Operating case · Sales and governed AI agents

Governed AI sales agents for B2B in Mexico: how we designed an SDR that qualifies leads without cold-calling

By
~14 min read

An operating case from Transgenia itself, not from a client. After a cold outreach campaign that produced 193 invitations sent, about a 1% conversion rate, and zero meetings, we made an uncomfortable decision: shut down the motion and redesign it. This text documents the system that took its place, a sales agent governed by hard rules, and why we believe it is the replicable pattern for B2B SMEs in Mexico that want to use AI in sales without turning into an industrial spam generator.

Short answer. A B2B AI sales agent in Mexico, done well, is not an email auto-sender. It is a system with four pieces: (1) a CTA that offers an interactive demo instead of asking for a call; (2) a gate that requires a corporate email address (rejecting Gmail, Hotmail, Yahoo, and domains without MX records); (3) a weekly digest that translates interactions inside the demo into a prioritized list of informed calls; and (4) a human who authorizes each send. The combination moves the critical moment of the funnel from closing on a cold call to the beginning, where there is already a signal of intent before the phone is picked up.

By 2026 almost every B2B SME in Mexico has tried or considered using AI to sell more. The market's offering, broadly, splits in two: on one side, mass generators of personalized email ("our agent writes a thousand unique emails a day"); on the other, light integrations with existing outreach platforms that turn an LLM into a slightly variable template. Both share an assumption that is rarely questioned: the bottleneck of the B2B funnel is the volume of the first touch.

Our experience says the opposite. The bottleneck, at least for the average ticket of a Mexican trading company or private clinic, sits much further downstream: in the step from "interested reply" to "effective meeting", and from there to "signed proposal". An engine that fires more emails at the top of the funnel does not fix that bottleneck; it hides it under a louder noise of vanity metrics. This text describes how we reached that conclusion with our own data, and what we built afterward.

1. "The cold died at the close": autopsy of 193 invitations

In mid-2026 we ran with an external vendor a LinkedIn outreach campaign designed by them, executed by them, measured by them. The hypothesis was standard: send personalized invitations to decision-makers at Mexican SMEs and convert acceptances into discovery meetings. The experiment produced, according to the vendor's own report:

A 1% funnel is not a bad number in isolation for cold outreach; the problem is that the 1% evaporated between "replied with interest" and "accepted a meeting". The first contact did generate curiosity; the call was the wall. When we reconstructed the cycle with the campaign notes, the question left on the table was: why does a prospect who answers an invitation with "tell me more" not accept a call fifteen minutes later?

The most honest answer, and the one that gives structure to the rest of this text, is that the information asymmetry does not justify the call. The prospect does not know enough to decide to invest time, and the seller cannot transfer useful context in the little space available in a LinkedIn message. The call, then, is not a natural step; it is a leap of faith. Under normal conditions of the Mexican B2B market, almost no one takes that leap.

After an internal review we decided there were two paths: increase volume (more invitations, more A/B testing of the first message, more pressure on the closer) or change the motion. We chose the second. The new motion does not ask the prospect for a call; it offers an interactive demo with synthetic data that they can walk through in their own time. The call comes later, when there is already a measurable signal of intent.

2. What is a "governed" sales agent (operating definition)

There are two popular uses of the term AI SDR or AI sales agent in the market. One refers to a system that generates and sends emails almost autonomously; the human only approves the launch of a campaign. The other refers to a chatbot that responds in real time on the website. Neither of these definitions describes what we built.

