AI does not win customers: it speeds up four of the eight parts of a prospecting system. This guide builds all eight, with the public sources you can actually use in Mexico and Paraguay, the email infrastructure that keeps you from burning your domain, sequences that do not read like a template, and a way to tell in thirty seconds where the funnel broke.
Short answer: A B2B lead generation system with AI links eight parts: ideal customer profile, legitimate data sources, enrichment, email verification, multichannel sequences, qualification, meeting booking and measurement. AI speeds up research, writing, classification and diagnosis, but it does not replace the sources or the verification, and those are what decide whether it works. Buying a list is not a system.
- What is a B2B lead generation system with AI, and how is it different from buying a database?
- How do you define your ideal customer profile without making it up?
- Where does legitimate data come from in Mexico and Paraguay?
- What can you do on LinkedIn, and what can't you?
- Why are trade-show exhibitor lists the best source?
- How do you enrich a list without inventing data?
- How do you verify email addresses, and why does everything depend on it?
- How do you write sequences with AI without landing in spam?
- Which channel should you use, with what consent, and how do you get WhatsApp opt-in?
- How do you qualify and score a lead without inventing the scoring?
- How do you get from a reply to a booked meeting?
- Which metrics measure the funnel, and how do you know where it breaks?
- What does compliance require in Mexico and Paraguay?
- Tricks that are not in the manual
- What should you do this week?
What is a B2B lead generation system with AI, and how is it different from buying a database?
A B2B lead generation system with AI links eight parts: ideal customer profile, legitimate data sources, enrichment, contact verification, multichannel sequences, qualification, meeting booking and measurement. AI speeds up four of them —research, writing, classification and diagnosis— but it replaces none. Buying a database is not a system: it is skipping the first seven parts and hoping the eighth forgives you.
Almost every project that fails in Mexico and Paraguay fails the same way: someone bought a list, asked for a hundred personalized emails and waited for meetings. What arrived instead was bounces, spam complaints and a burned domain. AI does not fix a bad list: it amplifies it, and fast.
Of the eight parts, two are beyond any AI: where the data comes from and whether the email address exists. They decide whether the system works, which is why they come first here. The other six do speed up: AI finds patterns in your closed deals, structures public information, writes without clichés, classifies replies against a schema you define, prepares the brief before each meeting and diagnoses the funnel. None of this replaces the trade-show culture and personal relationships of both markets: it organizes them.
How do you define your ideal customer profile without making it up?
By looking backward. The usual mistake is asking "who do we want to sell to?": the answer is always too broad. The right question is "who have we already sold to successfully?", and you answer it with data you already have. Take your last thirty closed deals —won and lost— from the past eighteen months and record eight fields: industry, headcount, city, job title of the person who signed, job title of the person who pushed back, what was happening at the company when they contacted you, cycle length and deal size. Thirty rows show the pattern; ten do not. Separate won from lost and look for the difference: a factor nobody had named usually shows up.
Ideal customer profile template (copy it and fill it in)
1. Firmographics. Industry or activity code · headcount range · estimated revenue · states or departments.
2. Operations. What they use today for the problem you solve · minimum volume that justifies your solution · observable signal of that volume.
3. People. Who signs · who uses · who blocks · who introduces you.
4. Triggers. Three public facts that signal good timing: a new plant, an appearance on an exhibitor list, a specific job opening, a public contract award, a change of CEO.
5. Anti-profile. Three characteristics that disqualify automatically. Being explicit here is worth more than everything above.
6. One-line statement. "We sell to ___ with ___ employees in ___ who ___, usually when ___."
Prompt to extract the profile from your closed deals
You are a B2B sales analyst. I am giving you 30 closed deals. Three steps, showing each one:
STEP 1. Count: the distribution of WON and of LOST separately, in tables, field by field. No interpretation.
STEP 2. Compare: the three fields where the difference between the two groups is largest, with the exact figure.
STEP 3. Propose an ideal profile and an anti-profile, each statement with its support: "(supported by 11 of 18 won deals)". If there is none, write INSUFFICIENT DATA. Do not invent categories or add recommendations at the end.
