Sales email prompts: 8 templates to write emails that actually convert

Sales email prompts: 8 templates to write emails that actually convert

A good sales email prompt does not say "write a prospecting email". It gives the AI what it cannot guess: your target customer, the reason you are reaching out now, a concrete piece of proof, and constraints on length and tone. With those four elements, ChatGPT, Claude or Gemini will give you a usable first draft in a few minutes. Without them, you get the generic email your prospects already delete.

There is no shortage of "50 prompts to close more deals" lists. The trouble is that they all produce the same polite, interchangeable text. This article takes the opposite route: a reusable prompt structure, eight templates for the sales situations SMEs run into most often, a way to have the AI review its own draft, and the legal framework to check before you send anything. For the general principles of writing instructions, our prompt engineering guide for SMEs remains the reference for this series.

Why most AI-written sales emails fail

Because the AI fills the gaps in your prompt with the average of what it has read, and the average sales email is bad. A large language model (LLM) that knows nothing about your prospect writes for everyone, which means for no one.

The symptoms are easy to spot:

• an opening about you ("I'm reaching out because our company...") rather than about the reader;
• claims with no evidence ("innovative solution", "tailored support");
• three or four paragraphs where three sentences would do;
• a vague ask ("feel free to get back to me") that does not actually ask for anything;
• stock phrases that prospects now associate with automated sends.

Switching tools will not fix this. All three major assistants write well; our ChatGPT vs Claude vs Gemini comparison shows the real differences lie in how they fit your existing software. What changes the output is the material you give the model.

The structure of a sales email prompt that works

An effective email prompt has six blocks: role, recipient, trigger, proof, ask and constraints. It is the same logic described in our article on the anatomy of a good prompt, applied to sales.

The block people forget most often is the trigger. Yet it is what separates a prospecting email from spam: an announced hire, a new website, a regulatory change, a trade show where you met. If you have no trigger, the AI cannot invent an honest one, and that is a sign the contact may not be relevant yet.

The proof has to come from you. Ask an AI to "add a convincing statistic" and it will make one up. That is the same mechanism as hallucinations, and in a sales email your signature is on it. Our article on AI hallucinations and how to avoid them covers this risk in detail.

If you write the same kind of email often, put the stable blocks (role, offer, tone, banned phrases) in the tool's custom instructions or a dedicated project, and only change the recipient, trigger and proof each time. The difference between those two levels is explained in system prompt vs user prompt.

8 ready-to-adapt prompts for your sales emails

Copy each template as is, then fill in the brackets with your real information. Never leave a bracket empty: that is exactly where the AI starts making things up.

1. First prospecting email

You are [job title] at [company], which offers [one-sentence offer]. Write a first prospecting email to [recipient's job title] at [type of company, size, sector]. Reason for reaching out: [recent, verifiable fact]. Proof to use: [result achieved for a comparable client, with a number]. Goal: get [one action, for example a 15-minute call]. Constraints: 80 to 110 words, formal tone, first sentence about the recipient, no superlatives, no "I hope this email finds you well". Suggest 3 subject lines of 4 to 7 words.

2. Follow-up after no reply

Here is the email I sent [number] days ago, with no reply: [paste the email]. Write a follow-up of 50 words max that brings something new: [resource, figure, specific question]. Do not rephrase the first email and do not write "just circling back".

3. Final follow-up (break-up email)

Write a short final email (40 words max) to close the sequence with [prospect]. Respectful tone, no guilt-tripping. Leave a clear door open: [what the prospect can do later]. End with a closed question that can be answered in one word.

4. After a trade show or event

I met [name, job title] at [event] on [date]. We talked about [specific topic]. Write a follow-up email that recalls that specific point in the first sentence, suggests [next step] and stays under 100 words. Do not invent any detail of the conversation that I have not given you.

5. Reactivating a dormant customer

[Customer] last ordered [product or service] in [period]. Since then, we have [real new development]. Write a re-engagement email that does not read like a sales push: one question about their current situation, then the news in one sentence. 90 words max, direct tone.

6. Meeting recap

Here are my notes from the meeting with [customer]: [paste notes]. Write a structured recap email: the need in their own words, points of agreement, open questions, next step with a date. Flag in brackets any part of my notes that seems ambiguous instead of filling it in.

7. Handling a price objection

The prospect [job title, company] replies: "[paste the objection]". Our price is [amount] for [scope]. Factual arguments available: [measured gains, comparison with the alternative, possible funding]. Write a reply that acknowledges the objection without apologising, gives one argument backed by a number and offers a concrete option (smaller scope, staggered payments, trial). 120 words max.

8. Asking a happy customer for a referral

[Customer] achieved [result] through [project] completed in [month]. Write a short email that thanks them with one specific detail of the project, then asks whether they know one or two companies facing [problem]. No pressure, 80 words max.

These templates work even better if you add one or two emails that have already got replies for you, with the instruction "follow the tone and length, not the content". That is the principle of few-shot prompting, which is very effective at getting the model to adopt your style rather than its own.

Have the AI review the email before you send it

The second pass matters as much as the first: ask the AI to critique its own email against a precise grid, then rewrite it. A model judging a text against explicit criteria will catch flaws it did not avoid while writing.

A review prompt that works well:

You are the recipient of this email: [job title, context]. You receive 40 sales emails a week. Assess the email below on 5 criteria, scoring each from 1 to 5 with a one-sentence justification: does the first sentence concern me; is the proof credible; is the ask clear and easy; is the length right; does it sound like a person wrote it. Then suggest a corrected version that fixes the two weakest criteria.

