GuidesAI at work
AI prospecting: what an agent does alone, what the salesperson keeps
AI prospecting in B2B: the agent targets, enriches, drafts and follows up, the salesperson chooses and approves. The French rules on email, phone and LinkedIn.
Contents
- What can an AI prospecting agent do, task by task?
- What does the salesperson keep in AI prospecting?
- How is an agent different from a prospecting automation tool?
- Does an AI SDR replace a salesperson?
- What rules govern B2B email prospecting in France?
- What does the GDPR change when data comes from a list or an enrichment tool?
- Can you prospect with data taken from LinkedIn?
- Can you prospect by phone in France in 2026?
- What does Balt do in prospecting?
- Where should a small business start?
AI prospecting in a small business that sells to other businesses splits cleanly: an agent can do on its own everything that stays inside the company, and the salesperson keeps whatever commits the relationship. The agent builds the account list from a signal, such as a company hiring a purchasing manager. It enriches the records, spots the clients in the CRM who stopped ordering, drafts first messages in your voice, prepares the follow-ups and logs every exchange. The salesperson decides who to approach, approves each message before it leaves and runs the call or the meeting.
That split follows where a salesperson’s time actually goes. In the sixth edition of its State of Sales report, published in July 2024 and based on 5,500 sales professionals in 27 countries, Salesforce says its respondents report spending 70% of their time on tasks that are not selling. The figure is self-reported and comes from a CRM vendor that also sells AI. It still points at the share of the work an agent can take over, the part before and after the conversation.
The legal framework is the part most guides leave out. In France, emailing a professional without their consent is allowed under conditions, whereas calling a consumer to sell them something has required their consent since 11 August 2026. A list enriched by a third-party tool also obliges you to tell each prospect where their contact details came from. What follows is general information on French law as it stands on 14 November 2026, not legal advice.
What can an AI prospecting agent do, task by task?
An AI prospecting agent can handle six tasks end to end, provided it has access to the CRM, a company database and the mailbox, and a person releases whatever goes out. The table sums up the split we recommend for a business of ten to fifty people.
| Task | What the agent does alone | What the salesperson keeps |
|---|---|---|
| Account list | Find the companies showing a signal, and say why each one is there | The choice of accounts to approach |
| Enrichment | Find the right contact, check they are still in post, keep the source | Dropping a contact whose source is unclear |
| Win-back | Spot the clients who stopped ordering and the likely reason | The decision to call, write or wait |
| First message | Draft one message per prospect, in the company’s voice | Reviewing and sending |
| Follow-ups | Keep the calendar and prepare each follow-up on the right channel | Approving each follow-up |
| Logging in the CRM | Record every message, reply and objection | The meeting and what comes after it |
Building the account list from a signal
A signal is a dated fact that makes a company more likely to buy now than three months ago. A manufacturer hiring a purchasing manager is about to review its suppliers, and a client whose buyer has changed is a relationship starting again from scratch. A salesperson notices these facts when they have time to read job ads and the local press, which is to say rarely. An agent looks for them every week in a company database or in published job ads, then matches them against the profile of your best clients.
Its weak spot is the date, because a job ad stays online for weeks after the role is filled. Every line on the list should therefore arrive with the sentence that says why it is there and how old the fact is. The salesperson can then challenge it in two seconds, where a score of 87 out of 100 says nothing about what was measured, and the final call stays theirs, since they know that one company turned down a quote last year for a reason written down nowhere.
Enriching records without inventing contacts
Enrichment adds to a company record the name of the right person, their role and their work contact details. Specialist tools return an address or a number from their own databases. The agent also checks that the person still holds the role, matches the record against what the CRM already knows and flags duplicates, which matters because it reasons on stale data with the same confidence as on good data, as we showed when looking at the data to check before plugging in an agent.
Every contact detail should also keep a trace of the tool or page it came from, since the GDPR requires you to give that source to the prospect, at the latest in your first message. An address inferred from a pattern, such as firstname.lastname@company.fr, remains a guess, and the agent should present it as one.
Winning back clients who stopped ordering
The most profitable seam is often in the CRM, among clients who stopped ordering without anyone noticing. An agent compares orders year on year and isolates the accounts whose volume has dropped. It looks for the likely reason in what you have given it access to, a buyer who left or a competitor’s quote mentioned in an email, and presents it as a hypothesis to check.