<deals>[paste your table]</deals>
That mandatory INSUFFICIENT DATA is what separates useful analysis from text that merely sounds good. It is the principle behind the official prompting best practices (consulted on 19 September 2026): documents at the top, question at the bottom, quotes before conclusions. More assignments like this one in the guide to prompts for sales teams.
Where does legitimate data come from in Mexico and Paraguay?
From four types of source, and none of them is a file someone sent you over WhatsApp: official statistical directories, public registries, chamber of commerce directories and trade-show exhibitor lists. All four can be cited and all four survive an audit.
In Mexico, INEGI's Directorio Estadístico Nacional de Unidades Económicas (DENUE, the national statistical directory of business establishments) released its 05/2026 edition with 6,138,075 active business establishments, according to INEGI's bulletin of 20 May 2026: location, SCIAN activity code and size by headcount bracket. It does not give you a person's name or email address, and that is fine, because that is exactly what keeps it from being a list of personal data. Use it to build your universe of accounts; the contact comes later.
| Mexico | What you get |
|---|---|
| DENUE (INEGI) | Legal name, SCIAN activity, headcount bracket and address: the universe of accounts by industry and size |
| SIEM | Companies registered through chambers, with line of business and products; cross-referenceable with DENUE |
| CANACINTRA · COPARMEX · INDEX | Members by sector, state and IMMEX status: an introduction through the industry association, not a cold list |
| Secretaría de Economía · Bancomext · Nafin | Sector programs, exporters and supply chains: a signal of capacity to invest |
One housekeeping note: ProMéxico no longer exists —it was wound down in 2019 and its functions moved to the Secretaría de Economía and the foreign ministry.
| Paraguay | What you get |
|---|---|
| DNIT — RUC lookup | Legal name and taxpayer status: confirms the company exists and is active |
| DNCP — public procurement | Government suppliers, awards and categories: the best buying trigger there is |
| datos.gov.py | Open data from public agencies, for cross-checks and verification |
| UIP · CCPB · ARP | Industry, trade with Brazil and agribusiness: segmentation plus an industry-association introduction |
| REDIEX · MIC | Export programs, trade delegations and maquila: companies going international |
In Paraguay the sequence that works does not start with a cold email, it starts with showing up where your buyer already is: one chamber member who introduces you converts better than fifty perfect emails. If you automate processes using data from Paraguayan customers, first review what changed under Ley 7593, Paraguay's personal data protection law.
Want a hand applying this?
Free 30-minute working session. Uniamos is a remote AI automation agency (Austin and Mexico City) working over video calls and WhatsApp.
What can you do on LinkedIn, and what can't you?
You can search, read public profiles, connect with a message you wrote yourself and use the platform's own tools. You cannot scrape data automatically or use extensions that simulate your activity: the user agreement prohibits it and LinkedIn maintains a page listing prohibited software and extensions. The risk is not a fine: it is losing the account of your best salesperson, along with their network and their history.
The line is simple: what a human sees, a human may write down; a program that opens a thousand profiles overnight may not. Sales Navigator is the sanctioned route for advanced search, and AI comes in after the data.
Prompt to prepare a first message from a profile you are looking at
Context: I sell [product] to [role]. I am pasting the public text of a profile and of their company. Give me back: (1) three concrete, verifiable facts from the text, no inferences, and if there are fewer than three, say so; (2) the most likely problem hypothesis, in one sentence, stating which fact it rests on; (3) a connection message of at most 300 characters that mentions ONE concrete fact, does not compliment the person and ends with a closed yes-or-no question.
Forbidden: "I loved your profile", "hope this finds you well", "revolutionary", "synergy", "digital transformation", exclamation marks.
<profile>[paste]</profile> <company>[paste]</company>
Why are trade-show exhibitor lists the best source?
Because they are the only public source where a company has already proven three things at once: it has budget, that sector is its priority, and it wants to talk to strangers. And the list is free, it sits on the organizer's website and it updates itself.
