Prospecting with AI Without Spamming: For 20 Targeted Companies, It Prepares and You Write
By Anis Hammouche·November 7, 2026
Yes, AI can help you prospect without spamming. It is useful for preparation: finding the right companies, understanding what they do, forming a hypothesis about their problem. It becomes a problem as soon as it writes the message in your place and you send it without rereading. For email prospecting toward professionals, the CNIL sets precise rules: the message must relate to the person's activity, say who is writing, and let them refuse simply.
Why spam arrives so fast with AI
A message generated in one second looks like every other message generated in one second. A recipient spots that kind of message quickly. The problem comes from how it is used: you ask "write an email to 500 prospects", and you send whatever comes out. Spam is a message sent to people who did not ask for it, unrelated to their activity, with no simple way to refuse. An opening line is the first sentence of the message, the one that decides whether the person reads on.
The opposite habit works better. You keep AI on the work that needs sorting and reading, and you keep your hands on the work that needs judgment: why this person, why now, and what exactly you say to them.
One point to check before you start: the files you prepare often contain data about people. What you can paste into a tool and what should stay out is covered in what you should never entrust to a public AI.
What AI does well, and what you keep
The table separates what you can delegate from what stays with you.
| Step | AI can | You keep |
|---|---|---|
| Targeting | Suggest criteria (sector, size, tools used), reformat a list | Choose the criteria and check they match your offer |
| Documenting | Summarize the "about" page or a recent article from the company | Check each fact against the original source |
| Finding an angle | Propose three hypotheses about a problem, based on the facts | Keep the one that holds and phrase it yourself |
| Writing | Produce a first draft to correct | Rewrite the opening line and anything that sounds off |
| Sending | Nothing | Reread, send, answer replies |
| Following up | Sort replies (interested, not now, refusal) | Decide what comes next and remove refusals from your lists |
The last row is often forgotten. A refusal should be taken seriously and noted. A tool that follows up with someone who said no will damage your prospecting in one go.
CNIL rules to know before you send
The CNIL distinguishes recipients. For an individual, consent is in principle required beforehand. For a professional, prospecting can rely on legitimate interest, meaning the legal basis that allows writing without prior consent. It requires that the subject of the message relate to the person's profession. The example given is software presented to the IT manager of a company. Selling payroll software to the manager of a clothing shop falls outside that frame.
Three other rules matter every day:
- Generic addresses (contact@, info@, orders@) belong to legal entities. The CNIL states that they are not subject to the same principles.
- Each message must make it possible to know which organization is writing.
- The person must be able to refuse further messages, through a simple means. The CNIL recommends a checkbox left unticked by default when a form collects the address.
A bought list is a set of addresses supplied by a third party. If you use one, make sure each person was told that their address could be used for prospecting. This point is often overlooked, and it must be checked before sending, not after.
I am not a lawyer. For your specific case, the CNIL sheet on prospecting by email is the reference to read. The CNIL page on electronic communications with prospects and customers sets out the same principles in another form. If you are unsure about a bought list or a mixed database, ask a legal professional before sending.
Personalize without inventing
Personalizing does not mean flattering. A good message contains one verifiable fact, read by you in a source, and a question about a plausible problem. The rest is filler.
The trap is asking AI to find those facts. It can invent one, or attribute to a company a project it does not have. The instruction below limits that risk: it supplies the sources and forbids writing the message.
Here is the text of the company's "about" page and its latest article (paste the text below).
1. Give me three precise facts about what they do, without adding any.
2. For each fact, copy the original sentence.
3. Propose two possible questions about a problem they might have.
Do not write any message.
You get a working sheet. You check each original sentence, then write the opening line yourself, in three or four lines, using the fact you verified.
For example, take a craftsperson who targets communication agencies. The agency's "about" page names three real-estate clients. The working sheet keeps that fact and a possible question: "How do you handle your real-estate clients' visuals when a property changes price?" The message starts from that question and the verified fact. It does not say "I love your work" or "I can automate everything for you". If the question does not match a real need, you send nothing.
How to do it, step by step
- Limit the list. Twenty well-chosen companies are worth more than five hundred addresses found at random. Define three criteria before you ask AI anything.
- Prepare the sheets. Use the instruction above, one company at a time. Keep the original sentences in a spreadsheet next to the company name.
- Write the first message by hand. Do not have it generated. Once you have your version, you can ask AI to flag empty phrases, without asking it to rewrite.
- Read it aloud. If a sentence would not work spoken, it does not work written.
- Check the frame. The subject must match the person's job. Your identity and a way to refuse must appear in the message.
- Send in small batches. A few messages a day, each reread, are better than a mass send you cannot correct.
- Note the replies. A refusal is information. Remove the person from the list and do not contact them again with another wording.
Limits to keep in mind
AI produces sentences that sound right. It can be wrong about a figure, a name, a date. It also generalizes: its openings end up resembling each other from one message to the next.
It also does not know the relationship. If you spoke to the person three months ago, if a mutual contact recommended you, if the company has just lost a contract, you are the only one who knows. That information belongs in your message, not in a generic prompt.
For details a model invents about a prospect, the article on the figures and sources AI makes up gives the checking method.
Frequently asked questions
Can you send a prospecting email to a professional without consent? In most cases, yes, if the message relates to the person's activity, if you identify yourself, and if you let them refuse simply. The CNIL sets out the conditions. For a personal address or a bought list, the rules are stricter.
How many messages can you send per day? The CNIL sets no figure for volume. The useful guide remains the quality of replies and the number of refusals. If refusals rise, reduce the volume and tighten your criteria.
What do you do with an address that replied "no"? Remove it from your active list and note it so you do not contact that person again. A refusal cannot be bypassed with another wording.