This is done account by account, never as a campaign. A generic follow-up sent to three hundred former clients turns dormant clients into lost ones, as we explained when writing about the revenue that comes from clients you do not call. The agent prepares the list, the reason and a draft for each case, which the salesperson rewrites in their own hand before choosing between a call and a message.
Writing the first message in the company’s voice
A personalised first message tells the prospect why you are writing to them, now, and what you can do for their business, in under a hundred words. The signal supplies the “now” and the record supplies the “them”. The agent writes it all in the voice of your best past messages, which it has learned to recognise by reading them.
Left to itself, an AI flatters, and an invented claim about the prospect’s company ruins both the message and the credibility of the person who signs it. That is the first reason for review, and it takes ten seconds if the agent shows where each fact came from. Here is the kind of draft to expect, from an industrial supplies distributor writing to the head of a small company that has advertised for a purchasing manager.
Subject: your new purchasing manager
Dear Ms Royer, I saw that Menuiseries Royer is hiring a purchasing manager in Annecy, and a new buyer often reviews suppliers in their first months. We already supply three joinery firms in Haute-Savoie with hardware and consumables, delivered within 48 hours, and I can send you our terms or visit next month if the timing is right.
Kind regards, [first name, surname, role, company]
We obtained your work address from [source], and what we do with your data is explained here: [link].
To stop receiving our messages, reply “stop”.
Each of the last two lines meets an obligation. The first gives the source of the data, which GDPR article 14 requires you to disclose, and links to the other information that article lists, from the company’s identity to how long the data is kept. The second carries the right to object, which article 21 says must be presented clearly and separately from any other information from the first message onwards.
Following up across channels without borrowing an account
Follow-up is where an agent is most useful, because it mostly takes consistency. It keeps each prospect’s calendar and prepares the follow-up at the agreed time, taking into account what has happened since, such as a reply to a colleague or a trade fair where you met. A follow-up that brings something new gets read, while “did you get a chance to read my message?” goes straight to the bin.
The channel changes the rules, and email follows the ones we set out below. On LinkedIn, the user agreement forbids bots that send messages or add contacts (section 8.2, version of 3 November 2025). The account at risk would be the salesperson’s, along with a network that cannot be rebuilt, so the agent drafts the message there and the salesperson sends it, for the reasons set out in what automating LinkedIn really costs. On WhatsApp, the WhatsApp Business messaging policy, last updated on 23 September 2026, only allows you to contact someone who has given you their number and agreed to receive your messages. It is a channel for following up, never for a first contact.
On every channel, a “no” or a “stop” ends everything, and the agent records it in the CRM so the next campaign does not forget it. The GDPR makes objecting to prospecting an absolute right: the person can exercise it at any time, and their data is then no longer processed for that purpose (article 21).
Keeping the CRM up to date
The CRM is where an agent’s prospecting becomes useful to the whole team. The agent records every message sent, every reply, every objection and the next step. That avoids the most common failing of separate sending tools, an objection known to the campaign tool but missing from the CRM, which lets the next campaign go out to the same prospect.
Retention gives it a background job that always gets put off. Under the CNIL’s reference framework for the management of commercial activities (deliberation no. 2021-131 of 23 September 2021), which is not binding but serves as a benchmark, a prospect’s data may be kept for three years from collection or from the last contact initiated by the prospect. Merely opening an email does not count as contact. Once that period is over, you ask whether they still want your messages, and without a positive answer their data is deleted or archived, a deadline an agent keeps track of better than a salesperson.
What does the salesperson keep in AI prospecting?
The salesperson keeps the acts that commit the company to someone outside it, starting with the choice of accounts. Deciding who to approach depends on what you can deliver this quarter and on the accounts a competitor holds too tightly to be worth the effort. The agent proposes a reasoned list, and the salesperson cuts it.
Approving each message before it goes out is the rule we apply to everything that leaves the company, and this is where it is most often questioned. Granted, reviewing forty drafts a day is tedious, and a salesperson who approves in bulk without reading is no longer checking anything. The answer lies in the shape of the draft, which arrives with its signal and the source of each fact, and in an approval step built into the tool, as explained in what an AI agent should not be able to do. The two kinds of error are not equal: a bad draft costs ten seconds of review, while a message sent with a false fact, or to a prospect who had objected, costs you the prospect.
