How to use AI to find customers. AI reads in an hour what takes a week by hand; the rules for writing to a company are the same as before.
Guides on using AI to find customers list tools and skip the question that decides everything: who you are allowed to write to. Here is the five-step sequence, with the prompts and the rules.
With AI you find customers in five steps: ideal customer, sources of names, checking the signal, first message, human approval. AI shortens the first four. The fifth stays with a person.
The rules for making contact do not change when a machine does the work. On 10 November 2026 the EDPB, the board of European data protection authorities, adopted the final version of its guidelines on legitimate interest: for email and messaging apps, unsolicited marketing requires prior consent.
In Italy consent applies to companies too, under Article 130 of the Privacy Code, and data found online does not become usable for promotion because it is public.
The sequence can be downloaded below in Word and PDF, with the four prompts, the approval log and the outline of the privacy notice, with no email required.
Written by: Mattia Esposito, AI and marketing consultant, founder of Itria AI. From Puglia, based in Turin since 2023, he works remotely in Italy and abroad. LinkedIn profile.
The five-step sequence, to download
A document in four parts: the five steps, with what AI does, what you check and the rule; four prompts to copy into ChatGPT or Claude; the approval log; the outline of the notice for people you find through public sources. The Word file is for filling in, the PDF for printing and keeping beside the screen.

| File | Contents | Link |
|---|---|---|
| Wordto fill in, in Italian | The five steps, the prompts with blanks to fill, a ten-row approval log and the notice in seven fields. | |
| PDFto print, in Italian | The same document, to keep at hand while you work with the assistant. |
| Step | What AI does | What you check | The rule |
|---|---|---|---|
| 1 · Ideal customer | Writes the profile from your description, with no names. | That it resembles your best real customers. | No personal data in the prompt. |
| 2 · Sources of names | Suggests public lists and reads websites. | That the source exists, who runs it, the date. | The platform's terms; for a person's name, a legal basis and a privacy notice. |
| 3 · Checking | Says yes or no and copies the sentence that proves it. | That the sentence is on the page, and recent. | An old signal, or one not found, counts as no. |
| 4 · First message | Writes three drafts from the checked sentence. | The opening line, the channel, the basis for contact. | Email and messages: consent. Phone: the Italian opt-out register. |
| 5 · Approval | Nothing. | A person reads and signs off, with name and date. | Anyone who objects comes off the list (GDPR, Art. 21). |
In September I took 249 websites from the ICE showcase, the public agency that presents Italian companies to foreign buyers. I had set myself the task in one line: “The mission: do what a buyer does. Go in, open a product, work out how to buy.” I told the story in my post of 30 September.
One of the links, belonging to a company in Puglia, led to a ranking of online casinos. The domain had expired and passed to someone else, and the showcase still pointed to it. A list goes stale even when a public body keeps it, and an assistant working from the list will not notice until somebody opens the page.
The rest I counted page by page. Of the 121 food and wine producers, 120 had a way to be contacted; 19 offered a technical data sheet to download, and only one set out the terms for ordering. Anyone can see facts like these by opening the site, and a signal you can check is built from facts like these.
The case of the 389 profiles gathered in an hour, with the rules around it, is in the article Is scraping LinkedIn legal?.
249 websites, one casino. Taken from the ICE showcase and read the way a buyer reads them: one led to a casino ranking because the domain had expired. Post of 30 September 2026.
19 out of 121. The food and wine producers offering a technical data sheet to download. One set out the terms for ordering. The full data is in the study of food producers' websites.
142 websites, 61 with no door. Reading the websites of 142 Italian companies that also sell abroad, 61 had no real way for a foreign buyer to get in touch. Post of 7 September 2026.
389, 30, 3. Twenty-two searches, 389 profiles, 30 to read one by one, 3 with the problem written in public. Told in my post of 15 September 2026.
How to use AI to find customers: what the machine does and what stays with you
AI shortens writing the profile, finding sources, reading pages and the first draft; choosing whom to contact and pressing send stay with a person. This split holds up both the results and the rules. The full sequence of B2B lead generation, from volume to conversation, is in the main guide of this section.