The Annual Exhibition Industry Study from AMPROFEC, presented on 5 June 2024 with 2023 data, puts the number of leads obtained by exhibitors in Mexico at 15.22 million a year, with 3.04 million deals closed and 170 contacts per exhibitor per event. That last figure justifies everything: 170 contacts per show cannot be managed with business cards and a good memory. And the CEIR report (2019, with 2015 and 2018 data) adds that only 39% of trade-show leads turn out to be qualified by the exhibitor's own criteria and that fewer than 10% use custom qualifiers. The problem is not capturing more: it is qualifying as you capture.
- Find the exhibitor directory for the current edition and the previous one: two editions in a row signals commitment to that channel.
- Copy company, booth number, category and country. The booth matters: size and location tell you about the budget.
- Run it against your anti-profile and cut without mercy. Of 500 exhibitors, 60 real ones will remain.
- Enrich those 60 with the official sources above and write ten days before the event, not after.
To get started today, with dates confirmed on the official websites on 19 September 2026: EXPO CIHAC (14–16 Oct 2026, Centro Citibanamex), Expo Paraguay Brasil (11–13 Nov 2026, Hernandarias), Mexico Business Summit (24–25 Nov 2026, Expo Santa Fe), EXPO MANUFACTURA (23–25 Feb 2027, Cintermex), INNOVAR (16–19 Mar 2027, CETAPAR) and Constructecnia (20–23 May 2027, CONMEBOL). Verify before you book travel.
The full calendar is in trade shows and lead capture with AI; what happens afterwards is in post-show follow-up in 72 hours.
How do you enrich a list without inventing data?
With two separate layers that never get mixed. The deterministic layer is verifiable truth: legal name, tax ID, activity and address come from DENUE, from SIEM or from the DNIT's RUC lookup. The probabilistic layer comes from B2B providers and is an estimate. The rule that prevents disaster: every field carries its source, and if a value is not in the source, the field stays empty. Empty is not a problem; invented is.
| Tool | What it is for and what to expect |
|---|---|
| Apollo.io | B2B database, enrichment and sequences; free plan, uneven coverage in Paraguay |
| Lusha | Contacts and verification: a second opinion on a doubtful contact |
| Clearbit | Company enrichment from the domain; now part of HubSpot |
| Dropcontact | Enrichment with a compliance focus, useful with European contacts |
| Clay | Orchestrates several sources in a waterfall; a real learning curve |
None of them publishes full open pricing and Latin American coverage varies by industry: test with fifty of your own accounts before you sign anything.
Enrichment prompt with an anti-invention rule
You are a B2B data analyst. I am giving you the public text of a company's website. Return only these fields:
legal_name · primary_industry · products_or_services (max. 5) · geographic_markets · size_signals · technology_mentioned · recent_trigger (with date) · icp_fit (high/medium/low/not determinable) · fit_rationale (one sentence quoting the text verbatim).
ABSOLUTE RULE: if a field does not appear verbatim in the text I gave you, write NOT_AVAILABLE. Do not infer the industry from the name and do not estimate headcount.
<website>[paste]</website> <icp>[paste your ideal profile]</icp>
Run that prompt on twenty accounts you know well and count the errors: if a single invented value appears, tighten the instruction before you scale it to a thousand. Save your profile and anti-profile as a reusable base; how to build one is in Projects as your company's brain.
How do you verify email addresses, and why does everything depend on it?
Because email providers have stopped tolerating careless senders. If your domain does not meet the requirements or your complaint rate goes over the line, your messages stop arriving —including the ones your team sends to real customers.
According to Gmail's official sender guidelines (consulted on 19 September 2026):
- The bulk sender threshold is 5,000 messages a day to Gmail accounts, and subdomains count together.
- Every sender needs SPF or DKIM; bulk senders need SPF, DKIM and DMARC, with alignment on the "From" domain.
- The spam rate must stay below 0.30% in Postmaster Tools; Google recommends operating below 0.10%, and above 0.3% you cannot even request mitigations.
- One-click unsubscribe is mandatory, along with a visible link in the body and processing opt-outs within 48 hours.