Asking the AI to spell out its reasoning criterion by criterion, rather than jumping straight to a "better version", leads to more relevant corrections. It is the idea behind chain of thought prompting, applied to editing.

Keep the final call. Read the email out loud: if a sentence does not sound like you, rewrite it yourself. The prospect who calls back will be talking to you, not to the tool.

What can go into the prompt, and what should stay out

Public, professional information about a prospect can go into an AI tool used with a business account; bulk exports from your CRM have no place in a free personal account. The issue is not AI as such, it is where the data ends up.

In practice, for an SME:

• No problem: your offer, anonymised case studies, the recipient's sector and job title, a public fact (new site, announced hire).
• With a properly configured business account: the contact's name, your email history, meeting notes.
• Avoid: a full prospect file export, consumer data, sensitive information about a customer, confidential pricing negotiated with other clients.

Check in your tool's settings that conversations are not reused for training, and that usage rules are written into your AI policy. Our article on AI and data confidentiality covers the settings for the main tools.

If your sales team works in Microsoft 365, Copilot can draw on Outlook history without copy-pasting; our Microsoft 365 Copilot beginner's guide explains how to start. On Google Workspace, the same logic applies with Gemini in Gmail, described in our Gemini use cases for SMEs.

The legal framework for B2B email prospecting in France

For B2B emails, the French data protection authority (CNIL) does not always require prior consent, provided three conditions are met together: the message relates to the recipient's professional activity, the recipient is told where their data came from and what the message is for, and they can easily opt out of further messages. These rules come from article L.34-5 of the French Postal and Electronic Communications Code and from the GDPR.

In practice, every generated email must keep three elements that the AI tends to drop when you ask it to "make it shorter":

1. A clear identification of the sender and the company.
2. A sentence saying where the contact came from (trade show, company website, referral).
3. A simple way to stop receiving messages, then honoured through a suppression list.

Add these to your prompt constraints rather than relying on proofreading. For example: "Always end with the line: If you would rather not hear from us again, just reply STOP."

A recent development: on 16 October 2026 the CNIL published guidance on reconstructing business email addresses of the firstname.lastname@company kind. It accepts that legitimate interest may apply on a case-by-case basis, provided in particular that the person could reasonably expect to be contacted this way, that the address is only kept if an actual first contact takes place within a reasonable time, and that any objection leads to immediate exclusion. Asking an AI to "guess" addresses in bulk does not fit that framework.

For the wider data protection rules, see our GDPR glossary entry. This summary is not a substitute for legal advice on your own situation.

Measure what actually changes reply rates

A prompt is judged on reply rate, not on whether the email feels "well written". Without measurement, you will be changing wording at random.

A simple method that works at SME volumes:

• Test one variable at a time: the subject line, the first sentence or the final ask.
• Send each version to at least 50 comparable contacts before drawing conclusions. Below that, one or two replies are enough to skew the picture.
• Track reply rate rather than open rate, which mail privacy features now distort.
• Record the exact prompt that produced the winning version in a shared document. That prompt, not the email, is the asset you keep.

An example test log: version A with the subject "Your second production line", version B with "Unplanned stoppages on line 2". Same body, same audience, same sending week. The gap in replies tells you which angle resonates with your prospects, and you feed that back into the constraints block of your prompt.

On terminology, the conversion rate of a sequence is calculated on meetings booked, not just on replies. A "not interested" reply is useful information, but it is not a conversion.

Frequently asked questions

Which AI tool should we use to write sales emails?

The one your sales team already has in its daily tools, with a business account. ChatGPT is the most versatile, Claude is often preferred for a more natural tone, and Gemini and Copilot have the advantage of reading the mailbox directly. Our pages on ChatGPT use cases for SMEs and Claude use cases for SMEs give examples by role.

Can prospects tell an email was written by AI?

What they spot is generic email, whether written by an AI or by someone in a hurry. An email that mentions a specific fact about their company, verifiable proof and a simple ask is indistinguishable from one written by hand.

How many follow-ups should we send?

Two or three in total, several days apart, each with something new. Beyond that, the damage to your reputation outweighs the likely gain. Honour any opt-out request immediately.

Can we fully automate sending AI-generated emails?

Technically yes, with an automation tool connected to your CRM and an AI model. We recommend keeping a human approval step until your prompts have been tested over several hundred sends, because a personalisation error will otherwise repeat at scale.

Should we give the AI our company knowledge base?

For one-off emails, pasting the two or three relevant facts is enough. If your sales team needs to draw on a large catalogue or technical sheets, a document retrieval approach becomes more suitable; see RAG vs prompting.

Does an AI-generated prospecting email have to say so?

No general rule currently requires you to state that a sales email was drafted with AI help, as long as it is reviewed and sent under the responsibility of an identified person. The prospecting rules above apply in every case. This may change with the European guidelines on the transparency obligations in article 50 of the AI Act, so keep an eye on their publication.

How GrowthPerf can help

The best sales prompts are the ones your team builds around its own customers. GrowthPerf, a Qualiopi-certified training provider, trains sales teams in SMEs and non-profits to write, test and share their prompts: a context sheet for your offer, a library of validated templates, a review grid and compliance rules suited to your prospect file.

Our Operational AI course includes hands-on workshops on your real emails. For a programme focused on prompting techniques, see our page on prompt engineering training. In France, these courses can be funded through your OPCO.

To go further on method, return to the complete prompt engineering guide for SMEs.

Is your sales team already using AI for emails, with no shared method? Book a free 30-minute audit: we review your current emails, reply rates and the prompts in use, and tell you what is worth standardising.