+33 1 75 43 77 01info@ediware.netFree account: 1,000 emails per month FrançaisFR
Talk to an expert

Home›B2B email marketing›Writing emails with AI

Writing B2B emails with AI without losing your voice

In brief: AI can draft a B2B email in ten seconds, and that is exactly the problem: everyone ends up receiving the same one. To make it write like you, it needs a voice sheet, verified facts and a review that always asks the same six questions.

Illustration: Writing B2B emails with AI without losing your voice

In 2025, 18% of French companies with 10 or more employees said they used at least one artificial intelligence technology, up from 10% a year earlier. The figure comes from the ICT survey run by Insee, the French national statistics office, published in July 2026. Of the companies that have taken the step, 43% use generative AI tools for text, and marketing is among their top three uses.

You can see it in your own inbox. Prospecting emails look more alike every month: the same polite opener, the same three-beat paragraph, the same “feel free to get back to me” to close. The tool is not at fault. It writes the average of everything it has read, which is exactly what you get when you give it nothing else.

Prompts are not the subject here. What interests us is a method that holds up when you send every week, as a team, and want each message to sound like the company rather than a language model. It sits inside a B2B email marketing strategy that, for its part, remains yours.

What AI writes well, and what it writes badly

Not every email can be delegated in the same way. The question is not whether AI can write; it can. The question is what it can write without knowing your customer.

Type of email What AI does well What stays with you
Newsletter Formatting existing content, suggesting subheadings, cutting length Choosing the topic, the opinion, the example drawn from your customers
Product announcement Structuring, turning a spec sheet into benefits Dates, prices, what is actually available
Cold prospecting Offering variants of the opener and subject line The reason to contact this company, now
Follow-up Varying the wording from one follow-up to the next The new element that justifies following up
Onboarding and automated emails Writing a first draft of each step The coherence of the sequence and its trigger conditions
Campaign translation Producing a clean version in English, German or Spanish A review by someone who speaks the language

The split is obvious. AI excels at the form of content that already exists. It gets weak as soon as a fact is needed that it cannot know: your price next year, the name of the customer who signed last week, the reason this sales director should reply to you rather than to your competitors.

Hence a fairly simple rule of thumb. The more an email relies on information specific to your company or your recipient, the more the AI’s share shrinks to rewording.

A voice sheet, written once and pasted every time

The first instinct of most teams is to write a new prompt for every email. The result is a warm message on Monday and a very corporate one on Thursday, because two people phrased their request differently. The reader sees one thing only: a company whose voice changes from one send to the next.

The fix fits on one page. A voice sheet, written once, that everyone pastes at the top of every request. It contains:

  • the register: formal “vous” throughout, short sentences, no superlatives;
  • the length: for example 120 words for a prospecting email, 300 for a newsletter;
  • three emails that worked well, copied as they are;
  • banned words, the ones in your sector that ring hollow and the phrases nobody on your team can stand any more;
  • what you never promise: a deadline, a discount, a figure that the sales team has not approved;
  • the signature and the way you sign off.

The three examples weigh more than everything else. A language model imitates far better than it obeys. Ask for a “direct and warm tone” and it will produce its own idea of warmth, not yours. Give it three of your emails and it will pick up the way you start a sentence.

The sheet has another merit, less visible. It forces the team to agree on what its voice is, and quite a few companies discover in the process that they never really had one.

AI only knows what you give it

A language model checks nothing. It produces the most plausible string of words, and a plausible statistic looks very much like a real one. Ask for an email about deliverability without supplying figures and there is a good chance the tool will invent one, complete with a source.

The remedy is a factual brief, kept separate from the voice sheet. The exact offer, with its price and conditions. The proof point to feature, a real customer case or a figure whose origin you know. The event date, the link to the page. And an explicit instruction: no figure, no customer name and no quotation that does not appear in this brief.

The CNIL, the French data protection authority, sums it up in one sentence in its practical guide on AI for very small and small businesses: “The final decision remains human: it is necessary to step back and challenge the answers provided” (our translation). An email announcing a feature that does not exist, or a price two years out of date, costs far more than the time saved by drafting it.

The tics that give away AI-written emails in French

If you send campaigns in France, this section is for you. It is about French, not English: the tics below are what AI produces when it writes in French, and your French recipients spot them more and more easily. Language models learned most of their style from English texts, and it shows. Some phrases are calques, others are quirks of the machine.

