The reason most automated email feels robotic isn't the technology. It's the prompt. AI tools default to a tone that splits the difference between LinkedIn enthusiasm and customer service script — warm enough to feel polite, hollow enough to feel fake. Once you know what to fix, you can fix it in fifteen minutes.
This is a practical guide to making AI-drafted email sound like you, not like the average of every business email on the internet.
The single biggest fix
Almost every off-the-shelf email AI produces output with these traits:
- Opens with "I hope this email finds you well" or some variant
- Uses three or four exclamation marks per email
- Says "circle back," "touch base," "loop in," "deep dive"
- Has a closing line like "Looking forward to hearing from you!"
- Sounds vaguely upbeat regardless of the topic
This is the LLM averaging over the training data. The training data is corporate email. Corporate email sounds like this. Yours doesn't have to.
The fix is a single section in your prompt or your AI tool's system instructions:
Voice rules:
- Never use exclamation marks
- Never say "I hope this email finds you well" or any variant
- Never use the words: circle back, touch base, loop in, deep dive, reach out, synergy
- No "Looking forward to..." sign-offs unless the email is actually about an upcoming meeting
- Match the energy of the email's subject. Routine email = calm. Apology = direct. Bad news = brief.
- Use contractions. "We're" not "we are."
- One thought per paragraph. Short.
Paste this into any AI tool that drafts email — ChatGPT, your CRM's AI feature, your sales tool's email assistant — and most of the robot smell disappears immediately.
The structure that works
Beyond tone, structure matters more than people realize. Good business email follows a pattern:
- What this email is about — one sentence, no preamble.
- The specific ask or information — the meat.
- The next step, clearly named — what you want them to do or what you're doing.
That's it. Most emails are 60% longer than they need to be because the AI was trained on examples that buried the point.
Tell your AI tool:
Structure every email as:
1. One-sentence opener that names what this is about
2. Two or three sentences with the relevant detail
3. One sentence with the next step
Total length: under 100 words unless the topic genuinely requires more.
The 100-word rule is the most important. Emails that are short and direct get read and replied to. Emails that wander get scanned and forgotten.
Different tones for different jobs
A common mistake is using one tone for everything. The same email tool drafting a "thank you for your purchase" should not draft your "your payment is 60 days overdue" letter in the same voice.
Build three to five preset tones into your tool:
The "warm but professional" — for new customers, partners, first emails
Friendly but not cute. Use the recipient's first name. Acknowledge them as a person
once. Get to the point in the second sentence. Don't oversell what's in the email.
The "direct and brief" — for routine business, follow-ups
Skip pleasantries. State what this email is for. Provide the information or ask.
End with the next step. Under 80 words.
The "firm but fair" — for overdue payments, broken commitments, second-time complaints
Open with the specific issue, not feelings about the issue. Be factual. State what
needs to happen and by when. No softening language like "I understand things happen."
Close with a clear next step.
The "apology that doesn't grovel" — for our mistakes
Acknowledge the specific thing that went wrong. One sentence on why. Two sentences
on what we're doing about it. Don't over-apologize. Don't offer free things unless
told to.
The "bad news, said well" — for declining requests, rejecting proposals
State the decision clearly in the first sentence. One sentence of reasoning if it
helps the recipient. Don't pad with sympathy. Offer one constructive next step if
appropriate, none if not.
Build these as separate options in whatever tool you use. The drafts will be dramatically better, immediately.
The guardrails that matter
Beyond tone, there are a handful of practical guardrails that prevent AI email from going wrong:
- Block customer names in subject lines unless they're confirmed correct. AI getting a name slightly wrong in the body is recoverable. In the subject line it's the first thing the recipient sees.
- Never let AI quote prices. It will guess and be wrong. Pull the price from a real source, paste it into the email yourself.
- Be explicit about who you are. AI doesn't always know whether to write "I" or "we." Tell it. "I'm Sarah, the owner. Sign off as 'Sarah.'" Otherwise it'll sometimes default to a generic "the team" voice that nobody trusts.
- Confirm details before AI uses them. If the prompt says "follow up on the meeting we had on Tuesday," and there was no Tuesday meeting, the AI will write a confident email referencing a meeting that didn't happen. Garbage in, embarrassing out.
The one settings change that quietly fixes most email AI
If you're using an AI tool inside your CRM or email platform (HubSpot, Zoho, etc.) that has any kind of "default tone" or "AI personality" setting, change it from whatever it currently is to "direct" or "concise" if those are options. If the only options are "friendly" and "professional," choose professional. The defaults are tuned for engagement metrics, not for sounding like a real person.
You will not get every email right with these rules. But the floor of how bad an AI-drafted email can be — for your brand, for your reputation — goes up dramatically when you give the tool real guardrails instead of letting it default to the trained-on-average voice.
The honest test
Read three of your recent AI-drafted emails out loud. If you wouldn't say it that way on the phone to that person, the tone is wrong. Adjust until you would. That's the whole skill.
You don't need a course on prompt engineering. You need to actually read what your AI is sending in your name, and have an opinion about it.