The minimum infrastructure comes down to five decisions. A separate domain for prospecting, so that billing and support keep getting through if something goes wrong. A three- to four-week warm-up starting at ten or fifteen emails a day per mailbox. A ceiling of thirty to fifty cold sends per day per mailbox. Verification before every send with Hunter, ZeroBounce or Dropcontact, setting "catch-all" addresses aside. And watch bounces, not opens: above 3% hard bounces you have to stop and clean, while open rates have been inflated for years by systems that load images on their own.
Sender health checklist (every Monday)
☐ SPF, DKIM and DMARC configured and aligned. ☐ Spam rate below 0.10%. ☐ Hard bounces below 3% over 7 days. ☐ Unsubscribe link visible and working. ☐ Opt-outs processed in under 48 hours. ☐ No mailbox above its daily ceiling. ☐ Suppression list up to date, permanent and applied across every channel.
How do you write sequences with AI without landing in spam?
By putting one verifiable fact and a single objective in each message, and banning brochure vocabulary. The difference between an email that gets a reply and one that gets deleted is almost never the offer: it is whether the reader senses that somebody looked at their company before writing.
- One fact per message: the trade-show booth, the contract award, the new plant, the job posting.
- One objective per message. The first one is after a reply, not a meeting: asking a stranger for 45 minutes in a first email is proposing marriage on a first date.
- Between 60 and 90 words, with no attachments or links in the first send.
- One closed question at the end. "Would it help if I sent you two lines on how a company your size solved this?" gets answered. "Looking forward to your thoughts" does not.
| Touch | Day | Channel | Objective |
|---|---|---|---|
| 1 | 0 | Concrete fact + problem hypothesis + closed question | |
| 2 | 2 | Connection with no pitch, mentioning the same fact | |
| 3 | 4 | How a similar company solved it, without naming them | |
| 4 | 8 | Phone call | Thirty seconds: who you are, why you are calling, a question |
| 5 | 14 | "I'll close this out unless you tell me otherwise" |
Prompt to generate the full sequence
You are a B2B prospecting copywriter writing in US English, and you write the way one busy person writes to another.
CONTEXT. We sell: [one sentence]. Proof: [one sentence, no customer names]. Target profile: [title, industry, size]. Concrete fact about THIS company: [paste]. Country: [Mexico / Paraguay].
ASSIGNMENT. The 5 touches in the table. Maximum 90 words per email; one fact per message; a single closed question at the end; no attachments or links in touch 1; a 6-word subject line, lowercase, no exclamation marks.
FORBIDDEN: "hope this email finds you well", "revolutionary", "market leader", "end-to-end solution", "synergy", "digital transformation", "disruptive", emojis, all caps, promises expressed as percentages.
DELIVERABLE. The 5 messages; then the 3 weakest sentences and why; then rewrite only those 3.
That last instruction —have the model critique its own text and rewrite it— is the self-correction pattern from the official documentation. One language note: if you write in Spanish, Paraguayan Spanish uses voseo in everyday address, so asking the model to "adapt this text to how people normally address each other in Paraguay, keeping a professional register" costs a second and shows.
Which channel should you use, with what consent, and how do you get WhatsApp opt-in?
The order depends on a single variable: the permission you have. Cold business email is the channel with the least legal friction in B2B, LinkedIn requires you to respect the platform's rules, and WhatsApp requires prior consent. Starting with WhatsApp because "everyone here uses it" is the most expensive mistake in both markets.
| Channel | What it is good for | What permission you need |
|---|---|---|
| Business email | Cold first contact at scale | Privacy notice published and objection rights honored |
| Introduction and context; works without an email address | Comply with the user agreement; nothing automated | |
| Following up with someone who already knows you | Explicit prior opt-in, always | |
| Phone call | Unsticking and qualifying in two minutes | Nothing special in B2B, with clear identification |
| Trade shows and events | Capturing with consent on the spot | Consent checkbox on the booth form |
WhatsApp's policy requires that the recipient has your number and that you have received their opt-in confirming they want to receive your messages; it also prohibits confusing, deceiving, spamming or surprising people. It is set out on the official policy page and in Meta's technical documentation. The four places where you get solid opt-in: the booth form, your website form with a separate, unchecked box, a reply to an email in which the contact asks you to write to them there, and a printed QR code that opens a chat the user initiates.