The person at the end also matters for commercial reasons. Forrester, in its survey of almost 18,000 B2B buyers presented in January 2026, writes that 94% use AI when they buy, and then check its answers with people they trust. A message that sounds generated lands with a reader who already uses AI.
Step 1: the ideal customer, written with ChatGPT or Claude
The first prompt turns what you know about your best customers into a five-line profile: sector, size, area, the problem visible from outside and where you can read it. The last line matters most, because it tells you what to look for in the steps that follow. Without it, AI hands you back a list of sectors.
I sell [product or service] to [type of company]. My three best customers are [description, no names]. Write the ideal customer profile in five lines: sector, size, area, the problem visible from outside and where you can read it. Mark everything you infer with TO CHECK.
Keep customer names and contact details out of the prompt: describe the companies, not the people. The rules on what to paste into an assistant, and where the data ends up, are on the page about ChatGPT, the GDPR and the Cloud Act. If the profile starts from a market you do not know, begin with market research with AI.
Step 2: where the names come from, including B2B customers abroad
The best sources are the ones where a company has chosen to be found: exhibitor lists, trade associations, public databases of business partners, the company website itself. AI helps you list them and read them. Write down each source with who runs it and the date, because step 3 starts from there.
| Source | What you will find | The rule to check |
|---|---|---|
| Company website | Products, markets, job adverts, general contact details. | A company info@ address and a firstname.lastname@ address are two different things: the second is personal data. |
| Trade fair exhibitor list | Companies looking for a market in those months. | The terms of use of the list, set by the organiser. |
| Enterprise Europe Network | Company profiles with “Business Offer” and “Business Request” filters, by size and type of agreement. | Contact goes through the network, not a direct email. |
| ICE, foreign customer and partner search | Local operators selected by the overseas offices to match the company's profile. | A service on request and paid: ICE opens the contact. |
People and companies, with the platform's own searches and filters. | The contract bans scraping and automated activity (§8.2). | |
| Bought contact databases | Names, roles, emails and mobile numbers, ready made. | A supplier's mistake does not clear the sender (Garante, 14 May 2026). |
For foreign buyers, ICE describes its service as finding “foreign customers and partners potentially interested in starting business contacts with your company”. The rules for the first contact are built in, because the agency makes the introduction. Other routes for exporters are in the guide to finding foreign importers for free.
Step 3: checking, because AI makes things up with confidence
An assistant can tell you a company “is looking for a supplier” when no page says so. These are called hallucinations, and in this work they are expensive: a first message built on a false fact ends the conversation before it starts. The defence is the technique Anthropic recommends in its guide to reducing hallucinations: give the model permission to say it does not know, and ask it for word-for-word quotes before the answer.
Here is the text of the page [paste the text]. Is the signal [the criterion, in one sentence] there? Answer yes or no and copy the exact sentence that proves it. If the sentence is not there, answer no.
Then open the page and look for the sentence. If it is there, record the link, the date and who checked it in the log. If it is a year old, it counts as no: an expired job advert or a domain that has changed hands, like the casino one, describe a company that no longer exists. Pasting the text instead of the link also stops the assistant reading an old version of the page.
How to use AI to find customers on LinkedIn
On LinkedIn, AI can help you write the customer profile and the draft message; it cannot collect profiles on your behalf. The User Agreement, at section 8.2, bans software, crawlers, plug-ins and bots that copy data, and the page “Prohibited software and extensions” extends the ban to extensions that automate activity on the site.
The ban is also a privacy matter. On 5 December 2024 the French regulator, the CNIL, fined KASPR 240,000 euros: a Chrome extension that collected contact details from LinkedIn profiles, including details members had made visible only to their connections. The three questions that decide whether collection is lawful are in the article on scraping LinkedIn.
What the platform allows remains: search with its filters and write in person, one profile at a time. For the Garante (14 May 2026), accepting a connection request is not consent to receive promotions. How to build the profile a buyer opens after your message is in the guide to personal branding on LinkedIn.