What the AI writes Why it sounds off What to write instead
« J’espère que ce message vous trouve bien » A literal calque of “I hope this email finds you well” Nothing: get straight to the point
« Je me permets de revenir vers vous » (I am taking the liberty of coming back to you) A worn-out formula with no content The reason for the follow-up, in the first line
« N’hésitez pas à me contacter » (do not hesitate to contact me) A catch-all closing line One precise question, answered in a single line
« Dans un contexte où… » (in a context where…) An essay-style opening The fact itself, stated directly
A Subject Line With A Capital On Every Word An English convention A capital on the first word only, nothing more
« Simple, rapide et efficace » (simple, fast and effective) Three adjectives, a mechanical rhythm A single adjective, the one that is true
Long dashes in the middle of sentences Punctuation rarely used in everyday French A comma, or two sentences
English quotation marks, question mark stuck to the word English typography French typography wants « » quotes and a space before ? ! : ;

Using “tu” instead of “vous” deserves a mention of its own. Depending on the texts you feed it, AI sometimes slips from the formal to the informal form halfway through a message, especially when your examples mix the two. One line in the voice sheet settles the matter.

The aim is not to hide the tool. It is to stop your email from looking like the one sent by fifty competitors using the same model.

The review, in six questions

Proofreading “to see if it works” is not much use. A useful review asks the same questions, in the same order. Six are enough.

  • Is every fact correct? Prices, dates, feature names, the customer quoted. This is where mistakes cost the most.
  • Does every figure have a source you could give if asked? If not, remove it or reword it.
  • Does the email ask for one thing only? AI loves to add a second link, a second invitation. One request per message.
  • Read aloud, does it sound like you? The quickest test for spotting the sentence nobody on your team would have written.
  • Do the personalisation variables hold up when the field is empty? A “Hello , " gives away an automated send far more surely than an awkward turn of phrase.
  • Does the test send render properly on mobile, and does every link go where it should? A reworded text is often longer than the original, and the layout suffers.

Truth be told, nothing in this list is specific to AI. The volume, though, changes everything: when you write three times faster, you are tempted to review three times less. Deliverability rules have not moved either. Since February 2024, Google has asked senders to keep their spam complaint rate below 0.30%, and a generic email sent in bulk attracts more complaints than a useful one.

Writing in a chatbot or in the sending platform?

Few guides cover the question, yet it changes day-to-day work. Some teams write in a general-purpose assistant and then paste the text into the email editor. Others use the AI built into their sending tool.

Criterion General-purpose assistant AI built into the emailing platform
Contact data The temptation to paste names and companies into the prompt The text is worked on where the data already lives
Layout Copy and paste that often breaks the HTML The content drops into the company’s own template
Brand context To be supplied again in every conversation A previous campaign can serve as the starting point
Translation Text only, to be reformatted The whole campaign, layout included
Freedom of request Total Focused on the email and its presentation

The first point is the most serious. The CNIL repeats it in the same guide: with free consumer tools online, “everything you enter in your requests or share may be used as training data for the AI” (our translation again). Its instruction is clear: do not enter personal data, or anonymise it. Pasting your prospect list into a chatbot to “personalise” your messages is precisely what you should not do. Personalisation goes through your platform’s variables, not through the prompt.

It is with this logic in mind that we designed Ediware’s AI. The assistant works inside the campaign editor: it fits a text into an existing template or a previous campaign, corrects spelling, adjusts the tone, restructures a message using a method such as AIDA and translates the whole campaign. It offers a new version, which you apply or not. We also advise duplicating the campaign before using it, more out of habit than mistrust.

That leaves the AI Act. Its Article 50, applicable since 2 August 2026, requires AI-generated text to be flagged when it is published to inform the public on matters of public interest, unless it has been through human review and someone holds editorial responsibility for it. A commercial email reviewed by your team is therefore, in principle, not concerned. One more reason never to skip the review.

Measure before you generalise

One last point, often forgotten: you cannot know whether AI improves your emails until you have measured it. The team’s impressions are no substitute for a test.

The method is that of a classic A/B test. On a single campaign, one half of the segment receives the version written as usual and the other the version produced with AI and then reviewed. You judge on clicks and, in prospecting, on replies. Not on opens, which have been skewed ever since Apple’s privacy protection started triggering the tracking pixel automatically. And you repeat the test over three or four sends before concluding, because a single test proves nothing.

The AI version may well come out level. That is already good news if it saved you an hour. It may do worse on one specific type of email, cold prospecting most often, and you will then know where to confine it.

AI writes quickly and cleanly. What makes a prospect reply is still what you know about them, and what only you can prove.