The call and the meeting, finally, are the sale itself. That is where the prospect says what they will never put in writing and the salesperson decides what they can concede. The agent prepares the call, with the account’s context and the questions to ask, then writes up what was said.
How is an agent different from a prospecting automation tool?
A prospecting automation tool sends a sequence written in advance to a list, whereas an agent reads each prospect’s context and decides what to propose. The sequencer runs three emails four days apart perfectly well, but it does not know that the prospect replied to a colleague last week and sends the scheduled follow-up anyway. In a small business the two work side by side: the sequence remains the sending tool, and the agent decides who enters it, adapts each message and takes out the prospect whose situation has changed. The rule that tells them apart is the one we set out in what separates an agent from an automation: if you can write the script in advance, automate, and if you cannot, delegate.
Does an AI SDR replace a salesperson?
An AI SDR takes over the front end of the SDR’s job, the research, the first message and the follow-up, and leaves qualification and selling to the salesperson. The sales development representative is the role American companies created to separate prospecting from selling: the SDR books the meeting, then hands over to an account executive who closes. In France the title is usually business developer, and most small businesses have none, since prospecting is done by the owner or by a salesperson between two quotes. An AI sales agent, as some vendors call it, mostly takes on preparation that used to be skipped for lack of time.
What the label implies beyond that, an agent holding hundreds of conversations on its own and filling the diary, runs into qualification. Knowing that a prospect has a budget, a deadline and a decision-maker means asking questions and hearing what they leave unsaid. The agent prepares those questions and files the answers, and the salesperson is the one who asks them.
What rules govern B2B email prospecting in France?
In France, emailing a professional without prior consent is allowed when the message concerns their job, they have been told their address might be used and they can object at no cost. The starting rule is strict, though. Article L34-5 of the French postal and electronic communications code prohibits email prospecting that uses the contact details of a natural person who has not consented beforehand, and it makes no distinction between a private individual and an employee reached at their work address. The CNIL, France’s data protection authority, draws the line: prospecting aimed at professionals can rest on the company’s legitimate interest when the subject of the message relates to the recipient’s profession.
Its example is a message presenting the merits of a piece of software to a company’s IT director, and the same director approached about a personal holiday would fall outside that frame. When the addresses were acquired from third parties or are already held by the company, the CNIL asks you to make sure the person was told they might be used for prospecting and is able to object.
Every message must identify the company sending it and offer a way to refuse further messages at no cost beyond that of sending the request. Article L34-5 also forbids hiding the sender’s identity and using a subject line unrelated to the service offered. Two cases fall outside these rules or soften them. Generic addresses such as contact@ or orders@ belong to a legal entity and are not subject to these principles, according to the CNIL. Your existing clients, for their part, can receive messages about products or services similar to those they bought from you, if their details were collected at the time and they could object both then and in every message.
Breaches fall to the CNIL and to the DGCCRF, the consumer protection authority, which can impose an administrative fine of up to €75,000 for an individual and €375,000 for a company. On 25 June 2026, the Conseil constitutionnel ruled that allowing several authorities to punish the same facts with penalties of the same kind breaches the Constitution (decision no. 2026-1210 QPC, brought by Orange). The repeal of the paragraphs concerned is postponed to 31 October 2027, and until then a company already prosecuted or penalised by one of these authorities cannot be pursued by another for the same facts. The substantive rule is untouched.
What does the GDPR change when data comes from a list or an enrichment tool?
Contact details bought or enriched through a third-party tool oblige you to inform each prospect at the latest in your first message, telling them in particular where their data came from. Legitimate interest remains the usual legal basis for B2B prospecting: GDPR article 6 allows it unless the person’s interests or rights override it, and recital 47 states that processing for direct marketing purposes may be regarded as carried out for a legitimate interest. The balancing test assumes the prospecting stays within what a professional can reasonably expect, which is exactly what the CNIL refused to grant KASPR.
GDPR article 14 covers data not obtained from the person themselves. It requires you to give them a set of information within a reasonable period of no more than a month or, if the data is used to communicate with them, at the latest at the first communication. For a prospect found by an agent, that means the first email carries this information: who you are, why you are contacting them, the legitimate interest you pursue, how long you keep the data, their rights and the source of the data, including whether it is publicly accessible.