Copy-and-paste consent text (adapt it to your legal entity)
☐ I authorize [Legal entity] to contact me on WhatsApp at the number I am providing, to follow up on this request and to send me related commercial information. I understand that I can opt out at any time by replying STOP to that same number. I have read the privacy notice available at [URL].
Store alongside the checkbox: date and time · channel where it was obtained · the exact version of the text accepted · name of the event or form · who captured it. Without those five fields you cannot prove anything six months later.
Messages the company initiates require approved templates; once the contact replies, a service window opens in which the conversation is free-form, so that first template has to be written to provoke a reply. To operate through the API you need an official provider: 360dialog, Twilio or the WhatsApp Business Platform itself. The full architecture is in the definitive guide to AI agents on WhatsApp Business.
How do you qualify and score a lead without inventing the scoring?
With two axes that are never blindly added together: fit (how closely they resemble your ideal customer) and signal (what they have done that indicates interest). High fit with low signal is worked patiently; high signal with low fit gets a polite response and no assigned rep. Confusing the two is the origin of the classic "marketing keeps sending me junk".
| Axis | Criterion | Points |
|---|---|---|
| Fit | Industry within the ideal profile | 15 |
| Fit | Size within the target range | 15 |
| Fit | Located in a territory you serve | 10 |
| Fit | Contact's role: decides or uses | 10 |
| Signal | Replied with substance, not just an acknowledgment | 20 |
| Signal | Mentioned a specific problem or a deadline | 15 |
| Signal | Recent public trigger | 10 |
| Signal | Asked for information on their own initiative | 5 |
Thresholds: 70 or more, call within 24 hours; between 40 and 69, educational sequence and a review in thirty days; below 40, keep in the database. And one condition overrides everything: if the lead falls into your anti-profile, it is disqualified regardless of score.
Prompt to classify free-text replies
You are a classifier. I am giving you a contact's reply verbatim. Return exactly these fields and nothing else:
intent: [interested_now | interested_later | referral | not_interested | opt_out_requested | out_of_office] · problem_mentioned · timeline_mentioned · who_decides: [verbatim quote or NOT_AVAILABLE] · objection: [price | timing | already_have_a_vendor | not_a_priority | none] · next_action: [call_within_24h | send_proposal | reschedule | suppress_from_all_lists] · confidence: [high | medium | low]
RULES: do not invent content that is not in the reply. If the contact asks not to be contacted, the intent is opt_out_requested and the action is suppress_from_all_lists, no exceptions: that is the point where automation can land you in legal trouble. When in doubt, pick the most conservative category and mark confidence: low.
<reply>[paste]</reply>
Want a hand applying this?
Free 30-minute working session. Uniamos is a remote AI automation agency (Austin and Mexico City) working over video calls and WhatsApp.
How do you get from a reply to a booked meeting?
By replying fast, offering few options and preparing the conversation before you have it. Speed matters most, and it takes no technology: it takes someone being assigned the task.
- Two specific slots, not an open calendar. "Does Tuesday at 11:00 or Thursday at 16:00 work for you?" closes better than a link; the link goes out afterwards, to confirm.
- Twenty minutes. If it deserves more, it will run long on its own.
- The agenda inside the message: what you want to understand, what you will show, what gets decided. It cuts no-shows because the contact knows what they get.
- A reminder the day before, on the channel where they replied to you.
Pre-call brief prompt (five minutes before the call)
Prepare a one-page brief for a 20-minute meeting. I am giving you the full thread with the contact, the account record and my ideal customer profile.
Structure: (1) What we know for certain, with a verbatim quote from the thread. (2) What we are assuming, labeled ASSUMPTION. (3) The three questions to qualify volume, timeline and decision-maker. (4) The most likely objection and a two-sentence response. (5) The outcome that counts as success.