For the purposes of this text, a governed sales agent is a system that combines language models, hard rules, and humans-in-the-loop, with this architecture of responsibilities:

This definition is not cosmetic. During 2024-2025 there were enough cases of AI agents that sent compromising emails at scale, in different parts of the world, for hard rules to become the serious differentiator between a system usable in production and a covert legal liability. A governed sales agent does not compensate with AI for what the human operator lacks; it gives the human operator a larger surface on which to apply judgment, without scaling errors.

flowchart LR
  subgraph AI[Language model]
    A1[Drafts messages]
    A2[Personalizes context]
    A3[Summarizes digest]
  end
  subgraph Rules[Hard rules]
    R1[Freemail blocklist]
    R2[MX verification]
    R3[Stop-on-demo-entry]
    R4[Turnstile antibot]
    R5[Immediate opt-out]
  end
  subgraph Human[Human-in-the-loop]
    H1[Sign-off per wave]
    H2[Approves final list]
    H3[Receives Monday digest]
  end
  A1 --> Rules
  A2 --> Rules
  A3 --> Human
  Rules --> Human
  Human --> Send[[Authorized send]]
  Send --> Effect[(demo_events real row)]
A governed sales agent: the model drafts, hard rules filter, the human authorizes. The system measures effect, not artifact.

3. Four verifiable components of the system

The system as it is deployed today has four components with explicit contracts. None is optional; removing one breaks the design.

3.1 The gated demo as a qualification filter

The primary CTA of the outreach is not "schedule a call". It is "enter the demo", with the URL https://transgenia.org/demos/ as the destination. The demo is a functional simulation of a real case from Transgenia's operation (specifically, the clinical follow-up we built on Odoo for a private bariatric clinic in Mexico City; aggregated and anonymized, with a fixed "SIMULATED DATA" badge). The prospect walks through it in their own time, looks at the flows, interacts with the synthetic data, and decides whether it is worth continuing.

The demo fulfills three functions that a discovery call cannot:

  1. It transfers context without cognitive cost to the prospect. A 15-minute video and a call both demand agenda; an interactive demo can be walked through in 5 minutes between meetings.
  2. It behaves as a passive filter. The prospect who takes the trouble to log in and navigate is qualitatively different from the one who opened an email and did nothing else. The system records the difference without asking the prospect a thing.
  3. It generates auditable evidence. Every interaction is registered in demo_events, a log that complies with the LFPDPPP (Mexico's federal personal data protection law) notice declared on the site's legal page. It is not an opaque cookie tracker; it is a record with an explicit commercial purpose.

Demo access is valid for fifteen days from registration. A daily cron purges expired access. Limited validity does two things: it closes the site's exposure surface and tells the prospect there is a window to act. The psychology of a soft deadline works here just as it does in any SaaS product with a free trial.

3.2 Mandatory corporate email

The demo access gate rejects three classes of email address:

This rule is hard by design. Rejecting a gmail.com means closing the door to potentially valid leads who do not yet have a corporate email; we take that cost consciously. The reason is that the demo generates a lead the system will push toward a human call, and that call has a cost (agenda, context, follow-up) that does not amortize on an individual representing only themselves. Corporate email does not guarantee buying authority, but it is the cheapest filter that separates a legitimate B2B contact from noise.

An interesting side effect: when the gate rejects a freemail address, the conversation with the prospect is not automatically closed. The page tells them, in plain text, that the demo requires a corporate email and suggests other paths (public documentation, direct contact). The demo's yes is a decision with a cost; the gate's no is a polite redirection. Both are signals.

3.3 The weekly digest

Every Monday at 08:00 Mexico City time, a cron consolidates the interactions recorded in demo_events from the previous week and produces a prioritized digest. The digest is made of these artifacts:

The digest is delivered by internal email to the sales team and to Efraín Carreón (managing director). It never leaves the organization. It is raw material for a human; it is not an automated action. A lead that appears with a high score in the digest is a candidate for an informed call, in which the seller already knows which section of the demo the prospect walked through the most, what questions their path implied, and can open the conversation at a useful point instead of from scratch.

This is the piece that replaces the traditional work of a human SDR. An SDR researches leads, prioritizes, and schedules calls; the digest does the research-and-prioritization part with data the human does not have time to aggregate manually. A human still makes the call, because no AI today has the sensitivity to read a Mexican B2B negotiation with the ambiguity and cultural context it requires.

