Personalising prospecting emails with AI without sounding fake

There are two ways to get personalisation wrong in a prospecting email. The first is not to personalise at all and send the same text to 2,000 people. The second, more recent, is to hand personalisation to an AI with no framework and send 2,000 generic compliments that everyone spots at a glance: "I was impressed by your remarkable track record and your commitment to innovation."

Used well, though, AI lets you do what no sales team has time for: read each prospect's website and pull out one accurate sentence. Here's how we go about it.

What AI does very well

Today's language models are excellent at one specific task: reading a text and pulling out the element most relevant to a given question. On a company's home page, About page or news section, they quickly spot:

• a new office opening, a product launch, a funding round in the news;
• a live job advert for a role related to what you sell;
• a distinctive positioning ("specialists in family businesses", "construction only");
• a specific customer base or region.

They can also judge, from those elements, whether the company matches your target. That's what lead scoring in SmartOutbound does: every contact gets a relevance score, so you can rule out the ones that have no place in the campaign before you write a word.

What AI does badly

Left without instructions, an AI tends to:

• flatter: it congratulates, admires, pays tribute, which sounds fake by the second time;
• generalise: "your innovative company", "your customer-centric approach", phrases that apply to everyone;
• make things up: if the website says nothing interesting, it fills the gap with a plausible but false detail. This is the most dangerous flaw;
• write long: three sentences where one would do;
• adopt a recognisable style: certain turns of phrase, punctuation that is too regular, relentless enthusiasm give a generated text away instantly.

The whole challenge is to break it out of those habits.

Framing the hook: one sentence, one fact

The brief we give the AI comes down to a few rules:

1. one sentence, 15 to 25 words;
2. it rests on a specific, verifiable fact found on the prospect's website;
3. that fact relates to your offer and the problem you solve;
4. no compliments, no gushing adjectives;
5. if nothing relevant turns up, it must not invent anything: better a more general hook about the sector, or no hook at all.

Compare:

Congratulations on your impressive growth and your remarkable commitment to your customers.

I saw you're opening an office in Bordeaux in March and that three consultant roles are already advertised there.

The first could be about any company. The second is about that one only, and it sets up the rest of the email naturally.

Where should AI look?

The company website is the most reliable and easiest source to work from: it is public, it is written by the company itself, and it rarely misrepresents what the business does. The most useful pages are the home page, About, job adverts and news.

Personal information about the recipient (their career, their posts, their private life) needs far more caution. A hook about the company reads as solid homework. A hook about the person can quickly come across as intrusive.

Reviewing at scale

AI writes fast. It sometimes gets things wrong. The only safeguard is review, at least at the start of a campaign. In SmartOutbound, copilot mode shows you each first email before it goes out, with its hook, and you can approve it, fix it or drop the contact. In practice, after fifty or so approvals you have a very clear idea of the quality of the campaign's hooks:

• if almost everything is right, you can approve in batches;
• if several hooks are clumsy, it is often the target or the objective that is badly described, and it is better to fix the brief than to review every message.

Watch in particular for name errors (an AI can misspell a company name it has read from a logo or a headline), out-of-date information (last year's event presented as current) and hooks that are too similar to one another.

The markers that betray an email written by a machine

Your prospects receive more and more generated emails. They have developed a real nose for it. Avoid:

• enthusiastic openings and compliments;
• words no salesperson uses out loud: "unmissable", "a real lever", "constantly evolving";
• reflex three-point lists;
• em dashes used as punctuation in every sentence;
• conclusions that recap what has just been said;
• a generic signature with no first name and no real contact details.

The best test is still to read the email aloud. If you wouldn't say it to a prospect on the phone, don't write it.

AI across the whole sequence

The hook is only part of the job. AI can also draft a first version of the whole sequence from your website and the campaign objective, sort replies (interested, not interested, out of office, wrong contact, unsubscribed) and pick out from an auto-reply the name of a colleague worth contacting. It saves you hours, provided you keep the final decision: you are the one who reviews, approves and sends.

To write a good first email, structure matters more than the tool: our guide to the first email sets it out in detail. To see personalisation at work on your own prospects, create a free account: the trial includes 1,000 prospects processed by AI.