Maximum 300 words, no generic industry background. If the thread does not support a section, write "no information".
That brief turns an ordinary rep into one who looks like they studied the account for half an hour: it is the highest immediate-return use of AI in the entire sales process.
Which metrics measure the funnel, and how do you know where it breaks?
Eight linked rates, each with a different symptom when it fails. The value is not a pretty dashboard: it is answering "why were there no meetings this month?" in thirty seconds, without arguing over hunches.
| Metric | Sign it is healthy | If it fails, the problem is in… |
|---|---|---|
| Delivery (delivered ÷ sent) | Above 97% | A dirty list or misconfigured authentication |
| Hard bounces | Below 3% | Verification: you are sending to addresses that do not exist |
| Spam complaints (Postmaster Tools) | Below 0.10% | Relevance and permission: you are writing to the wrong people |
| Reply rate (÷ delivered) | Stable or rising | Segmentation or copy |
| Positive replies (with intent) | One in four | A poorly defined ideal customer profile |
| Meetings booked | More than half | Response speed and how you propose the meeting |
| Attendance (held ÷ booked) | Above 70% | No agenda sent in advance and no reminder |
| Opportunities (÷ held) | Depends on your cycle | Qualification: you are meeting people who do not buy |
Diagnosis is four questions in order. Do they arrive? If delivery drops below 95%, stop everything: every other metric is contaminated. Do they reply? Change the segment first, it moves the needle more than the copy. Do they reply yes? Lots of replies with no intent means your ideal profile is wrong. Does it turn into an opportunity? If you book meetings and nothing converts, the failure is in qualification. Golden rule: one variable every two weeks, measured in weekly cohorts.
Weekly funnel diagnosis prompt
I am giving you the figures for the last 8 weeks, week by week: sent, delivered, hard bounces, complaints, replies, positive replies, meetings booked, meetings held and opportunities.
Return: (1) A table with the 8 rates by week. (2) The first point in the chain where the rate falls below the threshold I give you. (3) The two most likely causes of THAT drop and how to tell them apart using a data point I can look at tomorrow. (4) A single experiment for the next two weeks, with the metric that judges it and the number that declares it a winner.
Do not recommend "improve personalization" without saying what to change. If the data does not support a conclusion, say so and ask for the missing data point.
<thresholds>[paste]</thresholds> <figures>[paste]</figures>
These rates line up with the scorecard in measuring the return on AI training.
What does compliance require in Mexico and Paraguay?
Three things that cost little and prevent a lot: an accessible privacy notice, a clear way to object, and a record of where each data point came from. Both countries ask for the same thing: being able to show what you do with data and why.
Notice. This section is informational and does not constitute legal advice. Before launching a campaign using personal data, consult a qualified lawyer in the relevant country.
Mexico: the law changed in 2025 and the authority is no longer the one you think
The Ley Federal de Protección de Datos Personales en Posesión de los Particulares (LFPDPPP, the federal law on protection of personal data held by private parties) now in force was published in the DOF —the official federal gazette— on 20 March 2025, with an amendment of 14 November 2025. INAI was dissolved and its functions passed to the Secretaría Anticorrupción y Buen Gobierno (the ministry of anti-corruption and good governance). If your privacy notice still names INAI, it is out of date.
- The privacy notice is mandatory: controller, data processed, purposes, options to limit use and mechanisms for exercising rights (articles 14 to 16).
- As a general rule, tacit consent is valid: once the notice has been made available, if the person does not state otherwise, consent is deemed given (article 7). Financial and asset data require express consent; sensitive data, express consent in writing (article 8).
- ARCO rights (access, rectification, cancellation and objection) must be answered within 20 days and acted on 15 days after that (article 31). Penalties are expressed in UMA units and can double where sensitive data is involved (articles 58 to 60).