3.4 Demo-CTA cadence and stop-on-demo-entry

The outreach that pushes toward the demo has a three-touch cadence: day 0, day 4-5, day 10-12. Touches are logged as d1, d2, d3 in the send log, distinct from the classic cold outreach cadence (1..4) that also exists in the system for cases where the demo motion does not apply.

The critical rule of this cadence is stop-on-demo-entry: if the lead appears in demo_events at any point in the sequence, the remaining touches are canceled automatically. The lead moves to the digest list and is handled with an informed call, not with more email. The reasoning is direct: sending a fourth generic email to someone who already entered the demo is a channel degradation.

Beyond stop-on-demo-entry, the cadence also respects:

sequenceDiagram
  autonumber
  participant V as Sender (send_email.py)
  participant L as Corporate lead
  participant D as Gated demo
  participant B as Log (demo_events)
  participant G as Weekly digest

  V->>L: d1 (day 0)
  Note over V,L: If the lead enters the demo,
stop-on-demo-entry
  L->>D: Registers (corporate email, OTP)
  D->>B: Row with interactions
  V--xL: d2 and d3 canceled
  B->>G: Monday 08:00 CDMX consolidation
  G->>V: Lead prioritized for informed call
Demo-CTA cadence with stop-on-demo-entry. A lead that enters the demo leaves the chain and enters the digest.

4. Five hard rules that keep it from becoming spam

A poorly designed sales agent degrades fast. Degradation is not mathematical (more volume, worse quality); it is cultural (operators get used to looking at the system as a fire hose). To avoid that curve, the system has five hard rules that are not relaxed even under pressure for results.

Rule 1: Draft-first, always. Every email wave is prepared as a set of drafts and a manifest listing recipients, subject lines, and attachments. Sending requires an explicit sign-off per wave. Nobody runs --send without having received that authorization. An operator in a rush who wants to "approve everything quickly" finds that the sign-off signature is literal (a standard phrase), not a comfortable button.

Rule 2: Zero fabrication of figures. The only numbers that appear in the email are those from the published case, aggregated and anonymized, with a citable source in the demo link. Naming a client in writing requires a new sign-off per case. Presenting third-party metrics as our own, even implicitly, is forbidden.

Rule 3: A single canonical sender. All outbound email passes through a single script (send_email.py) with the following built-in controls: dedupe by (lead, touch), suppression, complete log, footer with visible opt-out, daily cap. Sending commercial email through any other channel, ad-hoc script, or external tool is forbidden. The sender's discipline is what makes the rest of the rules auditable.

Rule 4: Correct attachments in the right touch. The first commercial touch (d1) attaches the presentation letter and the sector's service offer; subsequent touches are short and without attachments, except in specific cases. A commercial email without its declared attachments is considered incomplete and is not sent.

Rule 5: Opt-out is sacred. Any reply containing "opt out", "unsubscribe", "no me contactes", or equivalents is recorded in the suppression list and applied before the next send. Opt-out is not negotiated. Silence, on the other hand, is not interpreted as consent or as rejection: it is silence, and the cooling-off period treats it as such.

flowchart TB
  D1[Draft-first] --> V[Human sign-off
per wave]
  D2[Zero fabrication] --> V
  D3[One canonical sender] --> V
  D4[Correct attachments] --> V
  D5[Opt-out sacred] --> V
  V --> S[[Authorized send]]
  S --> L[(Auditable log)]
Five hard rules that converge on a single human gate. None is optional.

5. The two lanes: private clinics vs. B2B trading companies

The system does not sell "AI for SMEs" in the abstract. It runs two lanes with distinct value propositions, tickets, and messaging.

Private clinics lane (healthcare sector). The anchor case is digital clinical follow-up, the application the demo simulates. Average ticket around six figures MXN for diagnosis and setup, with monthly retainer. Sales cycle of weeks to months; typical decision-maker is the medical director or clinic owner. The demo works as the primary piece because the problem is tangible on screen (a follow-up chart with complete data vs. the loose sheet that exists today in most mid-sized clinics). Outreach enters through the corporate email of administrative or medical staff.

