Why AI drafts sound generic, and what fixes it
By Sam Kim · · 6 min read

Why do AI-written emails sound so generic?
Because a one-line request into a blank box tells the tool nothing about you, your client or how you talk, so it writes what any business might say. The fix is better inputs rather than a better product: real examples of your own replies, a short page of voice rules, the facts of this particular message and the outcome you want. A person still reads anything that goes to a client, because a draft can sound exactly like you and still be wrong.
Why does AI write like that in the first place?
With nothing specific to go on, it plays safe. Safe means polite, padded and free of detail, which is how you end up with "I hope this email finds you well" and "please don't hesitate to reach out".
You'll know the other habits once you look for them. Three reasons where one would do. A warm restatement of what the client just said. An upbeat closing line nobody asked for. A tradie who normally replies "Yep, Thursday 7am works, can you send the gate code" gets back three paragraphs about looking forward to the opportunity.
Clients notice, even if they couldn't tell you why. A long-standing client who gets a message that doesn't sound like you may wonder whether you read theirs.
What does the tool need before it can sound like you?
Four inputs, and most people supply none of them:
- Real examples. Five to ten replies you actually sent, including the short ones. Cover the usual situations: a quick yes, a no, a follow-up on a quote, an apology, a chasing email. Examples work better than descriptions, because "friendly but professional" describes nearly every business in the country.
- Voice rules. A page of plain instructions about how you write. More on these below.
- The facts of this message. Who it's going to, what they asked, what happened, what they've already been told. Without these the tool fills the gaps with pleasantries, or worse, with invented detail.
- The outcome. What you want the person to do after reading it. Pay the invoice, pick a date, stop worrying.
The third one needs the most discipline, since it changes every time. Thirty seconds of dot points before you ask for a draft is usually the difference between something you can send and something you have to rewrite.
How do you write voice rules that work?
Read your own sent folder and write down what you actually do. Adjectives won't help the tool. Observable habits will.
Take half an hour and twenty sent emails, and note:
- How you open. First name and straight in, or something warmer?
- How long a normal reply runs.
- How you sign off.
- Words you'd never use.
- How you say no.
- How you deliver bad news: apology first, or the fix first?
- Whether you use contractions, bullets or exclamation marks.
Then turn those notes into instructions, one per line. Yours might look something like this:
- Two to four sentences, unless the client asked several questions.
- Put the answer in the first sentence.
- Never open by hoping they're well.
- Say no straight away and give the reason next.
- Ask one question per email.
- Don't thank anyone for reaching out.
Then test it on your own history. Give the tool an old client email and your rules, ask for a reply, and compare it with what you really sent. Where it misses, add a rule or an example. Two or three rounds on a handful of emails will show you most of what's wrong.
Why does it slide back to sounding generic?
Usually because the setup lived in one chat window or one person's head, and the next conversation started from nothing. Keep the rules and examples in one place, and make sure they're used every time. Check whether your tool can store standing instructions. If it can't, one shared document works.
If several people send email for the business, this matters more. Five staff with five private ways of asking will give your clients five different versions of you. One page, one named owner, and a review whenever someone spots a draft that sounds off. Getting a team to actually use it is its own problem and worth reading before you roll anything out.
One caution before you paste in those sent emails. Real emails hold client details. Strip out names and anything identifying from your examples, and check what your tool's current terms and settings say about the data you put in. Professional firms have extra duties here, covered in AI and client confidentiality. This is general information, so confirm your own obligations.
What won't this fix?
Knowing what you know. The tool can learn how you sound, but it has no idea that this client was burned by the last delay, or that "no rush" from this particular person means Friday. I'd keep these for a person, with AI tidying the wording at most:
- Complaints, disputes and anything where the relationship is on the line.
- Bad news a client will remember.
- Any message where you haven't decided what you think yet. A draft can't supply a position, though it will happily pretend to have one.
Commitments need a different kind of care. Check every date, price and promise in a draft against the job itself. A draft that sounds like you is harder to catch out, because it reads as though you already checked it. Where AI gets it wrong goes further on this.
At the other end, some messages don't need your voice at all. An appointment confirmation can be plain and identical every time. Spend the effort on the replies clients actually read for tone.
The working arrangement I'd aim for is simple. The tool does the first draft of routine replies, the person reads, fixes and sends. For a straighter account of the limits, see what AI won't do.
How do you know it's working?
Measure the editing, because that's where the time goes. Count your routine replies for a week. Time ten written from scratch, then time ten edited from a draft.
Take the difference in minutes, multiply by your weekly count and multiply by 48 working weeks, then divide by 60. That's your hours a year, on your numbers.
If you're rewriting more than you keep, the inputs are still thin, or that type of message wasn't suited to a draft in the first place. Look at what you changed. Repeated edits are rules you haven't written yet.
What to do next
- Pull twenty sent emails this week and write your page of voice rules from them.
- Choose five to ten examples across different situations, with client details removed.
- Test on three old emails and compare the drafts with what you sent.
- Give the page an owner and keep it somewhere everyone drafting email can reach.
- Decide which messages stay fully human, and tell your team.
General information, not legal, tax or financial advice · Drafted with AI tools to my written rules · Terms