The right to object goes in that same message, presented clearly and separately from any other information as article 21 requires. For an agent, the consequence is a design one: every contact detail it adds to the CRM keeps its source, otherwise it can neither be given to the prospect nor used to answer an access request. KASPR was criticised on exactly this point, for answering people who asked where their data came from by citing only publicly available sources.
Can you prospect with data taken from LinkedIn?
Reading a LinkedIn profile to prepare an approach is fine, whereas scraping profiles in bulk or collecting the contact details of members who have hidden them exposes you both to the platform’s terms and to the GDPR. LinkedIn’s user agreement, effective since 3 November 2025, forbids using software, scripts or robots to scrape or copy profiles and other data from the service (section 8.2). The penalty is contractual, a restricted or closed account, and it falls on the personal account of the person who used the tool.
The GDPR adds a public penalty, and the CNIL has given an example that bears directly on enrichment tools. On 5 December 2024, it fined KASPR €240,000 (deliberation SAN-2024-020). The company sold a paid Chrome extension that displayed the work contact details of the LinkedIn members whose profile the user was viewing, drawing on a database of around 160 million contacts. The CNIL held that collecting the details of members who had restricted their visibility had no legal basis, because that choice “necessarily binds third parties” and the collection ran against their reasonable expectations.
The decision also found a five-year retention period restarted at every profile update, often a change of job, and information to the people concerned that only began in 2022, by an email written in English. The CNIL closed its order on 4 March 2026, once the database had been deleted and collection from LinkedIn had stopped. The penalty targeted the tool’s publisher, but any company that imports enriched contact details into its CRM becomes responsible for them in turn. Before connecting an agent to an enrichment tool, ask its vendor two questions: where does your data come from, and were the people concerned informed?
Can you prospect by phone in France in 2026?
Since 11 August 2026, cold-calling a consumer requires their prior consent, while calling a professional about their line of work remains possible without it. A law of 30 June 2025 rewrote article L223-1 of the French consumer code, which now prohibits canvassing by phone any consumer who has not consented beforehand, and the Bloctel opt-out register has gone. According to the DGCCRF, that consent lasts a year at most and its proof must be kept for at least three years. Fines can reach €75,000 per call for an individual and €375,000 for a company.
These rules protect the consumer, whom the consumer code defines as a natural person acting for purposes outside their commercial, industrial, craft, liberal or agricultural activity. A craftsman called about his workshop is therefore not one. For professionals, the CNIL keeps an opt-out regime: the call can rest on legitimate interest if it concerns the person’s profession, provided they were told their number might be used and can object. Every call must also let them know who is calling and refuse further calls.
An agent placing these calls itself would face further rules, from announcing that the caller is an AI to the still open question of automated calling systems, which our guide to AI phone answering covers in detail. In the split we recommend, the call stays with the salesperson, and the agent prepares the list, the background on each account and the questions to ask.
What does Balt do in prospecting?
Balt, the AI employee for small businesses, takes on this preparation in the tools you have opened to him: HubSpot or Salesforce for the CRM, Apollo and LinkedIn for research, Brevo and Lemlist for campaigns, Gmail for email. He finds companies by signal, such as the local manufacturing firms hiring a purchasing manager, and hands you a list with one contact per company, leaving you to choose who to approach. He also compares orders year on year in your CRM, spots the clients who stopped ordering and gives you the likely reason behind each silence as a hypothesis to check.
He builds your campaigns from the contacts in your CRM, leaves out the ones who should not get them, writes the message in your voice and sets the campaign up in Brevo or Lemlist. Nothing goes out without your approval. He sends once you have signed off, then follows each prospect up on the right channel, by email, on LinkedIn or on WhatsApp; on LinkedIn he drafts the message and you send it, because automating sends breaches the platform’s terms and the account at risk would be yours.
He only changes a CRM record after asking you, and he remembers what he learns working with you, from the sectors that reply to the voice of your best messages. Every Monday, as a routine, he can send you the clients to follow up, the deals sitting idle in the CRM and the accounts worth approaching first.
Where should a small business start?
Start with win-back, because it works on your own data, can be checked at a glance and only reaches people who already know you. Ask the agent for the list of last year’s clients who have ordered nothing this year, with the likely reason behind each silence, and compare it with what you know. You will soon see how well it reads your CRM, on accounts you know by heart.