B2B trading companies lane. The anchor case is intelligent replenishment on Odoo, documented on a separate pillar of the site. Larger unit ticket, typically five figures USD for the initial setup project. Longer sales cycle; decision-maker is director of operations or general director. The medical demo does not apply here; the CTA of this lane redirects to the sector pillar /comercializadoras.html and to a QuickLook scheduled as a 20-minute guided call with prospect data on screen. The specific B2B demo is in backlog.

The separation between lanes is not a marketing luxury: it defines which hard rules apply with what weight. In the clinics lane, the 15-day demo validity is central; in the trading companies lane, what matters is the presentation letter attached in d1. Mixing lanes produces two symmetric degradations: offering the medical demo to a trading company (which will not use it) and asking a clinic for a QuickLook (which is not yet ready to pick up).

6. How success is measured

A commercial system without a measurable definition of success is indistinguishable from theater. To the demo-CTA motion we declared a formal goal in July 2026, with a six-month horizon.

Goal H2-2026 (declared internally). By the end of the semester we must meet at least one of these two criteria: (a) $10,000 USD in sales to the medical sector (private clinics), or (b) an individual ticket of $5,000 USD to a B2B trading company. Large pre-existing accounts (RackUSA, Qualtop) are explicitly excluded from the calculation by decision of the managing director.

The scoreboard that measures this goal is fed by three verifiable sources:

A number that is not on the scoreboard with a citable source does not, for the purposes of the goal, exist. This discipline is uncomfortable at first, because any seller is trained to add "conversations that are going well" to the forecast; the scoreboard only adds observable effect. Over time, the discipline pays: the October forecast is built on data that exists, not on wishful thinking.

A note on honesty: at the time this text is published, the demo-CTA motion has just officially launched. The goal measurement is in progress; this article does not report final results, it describes the system with which they are being generated. When there is consolidated evidence (positive or negative), we will publish it with the same level of detail.

7. What of this is replicable for other Mexican B2B SMEs

The system is not a patented recipe. Almost all of its pieces can be rebuilt with accessible components. What is proprietary, and does not transfer as a copy, are the policy decisions.

The replicable pieces:

The policy decisions that are not copied by instinct:

Our operating hypothesis, unproven but contrasted several times in conversations with Mexican SME owners in the services sector, is that the real competitive differentiator of an AI sales agent in Mexico in 2026 is not the language model it uses, but the quality of its hard rules. Models are commodity; policy is property.

8. Closing: why the demo-CTA motion changes the moment of truth

If there is one thing to take away from this case, it is that the traditional cold B2B outreach motion, in Mexico, died the moment the information asymmetry between seller and prospect became insurmountable at the call stage. A prospect who replies with interest to an invitation but does not accept a discovery meeting is not being rude; they are saying, in the language of daily SME operation, that they do not have enough information to justify thirty minutes of calendar time.

The demo-CTA motion relocates the moment of truth. Instead of demanding that the prospect take the leap of faith to the call, it hands them a demo they can walk through alone. If the demo matters to them, they enter; if it barely matters, they do not. In either case, the system learns. The call that follows, when it follows, happens on shared information: the seller already knows which section interested them, the prospect already has a common language with the seller. The conversation opens at minute forty of discovery, not at minute zero.

That shift in where the leap of faith sits is what a governed AI agent contributes to the B2B sales process in Mexico. AI does not replace the seller. It hands them a larger surface of shared context with the prospect, without scaling the errors of cold outreach. And, when the system is well designed, it does something a human SDR cannot do alone: it sustains the discipline of sending even on the days when the pipeline hurts.

If you recognize yourself in the problem, two honest paths: enter the medical demo (corporate email, fifteen days of access, no required call) to see the system in operation; or if your context is more B2B trading company than clinic, the Sectors › Trading companies pillar summarizes the analogous proposition we are building for that lane. For the background of the governed-agents stack that sustains this motion, the AI solutions pillar is the entry point.

Frequently asked questions

Does an AI sales agent replace the human SDR?