Using ChatGPT to find leads: what to ask and what not to ask
Ask ChatGPT for profiles, sources, readings of pages and drafts; do not ask it to “find me ten customers with their email”. In the first case it works on public company texts. In the second it produces names and addresses you cannot verify and have no basis to use. This is the difference between reading a company and processing a person's data.
Write three opening lines for a message to this company, starting from this sentence: [checked sentence]. No more than 30 words, a small question to close, no generic compliments.
Of the three drafts you keep one, and usually rewrite it. An opening line works when the reader recognises the fact as theirs. Examples of copy written with AI and corrected by hand are in the prompts for email marketing; getting the company's tone into the assistant is covered in the guide to brand tone of voice in ChatGPT.
What the law says: public data, legitimate interest, consent
Finding and reading a company is free; processing a person's name and contact details needs a legal basis; writing to them to sell, in Italy, needs consent, even when they work for a company. These are three separate things, and AI changes none of them. The Garante put it this way on 17 April 2026: data online is “not freely usable for promotional purposes simply because it has been made public”.
On 10 November 2026 the EDPB adopted version 2.0 of Guidelines 1/2024 on legitimate interest. At §108 it writes that direct marketing can be a legitimate interest, and adds: “This does not mean that direct marketing always constitutes a legitimate interest”. At §113, for email, SMS and messaging apps, unsolicited marketing requires prior consent.
At §116 the guidelines note that national laws can require consent towards companies too, “for example with respect to direct marketing towards professionals”. Italy does: Article 130 of the Privacy Code protects the “contraente”, the subscriber, and the term includes companies. Legitimate interest can support building the list, with a three-part test: a legitimate interest, necessity, and the balancing of rights.
People added to a list built from public sources must be told, under Article 14 GDPR: the source of the data, the legal basis and their rights, at the latest at the first communication. The right to object to marketing is unconditional (Article 21). The Word file has the outline in seven fields.
There is also an open case on contact databases: a July 2026 decision on a supplier of professional contact data has been removed from the Garante's website under an order of the Court of Rome, and the proceedings are pending.
| Topic | Italy | European Union | United Kingdom | United States |
|---|---|---|---|---|
| A list of people built with AI | Legal basis and a notice stating the source (GDPR, Art. 14). | Legitimate interest possible, with the three-part test (EDPB, 2026). | UK GDPR, same test (ICO). | No general federal law; individual state laws, not covered here. |
| Email or message to a company | Consent, even towards companies (Art. 130; Garante, 2026). | Consent towards individuals (EDPB, §113); for companies each country decides (§116). | To companies without consent; sole traders treated as individuals (PECR, ICO). | Opt-out, with an unsubscribe in every message (CAN-SPAM). |
| Phone calls with a live operator | The opt-out register, which also covers business numbers. | National rules. | TPS and CTPS lists. | B2B largely exempt from the Telemarketing Sales Rule. |
| Bought or enriched contacts | A supplier's mistake does not clear the sender (Garante, 14 May 2026). | CNIL against KASPR: 240,000 euros (2024). | The rules in the rows above apply. | The rules in the rows above apply. |
| LinkedIn’s terms | Scraping and automation banned (§8.2). | The same. | The same. | The same. |
The email rules, one by one, are in the guide to B2B cold email; the phone rules in the guide to B2B telemarketing and the Italian opt-out register. A voice agent that calls with no operator falls under paragraph 1 of Article 130, on automated systems, which requires consent.
The first message and human approval
No law requires, as such, that a person approve every message before it is sent; this rule is what lets you keep all the others. The approver checks that the signal is real, that the channel has a basis, that the notice is ready and that the name is not on the list of people who have objected. This is called human in the loop.
On a “written with AI” label, the AI Act separates two cases. A sales email written with AI and sent by a person to a company is not among the Article 50 cases. A chatbot talking to customers is, and from 2 August 2026 it must make clear it is an AI. The details are in the guide to the AI Act and generated content.
