Use an LLM to draft and tighten customer emails faster (without sounding fake)

Quote follow-ups, appointment confirmations, late-arrival notes, and review requests all fail the same way. Style adjectives do not constrain generation. Facts do. Open any chat interface that hosts a language model — ChatGPT, Claude, Gemini, or similar — with one real email you owe today.

Hallucination is not a tone bug

If the output is too long, too kind, and slightly wrong, the usual instinct is to ask for a more human voice. That is the wrong diagnosis. Extra unused context invites hallucination. A dumped thread is unused context. A system prompt full of “be warm” with no job and no facts is unused context.

Anthropic’s guidance is to let the model admit it does not know, to ground claims in provided text, and to validate anything high-stakes. OpenAI’s prompt-engineering guide is the same order of work: put the constraints and the reference material in the prompt, then generate. Neither vendor treats “make it perfect” as a control.

Grounded context is a short list, not the thread

The user message should carry only the slots the model must not invent: customer name and what they booked or asked for; date, time, address if needed; who shows up, if that matters; price or range already quoted; rules that already exist (deposit, cancel window, insurance); the one next act (confirm, reschedule, pay, reply).

Card numbers, health details, full charts, and the long thread do not belong in that message. They are not grounding. They are distractors. The smallest set of facts that makes a clear reply is the grounded context.

If the job — the speech act — will not fit in one sentence, do not open the chat. Confirm Tuesday 3pm cleaning and unlock the side gate. Follow up on Monday’s quote with one yes or no. Late, new arrival time, one extra already allowed. The model cannot choose the speech act. It can only write toward the one you named.

The system prompt is not the user message

Keep a short, reusable system prompt for voice. If the tool has no system field, paste the same block at the top of the user message and treat it as frozen: clear, calm, local, respectful; no hype, no slang the team never uses, no long apologies, no “as an AI”; short sentences; first-name sign-off.

That block is a prior. It should not change with today’s job. Today’s job and today’s facts belong in the user message. Mixing them is how a style adjective starts inventing a policy.

When a sent email gets a good reply, fold one line of it into the system prompt. That is a few-shot example, which beats style adjectives. It is retrieval of a known-good output, not a blank-page generation.

Two constrained generations beat a single shot

First user message, in order: system prompt for voice; job in one sentence; grounded facts; “Draft a short email. Do not invent facts. Mark anything missing as [NEED].”

[NEED] is a missing-slot protocol. It is not a vibe. If the model cannot fill a slot from the grounded context, it must surface the gap instead of hallucinating a value. That is the same idea as telling the model it may say it does not know.

Second turn, same thread: “Tighten this. Cut filler. Keep every fact. Put the ask in one sentence. Stay under 120 words.” Two constrained generations beat a single-shot request to “make it perfect.” If a line is wrong, edit that line. Do not ask the model to “sound more human” unless you name the failed phrase.

A person still sends the mail. Check four things: would you say this at the desk; is every price, time, and promise already approved; is there exactly one ask; did the model invent a policy, a discount, or a timeline.

If the email you owe is a chase on an open quote, here's a quotation follow-up email template with when to send each one.

The same draft-then-cut habit works for retention emails too, like this we miss you email template for martial arts schools that goes to parents when a student stops coming to class.

For benefits season, use the same draft-and-approve pattern on an open enrollment email to employees.

If that inbox is Outlook on Microsoft 365, use Outlook draft with Copilot so the first pass stays on the thread.

Unsupervised send is the real risk

Do not grant unsupervised send for refunds, medical advice, legal threats, or angry disputes. Drafting is fine. Those messages still need a person, because a hallucination in those frames is not an awkward sentence. It is a commitment.

Save the emails you actually sent, in folders named by job. Next time, paste one strong example into the prompt and say: “Match this length and directness. Use today’s facts.” Run the same workflow on this week’s real volume — 48-hour quote follow-up, same-day late note, thank-you plus review link, sick-tech reschedule, deposit reminder. Speed is repetition of a constrained loop, not a cleverer first prompt.

Nielsen (1997) measured that people scan. One idea per paragraph is not decoration. It is how the output will actually be read.

If the email you keep putting off is a quote that went quiet, here's an estimate follow-up email for landscaping quotes, with the timing.

If the emails you draft are shipper updates on booked loads, these delivery status update email templates for freight brokers show which facts to pull from your TMS and which update still needs a person.