.
- It answers everything they asked, not only the first question.
- It reads like someone on your team would say it out loud.
If the message involves legal claims, safety or a large sum of money, have a person write or approve it.
Common questions
Can AI write customer service emails?
Yes, and it is good at the wording. It is not good at deciding what your company should do, so give it the facts and the decision in a couple of lines, and read the draft before it goes out.
How do you make AI customer service replies sound less robotic?
Ask for the message a helpful person on your team would send, not "a professional reply". Give it specifics, ask for one reply with no preamble, and cut stock lines such as "we apologise for any inconvenience" if they appear.
How do you say no to a refund politely?
State the no clearly, give the real reason once, and offer what you can do instead, such as a repair, a credit or an exception route. Do not bury the no in apologies, because it makes the customer read the message twice.
Should customer service replies apologise?
Apologise once, for the specific thing that went wrong, when it was your fault. A single "I'm sorry the mug arrived broken" does more than several general apologies.
Is it safe to let AI send support replies on its own?
Only for simple, low-risk answers under clear rules. For refunds, complaints and anything involving money or policy, keep a person reading each reply before it is sent.