Paraguay: Ley 7593/2025 and the clock running to November 2027
Paraguay enacted Ley N.º 7593/2025 de Protección de Datos Personales (the personal data protection law) on 27 November 2025, with a twenty-four-month vacatio legis: fully enforceable from around 27 November 2027. It creates the national personal data protection agency, under MITIC. All processing needs a lawful basis —consent, contract, legal obligation or legitimate interest, which is the one B2B prospecting usually invokes and which requires documenting the balancing test against the data subject's rights—; the data subject may object to processing for marketing purposes; and security breaches must be notified within a maximum of 72 hours, with penalties expressed in minimum daily wages that escalate where sensitive data is involved.
Compliance checklist for prospecting
☐ Privacy notice published and linked in every email. ☐ Every record stores the origin of the data: source, date and how it was obtained. ☐ Opt-out visible on every channel, processed within 48 hours. ☐ Global suppression list: anyone who opts out on one channel comes off all of them. ☐ WhatsApp consent with date, channel and exact text. ☐ A contract with every provider that processes data on your behalf. ☐ CRM access per person, revocable when they leave. ☐ Nobody pastes customer data into unapproved personal tools.
That last point deserves a document of its own, with a copy-and-paste template: internal AI use policy and governance.
Tricks that are not in the manual
Twelve tricks that come from watching the same process break many times.
1. Write to exhibitors ten days before the show, not after
Everyone does post-event follow-up and nobody does pre-event preparation. "I see you're at booth 412; I'll be there Thursday, does it make sense for me to stop by for fifteen minutes?" gets answered in a way it never would on a Tuesday in February.
2. Put the anti-profile before the profile in every prompt
Models are eager to please: if you ask them to assess fit, they tend to find it. Starting with "automatically disqualify if…" cuts the list in half and raises the quality.
3. Require the verbatim quote before the conclusion
Swap "does this company fit?" for "first copy the sentences you are going to rely on, then conclude". The model cannot sustain a conclusion without the sentence that backs it.
4. Store the origin of every data point in a CRM field
A "source" field with values like DENUE, exhibitor-CIHAC-2026 or referral. It proves compliance and reveals which source converts best. You cannot reconstruct it after the fact.
5. Explicitly ban brochure words
A blacklist in the prompt —synergy, disruptive, end-to-end solution, market leader— improves the writing more than any instruction about tone.
6. Always send the fifth touch
"I'll close this out unless you tell me otherwise" is the message that generates the most replies and the one almost nobody sends. It is not giving up: it is giving permission to answer.
7. Use a sending domain that is not your billing domain
If prospecting goes badly, collections, support and payroll keep getting through. The cheapest insurance policy there is.
8. Ask the three qualifying questions every time, in the same order
Volume, timeline, who decides. At the booth, on the call and in the email. According to CEIR, almost nobody does: doing it puts you in the top 10%.
9. Ask the model to critique its own text before delivering it
"Write the sequence. Then point out the three weakest sentences and why. Then rewrite them." It is the self-correction pattern from the official best practices.
10. Check sender health on Mondays, not when it breaks
Finding out three weeks late that you are above 0.3% complaints costs a quarter of rebuilding.
11. Keep the suppression list separate from the CRM
Anyone who asks not to be contacted comes off everything: CRM, sending platform, WhatsApp and the new rep's spreadsheet. A master list checked before every send prevents the late email that ruins a relationship.
12. Change one variable every two weeks and write it down
One line per experiment —what I changed, when, which metric judges it, the result— turns prospecting into cumulative learning.
What should you do this week?
Five days, one task a day, without buying anything new. A small but complete system beats a big half-built one.
- Monday — The profile. Export your last thirty closed deals, run them through the prompt in section two and write your profile and anti-profile on one page. Have your two most experienced reps argue it out.
- Tuesday — The source. Pick one: DENUE in Mexico, the DNIT's RUC lookup or the public procurement portal in Paraguay, or the exhibitor directory for your industry's trade show. Pull a hundred accounts and cut them against the anti-profile.
- Wednesday — The infrastructure. Check SPF, DKIM and DMARC on your sending domain, open Postmaster Tools and, if the domain is new, start the warm-up today.