Not in the sense that is sometimes advertised. The governed AI agent this text describes replaces the repetitive work of an SDR (research, prioritization, drafting, sequence follow-up), but it does not replace the B2B conversation. A human still makes the call, because the Mexican cultural context and the sensitivity to read a mid-sized negotiation exceed what a language model can sustain in real time.

In terms of team structure: a human seller augmented by a governed agent can handle the pipeline that used to require a seller plus a human SDR, as long as the hard rules hold. If the rules are relaxed (sends without sign-off, figures without source, ignored opt-outs), the system degrades faster than a handmade one and causes reputational damage that is significantly larger.

Why require corporate email? Aren't you rejecting valid leads?

Yes, knowingly. An SME owner who only has gmail.com can be a valid lead; on average, however, the mix of leads that arrive via corporate email has a higher probability of converting to a paid project, and the mix that arrives via freemail has a higher probability of consuming time without closing. The rule is a statistical filter, not an individual disqualification.

In addition, requiring corporate email has three useful side effects: (1) it lowers the noise from bots and disposable addresses, (2) it signals to the prospect that the conversation is serious (a demo that requires corporate email looks more like a product than a piece of marketing content), and (3) it makes ownership verifiable: a [email protected] with valid MX is more traceable than a [email protected] without an owned domain.

How do you prevent the system from becoming a spam generator?

With the five hard rules described in section 4: draft-first, zero fabrication, single canonical sender, correct attachments, and sacred opt-out. The rule that does the heaviest lifting is the single-sender rule, because it concentrates all sends in one auditable script with suppression, dedupe, and daily cap. Without that chokepoint, any of the other rules can be skipped by accident or in a rush; with that chokepoint, skipping a rule requires deliberate action.

It also helps that the motion does not reward volume. A seller whose bonus depends on "emails sent" pushes toward spam; a seller whose bonus depends on closed projects with invoicing evidence has no incentive to saturate the channel. Compensation policy is part of the system's policy.

Does the same approach work in sectors outside technology?

It depends. The pattern (gated demo + digest + hard rules) is transferable as long as there is something demonstrable on screen. Sectors with a tangible physical product (manufacturing, distribution, retail) can replace the "demo" with a configurator or an interactive quoter with real prospect data; the rest of the system stays. Professional-services sectors with a non-digitizable deliverable (strategic consulting, boutique legal) have to redesign the first piece, because there is nothing palpable to walk through alone.

What does not change across sectors is the underlying mechanic: move the critical moment of the funnel from closing on a cold call to the beginning, with observable evidence of intent before the phone is picked up. That is the extractable lesson; the specific mechanism is contextual.

How common is it for a CEO to be involved in the hard rules of the system?

At Transgenia, the managing director signs every wave sign-off. It is not delegated. The reason is dual: (1) volume is low enough that the sign-off is not a burden (typically tens of emails per wave, not hundreds), and (2) the reputational risk of a misdirected send is high enough to justify that the final signature belongs to the person with skin in the brand.

In a larger organization, this sign-off can be delegated to an equivalent figure (commercial director, chief revenue officer), as long as clear ownership is maintained: a person with a name and last name is accountable for every dispatched wave. Diluting the signature in a committee is equivalent to not having one, because the system loses the human moment of stopping to look at the manifest before --send.

How much does it cost to implement a system like this?

The technical part is tractable. A backend with email OTP (PocketBase, Supabase), a weekly cron (any server with cron), an SMTP sender with the hard rules (a Python script of a few hundred lines), and a mid-range LLM for drafting and digest add up to a low infrastructure cost, typically under $200 USD per month for SME volume.

The expensive part is the discipline. Sustaining draft-first when the campaign runs late, sustaining zero fabrication when a competitor is exaggerating on LinkedIn, sustaining sacred opt-out when the prospect who unsubscribed looked like "a very good signal", all of that requires a trained operator and a decision apparatus that backs them. That cultural cost is what separates a system useful in production from an experiment that gets turned off in three months.

Sales Governed AI agents Mexico B2B Gated demo Operating cases
← Back to Blog