A client email is a poor place to discover that an AI tool has guessed wrong. The stakes are small in any single message — a wrong date, a promise you did not mean to make, a tone that reads colder than you intended — but they add up, because the client reads the email as your voice, not the tool’s.
The fix is not a longer prompt. It is a short, repeatable checklist you run before you ask the tool anything, and again before you send what it gives back. Use it for the five email jobs most small businesses repeat: replying to an enquiry, following up, explaining a delay, summarising a proposal, and confirming next steps.
1. Give the context the tool cannot guess
An AI tool does not know who the client is, what you already agreed, or where the conversation stands. If you leave that out, it will fill the gaps with plausible-sounding detail — which is exactly how wrong dates and invented commitments get into a draft.
Before writing the prompt, note four things:
- Who is this for? A new enquiry, an existing client, or someone you have already had a difficult conversation with? The same facts need a different opening in each case.
- What has already happened? One sentence: what the client asked, what you replied, and what is still open.
- What is this email for? One outcome only — answer a question, book a call, confirm a scope, or explain a delay. An email trying to do three jobs usually does none of them well.
- What must not change? Dates, prices, scope, and names that are already agreed. State them explicitly so the draft does not rewrite them.
A prompt that starts with those four points is already better than a clever prompt that starts with none of them. For the basics, see How to Write Better AI Prompts.
2. State the task, then the tone, then the constraints
Order matters. Put the task first, so the tool knows what kind of email it is writing; then the tone, so it knows how to sound; then the constraints, so it knows what to leave out.
A workable structure looks like this:
- Task: “Draft a reply to a client enquiry about [service], answering [their question].”
- Tone: “Professional and warm, not salesy. Short sentences. No jargon.”
- Constraints: “Keep it under 150 words. Do not promise a delivery date. End with one clear question: what time suits them for a call next week?”
- Facts to use: Paste only the facts the email needs, in your own words.
Tone deserves a concrete instruction, not an adjective on its own. “Friendly” can produce anything from casual to over-familiar. “Professional and warm, short sentences, no exclamation marks” gives the tool something it can actually follow — and gives you something you can check in the draft.
3. Keep client details out of the prompt
This is the checklist item most often skipped, and the one with the most consequence. Do not paste a client’s private details, financial information, contracts, or unreleased work into a public AI tool unless you have checked that tool’s data policy and are satisfied with where the data is stored and for how long. If in doubt, anonymise the example: “a client in retail” instead of a name, “last Tuesday” instead of a dated thread, and the substance of the question instead of the forwarded email.
The same rule applies in reverse. If a draft comes back containing a detail you did not supply, treat that as a warning, not a convenience. The tool has guessed, and the guess needs to be removed before sending, however plausible it reads.
4. Check the draft against the original question
Before editing for style, check for accuracy. Read the client’s message again, then read the draft, and ask:
- Does it answer the question that was actually asked, not a nearby question?
- Are all dates, names, and numbers the ones you supplied?
- Does it make any promise — a date, a price, a deliverable — that you did not authorise?
- Is there exactly one next step, and is it clear who takes it?
Most AI email drafts fail on the third question, not the first. A sentence like “We can have this ready by Friday” reads as helpful until you realise Friday was never agreed. Delete any commitment you did not intend to make, rather than softening it and hoping the client reads it softly too.
5. Five prompts you can copy and adapt
Use these as starting points, not as finished prompts. Fill in the bracketed facts from your own notes each time; a prompt reused word-for-word across clients is how the wrong detail ends up in the right template.
Replying to an enquiry. “Draft a reply to a new enquiry from [type of client] asking about [service]. Context: [one sentence on what they asked]. Tone: professional and warm, short sentences. Under 150 words. Answer their question directly, do not list services they did not ask about, and end by asking what time suits them for a brief call.”
Following up, without chasing. “Draft a short follow-up to a client who has not replied in [number] days about [topic]. Tone: polite, no pressure, one reminder of the decision or information still needed. Do not imply they have missed a deadline unless one was agreed. End with one question.”
Explaining a delay. “Draft an email explaining that [deliverable] will be [number] days later than planned because [brief, honest reason]. Tone: direct and apologetic, without over-explaining. State the new date, what is already done, and what happens next. Do not blame a third party.”
Summarising a proposal. “Summarise this proposal in email form: [scope in your own words], [agreed price or ‘price as discussed’], [next step]. Keep it under 200 words, use short paragraphs, and do not add benefits or deliverables that are not in the scope.”
Confirming next steps. “Draft a confirmation email after a call with [client type]. It should list: what was agreed, who does what, and by when — only the items I list here: [items]. Tone: clear and brief. If anything I listed is ambiguous, flag it instead of resolving it.”
In each case, the bracketed part is the work. The prompt template is the easy part; supplying accurate facts is what makes the email trustworthy.
6. The human review, every time
No client email should go out unchecked, however good the draft looks. This is the same rule as for automation generally: keep a human approval step before anything leaves the business, at least until a workflow has proved itself over many runs — and for client email, that step never really goes away, because every client and every situation is different.
A two-minute review covers it: read the draft aloud, check the names and dates, check the promise, check the next step, and check that it sounds like you. If you would not have written a sentence yourself, rewrite it or delete it. Clients forgive a plain email far more readily than a polished one that gets a fact wrong.
A one-page checklist to keep beside the keyboard
- ☐ One outcome for this email, stated in a sentence
- ☐ Context supplied: who, what has happened, what is still open
- ☐ Task, tone, and constraints stated in that order
- ☐ Only the facts the email needs, in your own words — no private client data pasted in
- ☐ Draft checked against the original question, not just read for fluency
- ☐ No promise, date, price, or deliverable the tool added on its own
- ☐ One clear next step, with an owner
- ☐ Read aloud once before sending
Used that way, an AI tool does not write your client emails for you. It gives you a first draft faster, and a checklist that stops the draft from becoming the email before you have checked it. That is the right division of labour: the tool handles the blank page, and you handle the judgement.
FAQs
Do I need a different prompt for every client?
No — reuse the structure, but refill the facts every time. The template stays the same; the context, dates, and commitments change with each client, and those are the parts that must never be carried over from a previous email.
Why do AI drafts sometimes invent details?
Because a prompt with gaps reads as an invitation to fill them. The more context you supply — who, what happened, what is agreed — the less the tool has to guess, and the fewer invented details you have to catch in review.
Should I let the tool send emails automatically?
Not for client email. Automatic sending suits internal, checkable jobs such as filing or drafting a report for your own review. Anything that reaches a client should pass a human check first, because tone, commitments, and facts are judgement calls, not formatting.
What if the draft sounds nothing like me?
Give the tone instruction in concrete terms — sentence length, formality, what to avoid — and keep one or two of your own sentences in the email, such as the opening or the closing question. A draft that is edited in your voice is more trustworthy than a fluent draft in nobody’s voice.
Leave a Reply