- Thursday — The message. Write the five-touch sequence and read it out loud. Any message you would be embarrassed to read out loud, rewrite it.
- Friday — Measurement and permission. Build the spreadsheet with the eight rates, add the "source" field to the CRM and the WhatsApp consent checkbox to your form, and publish the updated privacy notice.
The following week you start sending, in small volume, and measuring. If you would rather build it with support and train the team at the same time, AI training with Claude is that option.
Frequently asked questions
Is cold B2B prospecting legal in Mexico and Paraguay?
Yes, with conditions. In Mexico, the Ley Federal de Protección de Datos Personales en Posesión de los Particulares published in the DOF on 20 March 2025 generally accepts tacit consent where the privacy notice has been made available to the data subject and they have not objected, and it requires ARCO rights to be answered within 20 days. In Paraguay, Ley 7593/2025 accepts legitimate interest as a lawful basis and recognizes the right to object to processing for marketing purposes; it will be enforceable from around November 2027. You need an accessible privacy notice, a working opt-out route and a record of where each data point came from. Informational only, not legal advice.
Can I buy a company database and start tomorrow?
You can, and it is the fastest route to a burned domain. You do not know where the data came from, there are usually plenty of addresses that no longer exist, and the contacts have never heard of you. Bounces drive complaints, complaints drive filtering, and the filtering also hits the email your team sends to real customers. Building your own list from public sources takes two weeks longer and lasts for years.
How many emails a day can I send without running into trouble?
Gmail's official guidelines set the bulk sender threshold at 5,000 messages a day to Gmail accounts and require SPF, DKIM and DMARC beyond that point, a spam rate below 0.30% in Postmaster Tools and one-click unsubscribe. For cold prospecting the prudent number is much lower: thirty to fifty sends a day per mailbox, on a domain warmed up over three or four weeks.
Can I use a tool that scrapes contacts from LinkedIn automatically?
No, not if you want to keep your team's accounts. LinkedIn maintains an official page listing prohibited software and extensions, and its user agreement prohibits automated scraping. The real risk is not a financial penalty but a restriction on the rep's personal account, along with their network and their history. The sanctioned route is Sales Navigator, plus manual work for everything else.
Can I message someone on WhatsApp after meeting them at a trade show?
Only if they gave you explicit consent for that channel. WhatsApp Business policy requires that the recipient has your number and that you have received their permission confirming they want to receive your messages. The clean way is a consent checkbox on the booth form, storing the date, the channel, the text accepted and the name of the event. Their business card is not consent for WhatsApp.
What is the minimum budget to build this system?
Less than people usually assume: one additional sending domain, an email verifier, a CRM with a free tier and an AI subscription. Claude plans start at 20 dollars per seat per month on Team with annual billing, according to its official pricing page; that is as of September 2026, so confirm on their site, because prices change and vary by region. The expensive part is your team's time.
How long does it take to see results?
The first honest indicator arrives at four or five weeks: three weeks of domain warm-up plus two of low-volume sending. First meetings usually appear between week four and week six. Judging the system before you have worked a hundred contacts all the way to the fifth touch is like judging a trade show by its first morning.
Can AI run the whole prospecting operation on its own?
No, and serious attempts usually end up worse than the manual process they replaced. AI is excellent at researching public information, drafting variants, classifying replies against a schema and diagnosing a funnel. It is bad at the two things that decide the outcome: obtaining data from legitimate sources and verifying that an email address exists.
What do I do with leads that do not qualify?
Keep them with their disqualification reason and review them every six months, because the reason is usually timing, not the company. A lead disqualified on size today may fit next year; one disqualified on the anti-profile never will. Telling them apart with an explicit field keeps you from bothering people who are never going to buy.
Related articles
- Trade shows and expos: how to prepare and capture leads with AI
- Post-show follow-up in 72 hours with AI and WhatsApp
- Claude prompts and tricks for sales teams
- How to roll out Claude in your company, phase by phase
- Internal AI use policy: guide and template
- Measuring AI adoption and return in your company
- WhatsApp Business with AI for sales teams
- More from the blog