The last column of the log is the outcome. Anyone who says no, or asks not to be contacted again, comes off the list and stays off at the next AI search. Before the first email, also check that it arrives: SPF, DKIM and DMARC decide whether it lands in the inbox or in spam. The other guides in this section are in the index of marketing for SMEs.
Questions and answers
How do you find customers with AI?
In five steps: you write the ideal customer with AI, choose public sources for the names, have AI check the signal by copying the sentence that proves it, have it draft the first message, and a person approves the draft before it goes out. AI shortens the reading; the rules for making contact stay those of the GDPR and, in Italy, the Privacy Code.
Can I write to the companies AI finds on public websites and profiles?
In Italy a promotional email or message needs consent, even when the recipient is a company (Article 130 of the Privacy Code). On 17 April 2026 the Garante, the Italian data protection authority, wrote that data published online does not become usable for promotion simply because it is public.
Is legitimate interest enough for B2B marketing?
For email, SMS and messaging apps, no: the EDPB guidelines of 10 November 2026, at paragraph 113, require prior consent, and legitimate interest cannot be used there. Legitimate interest can support building the list, with the three-part test and the notice required by Article 14 GDPR.
Do I have to say an email was written with AI?
The AI Act does not require it for a sales email written with AI and sent by a person to a company. The Article 50 obligation covers, among other cases, chatbots talking to people, which from 2 August 2026 must make clear they are an AI.
Are there free sources for finding customers abroad?
The Partnering Opportunities database of the Enterprise Europe Network, run for the European Commission, lists companies looking for business partners, with filters for type of agreement and company size. The foreign customer and partner search offered by ICE, the Italian Trade Agency, is on request and paid.
Notes on sources
- European Data Protection Board, Guidelines 1/2024 on processing of personal data based on Article 6(1)(f) GDPR, version 2.0, adopted on 10 November 2026, §108, §113, §116, §118-120 (consultation page); source of the quoted sentence. Read on 12 November 2026.
- Garante per la protezione dei dati personali, decisions of 17 April 2026, doc. web 10252188, source of the quoted sentence (our translation), and of 14 May 2026, doc. web 10262105; guidelines on spam of 4 July 2013. The decision of 14 July 2026 (doc. web 10275035) appears as removed from the Garante's website under the interim order of the Court of Rome of 8 September 2026; only its status is reported here. Read on 12 November 2026.
- Italian Privacy Code, Article 130 (Normattiva, current text); GDPR, Articles 14 and 21. Read on 12 November 2026.
- CNIL, 240,000 euro fine on KASPR, 5 December 2024. LinkedIn, User Agreement, effective 3 November 2025, §8.2, and Prohibited software and extensions. Read on 12 November 2026.
- ICE Agenzia, Ricerca clienti e partner esteri, source of the quoted sentence (our translation); Enterprise Europe Network, Partnering opportunities. Read on 12 November 2026.
- Anthropic, Reduce hallucinations, official developer guide: let the model admit uncertainty, and ask for word-for-word quotes before the task. Read on 12 November 2026.
- Forrester, The State Of Business Buying, 2026, 21 January 2026: a survey of almost 18,000 B2B buyers worldwide. A research firm; the full report is for clients only. Read on 10 November 2026.
- United Kingdom: ICO, Business-to-business marketing. United States: Federal Trade Commission, CAN-SPAM Act: A Compliance Guide for Business. The Italian opt-out register and the Telemarketing Sales Rule are covered, with sources, in the guide to B2B telemarketing. Read on 12 November 2026.
- Mattia Esposito's figures come from his LinkedIn posts of 30 September, 7 September and 15 September 2026, and from the study of food producers' websites.
AI finds a hundred companies in an hour. Which ones deserve a conversation is decided by reading. Positioning is built starting from the market.
Mattia Esposito, who wrote this guide, also works with Itria on positioning: in-depth market analysis to find where competitors are exposed and take that ground. He did it on his own site first. In August, asked who had founded itria.io, ChatGPT couldn’t find him. From 11 October to 7 November 2026 itria.io pages appeared 13,803 times in Google and Bing results and 3,818 times in AI answers: Google’s AI Overviews and AI Mode, and Microsoft Copilot.