Why AI emails sound like AI
Ask a chatbot to "write a professional email" and you get the average of every corporate email ever written: "I hope this finds you well", "I wanted to reach out", "please don't hesitate to contact me". Recipients have learned to recognise this register, and it quietly signals that you could not be bothered. For a small firm whose entire advantage is being human and local, that is an expensive impression to make.
The fix is not a better model; it is better inputs. Left to guess, the model writes like everyone. Shown how you actually write, tools such as Claude, ChatGPT and Copilot in Outlook imitate you surprisingly well. The workflow below takes an hour to set up and then saves time on every routine email without making you sound like a press release.
Build a style sample pack
Your sent folder is a training set you already own. Building the pack takes four steps:
- 1. Pull ten to fifteen of your best sent emails covering different situations: a quote, a polite chase, a thank-you, a tricky no, a schedule change.
- 2. Strip out names, prices and anything confidential; the style survives, the data should not travel.
- 3. Note your quirks explicitly: how you greet, how you sign off, typical sentence length, whether you use contractions, words you would never use.
- 4. Store the pack somewhere persistent: a Claude Project, a custom GPT, or a pinned document you paste in.
Then add one standing instruction alongside the samples, for example: "Write in the style of these examples. Short sentences, contractions, no 'I hope this finds you well', no exclamation marks, sign off with 'Best, Raj'." Explicit bans work better than vague requests for a friendly tone, because the model needs to know which of its habits to suppress.
Need a hand with this?
Our team delivers Email Marketing for UK businesses — with a free initial consultation, transparent fixed quotes and no lock-in contracts. Tell us what you're working on →
The three-step drafting workflow
Step one: brief it like a colleague
Give the model who the email is to, your history with them, the outcome you want, the key facts, and the mood. "Chase the Hendersons for the March invoice, third reminder, they're usually good payers so keep it warm but firmer than last time" produces a dramatically better draft than "write a payment reminder".
Step two: generate against your style pack
Draft inside the Project or custom GPT holding your samples so every email inherits your voice without repeating the setup.
Step three: the sixty-second edit pass
Read it aloud; anything you would not say, cut. Delete hedging filler, check every fact and figure against reality, and rewrite the first line so it could only have been written to this person. The first line carries most of the human signal in any email.
Snippets for the repeatable eighty per cent
Most business email is a handful of situations in rotation: chasing an invoice, confirming a booking, declining politely, requesting information. For these, build reusable prompt snippets in a text expander such as TextExpander or the free Espanso, in Outlook Quick Parts, or simply as saved prompts in your AI tool.
A snippet is a fill-in-the-blanks brief, for example: "Invoice chase, tone [warm/firm/final], invoice [number] for [amount], due [date], history: [context]. Match my style pack." Typing a shortcut and filling four brackets takes fifteen seconds and produces a draft that needs almost no editing, because both the situation and your voice are fully specified. Over a month this is where the real time saving accumulates, not in the occasional long email.
Before and after
The difference the workflow makes is easiest to see side by side. A generic AI invoice chase reads: "I hope this email finds you well. I am writing to follow up regarding invoice #1042, which remains outstanding. We would greatly appreciate it if you could arrange payment at your earliest convenience. Please do not hesitate to reach out with any questions."
The same request drafted against a style pack and a proper brief: "Hi Sarah, quick nudge on invoice 1042 (£840, due 28 June). I know things get busy after half term, so no drama, but could you get it over this week? Shout if anything on it looks off. Best, Raj." Same message, forty per cent shorter, and unmistakably from a person.
Notice what changed: a named recipient, a concrete detail that proves familiarity, and the removal of every stock phrase. That is exactly what the sixty-second edit pass enforces when the draft falls short.
Key Takeaway
Build a style pack from ten to fifteen of your best sent emails, add explicit rules banning stock phrases, and store it in a Claude Project or custom GPT. Then brief like a colleague, generate against the pack, and finish with a sixty-second edit pass that personalises the first line and checks every fact. Use fill-in-the-blank snippets for repeatable emails, and write complaints, condolences, HR and legal messages yourself, every time.
When AI should never write the reply
Some emails are the relationship, and outsourcing them damages the thing they exist to protect. Keep AI away from:
- Complaints and angry customers, where a templated cadence reads as contempt.
- Condolences, serious illness and personal bad news.
- Redundancy, disciplinary and other HR matters, which also carry legal risk.
- Anything in or near a legal dispute, where wording may be quoted back at you.
- Apologies for your own mistakes; these only count when they visibly cost you effort.
A workable rule: if getting the tone wrong would cost more than the ten minutes the draft saves, write it yourself. And never let any tool auto-send; the human clicking send is your last and best quality control. If you want help setting up tone-matched drafting or wider email marketing that sounds like you, our team at Thind Global Services can help.