Then move on to searching by signal, with one specific signal and a single region, and after that to first-message drafts, approved one by one for the first few weeks. It is the order we describe for any delegation in where to start with an AI agent: first what can be checked quickly, then what calls for judgement. Budget comes up at that point, and it is best judged on the work done rather than on a number of seats, as what an AI agent costs explains.
It is also the order we recommend with Balt, and our page on automated prospecting shows what he takes on week after week. Once the meeting is booked, the next question is the quote that follows it, from writing it to following it up, which our guide to writing a quote that gets signed answers.
Frequently asked questions
Can an AI agent send prospecting emails on its own?
Technically yes, and French law allows you to email a professional without prior consent when the message concerns their job, they have been informed and they can object. We still recommend that a person approves every message before it goes out, because a false fact or the wrong recipient costs you the prospect and cannot be undone. A draft that arrives with its signal and the source of each fact takes seconds to check.
Do you need a business prospect’s consent to email them?
Not when the message relates to their profession. The CNIL then accepts the company’s legitimate interest as the legal basis, provided the person was told their address might be used for prospecting and can object at no cost. Every message says who sent it and how to stop receiving more, and a generic address such as contact@ is not covered by these rules.
Can you use LinkedIn data for prospecting?
LinkedIn’s user agreement forbids scraping profiles with software and using bots to send messages or add contacts. On the GDPR side, the CNIL fined KASPR €240,000 on 5 December 2024 for collecting the contact details of members who had restricted their visibility. Reading a profile to prepare an approach that the salesperson sends from their own account is still fine.
Does an AI SDR replace a salesperson?
It takes over the preparation: the account list, the enrichment, the message drafts and the follow-up calendar. Qualifying on the phone, the meeting and the negotiation stay with the salesperson, because the conversation is where the prospect decides. In a small business with no SDR, the agent mostly takes on preparation that used to be skipped for lack of time.
Can you cold-call businesses in France in 2026?
Yes. The prior consent rule in force since 11 August 2026 protects consumers, and calling a professional about their line of work can rest on legitimate interest, according to the CNIL. They must have been told their number might be used and be able to object, and every call has to say who is calling and let them refuse further calls.
Sources
- Légifrance, French postal and electronic communications code, article L34-5 (version in force since 26 July 2020)legifrance.gouv.fr
- Conseil constitutionnel, decision no. 2026-1210 QPC of 25 June 2026, Société Orange (overlapping penalties under article L34-5)conseil-constitutionnel.fr
- CNIL, Commercial prospecting by email, SMS-MMS and automated calling (page dated 10 June 2026, in French)cnil.fr
- CNIL, Commercial prospecting by phone (excluding automated calling): what are the rules? (page dated 10 June 2026, in French)cnil.fr
- CNIL, General Data Protection Regulation, recital 47 (prospecting and legitimate interest)cnil.fr
- CNIL, GDPR, chapter II, article 6 (lawfulness of processing)cnil.fr
- CNIL, GDPR, chapter III, articles 14 (data not obtained from the data subject) and 21 (right to object)cnil.fr
- Légifrance, CNIL restricted committee, deliberation no. SAN-2024-020 of 5 December 2024 concerning KASPRlegifrance.gouv.fr
- CNIL, Data scraping: KASPR fined €240,000 (19 December 2024)cnil.fr
- CNIL, Closure of the order issued against KASPR (6 March 2026)cnil.fr
- Légifrance, CNIL, deliberation no. 2021-131 of 23 September 2021 adopting a reference framework for the management of commercial activitieslegifrance.gouv.fr
- Légifrance, French consumer code, article L223-1 (in force since 11 August 2026)legifrance.gouv.fr
- Légifrance, French consumer code, preliminary article (definition of the consumer)legifrance.gouv.fr
- DGCCRF, The phone canvassing rules businesses must follow (27 July 2026, in French)economie.gouv.fr
- LinkedIn, User Agreement, section 8.2, effective 3 November 2025linkedin.com
- WhatsApp Business, Messaging Policy (last updated 23 September 2026)whatsappbusiness.com
- Salesforce, State of Sales, sixth edition: 5,500 sales professionals in 27 countries (25 July 2024)salesforce.com

