B2B prospecting with generative AI: an assistant, not an autopilot

Outbound prospecting has a poor reputation in SMEs. It was tried once, it produced nothing, and the conclusion drawn was that it does not work in this line of business. In reality, it rarely stops because it fails. It stops because nobody has time to prepare it a second time. Generative artificial intelligence does not change what good prospecting is. It changes what it costs to set up, and that is what finally makes it repeatable.

Why outbound prospecting stops

Setting up a campaign takes a series of tasks everyone recognises: defining the target, building a clean list, checking the contacts, writing a message sequence that does not read like a mail shot, scheduling it, tracking the replies and entering them into the CRM. Added together, that is several days of work for a single campaign.

In a twenty-person company, those days do not exist. The campaign goes out once, with an approximate list and a generic message. It produces little. And because preparing it was costly, nobody launches the second one, the one that would have corrected the aim.

Yet outbound prospecting only produces results through iteration. You test a segment, see who replies, adjust the message, and start again. As long as each iteration costs three days, the loop never gets going.

What the assistant changes

The reference build was carried out for a twenty-person research consultancy. Its principle: have an assistant hold a dialogue with the prospecting database, so that targeting is expressed in business language rather than in combinations of filters on a screen.

Targeting expressed the way you would to a colleague

You describe who you are looking for the way you would to someone on the team: the type of company, its size, its region, the role of the contact, and above all the signals that make the moment relevant, a site opening, a hire, a change of management. The assistant translates that description into a search of the prospecting database, builds the list, enriches it, and reports back on what it found and what it left out.

This is the most underrated gain. Prospecting interfaces force you to think in filters, and filters cannot express what makes a prospect interesting. Language can.

Messages written from the real offer

The assistant then drafts the sequence from the company’s real offer and the context of each recipient. Not a template with the company name as a variable, but a message that explains why this particular person is being written to, now.

This is where the quality of the positioning shows. A model cannot write a relevant message for a vague offer. If the company cannot say precisely what it delivers and to whom, the assistant will produce messages as vague as the brief it is given.

Execution and follow-up

The campaign is then run and tracked automatically. Replies flow into the CRM, into the same pipeline as inbound enquiries, under the same follow-up rules. A positive reply becomes a deal, and what follows can pick up the arrangement described in the guide from meeting transcript to sent proposal.

What stays with the human

The assistant prepares, it does not decide. The sales manager keeps control of the two things that matter: who gets contacted, and what they are told. They approve the list before sending, read through the sequences, and can stop a campaign in progress at any time.

This is a deliberate limit, and we never lift it. An outbound campaign puts the company’s image on the line with people who do not know it yet. A badly calibrated message sent to three hundred executives in the same sector costs more than not running a campaign at all, because those executives talk to one another. What gets automated is the preparation work, not the judgement.

This is also what sets this approach apart from tools that promise fully autonomous prospecting. They send a great deal, they damage the sending domains, and they end up harming the deliverability of the company’s ordinary email.

The rules to follow

In France, business-to-business email prospecting is allowed under precise conditions that must be applied from the very first campaign:

• the message is addressed to a person in the context of their role, and its subject relates to that role;
• the sender is clearly identified;
• every message offers a simple way to opt out, and that refusal is honoured from the first request, across every tool;
• prospecting data is kept only for as long as necessary.

These rules also make good commercial sense. A message that speaks to the recipient’s own work, signed by a real person, and that makes it easy to leave, gets better reply rates than a mass send. None of the above is legal advice, and your usual adviser can check the arrangement before it launches.

The prerequisites

Three conditions decide success, well before the technology does.

A clear offer and an identifiable segment. If you cannot say in two sentences who you are addressing and what you bring them, no tool will make up for it. You have to start with positioning, as we did before rebuilding the same consultancy’s acquisition.

An up-to-date CRM. Prospecting feeds the same pipeline as everything else. If the CRM is half empty, replies get lost.

Someone to answer. A campaign that produces replies nobody handles is a straight loss, compounded by damage to the company’s image.

What actually changes

The gain is not sending more messages. It is that a campaign can be set up in hours instead of days. It becomes achievable, then repeatable, and the learning loop starts turning: each campaign learns something from the one before.

This build is part of the sales and prospecting lever. If your outbound prospecting stopped for lack of time, it is often one of the first useful systems that can go into production.