EN22 Apr 20269 min read

I run an AI company from Guatemala. Here's the LATAM playbook nobody in Silicon Valley is writing.

WhatsApp-first sales. Spanglish models. Dollar pricing vs. quetzal pricing. Sales cycles measured in voice notes. The playbook for shipping AI to LATAM businesses doesn't come from SF. Field notes from Mozt, building from Guatemala City for the rest of the region.

When I tell US founders we are building Mozt from Guatemala, the second question is always some variant of "for now, right?" The first question is whatever they think they should ask. The second is the honest one.

Not for now. Permanently. And it's not a hedge or a fallback or a step on the way to relocating. The market is here. The customers are here. The work earns its weight here.

This is the post I've been meaning to write for six months. It's the playbook nobody in San Francisco is going to publish, because nobody there has the receipts to publish it. I'm 21, I run an AI ops company from Guatemala City, and what follows is what we've actually learned shipping to clinics, restaurants, and operators across Central America. Every line of this is a lesson with a scar attached.

If you're building AI for LATAM, save this. If you're a US founder thinking about the region, read this twice.

1. WhatsApp is the operating system

Not a channel. Not a "messaging app." The operating system.

The first time we tried to onboard a clinic to a web dashboard, the doctor opened it once, said "está muy bonito," and never logged in again. We tried email. Nothing. We tried SMS. The notifications got buried.

Then we wrapped the same agent in a WhatsApp Business number. The clinic checked it within four minutes. They've never stopped using it.

In Guatemala, Mexico, El Salvador, Honduras, and most of South America, WhatsApp penetration is effectively 100% among working adults. Banks send statements there. Hospitals send appointment reminders there. Customers DM brands there expecting same-day replies. If your AI product cannot operate inside WhatsApp, your CAC will be three times what it should be and your retention will be a third of what it could be.

What that actually means in product terms:

Your inbound flow is voice notes, image dumps, PDFs forwarded from a phone, and Spanglish typing. Not structured forms.
Your outbound has to handle media, audio, and threaded replies. Plain text isn't enough.
Your billing reconciliation has to read receipts photographed and sent to your support number. Customers will not log into a web app to dispute charges.
Your support agent has to answer in seconds during business hours, not in 24-hour SLAs. Customers expect human-pace responses.

US AI tooling assumes a typed-into-a-chat-window primitive. LATAM tooling needs a multi-modal-multi-channel primitive that's sticky on a phone. Build for the phone first, the dashboard second, or skip the dashboard entirely.

2. Spanglish is not a bug, it's the spec

Customers in Guatemala City do not write monolingual Spanish. They mix English technical terms in mid-sentence. "Hola, necesito el report de las consultas del Q1, please." That's a real message we got last week. Treating that as malformed input is how you ship a product that fails 30% of real conversations.

The fix is mechanical. Your prompts and your evals need to handle code-switching natively. Models in 2026 are good at this; the bottleneck is that most teams test only on monolingual prompts in dev and discover the failure mode in production.

Specific techniques that work:

Use the exact phrasing your customers use in your few-shot examples. Don't "clean it up" before inserting.
Run an eval on a sample of real production messages, not synthetic ones. The synthetic ones will be too clean.
Allow your agent to reply in the same code-switched register the customer used. Forcing it to "respond in formal Spanish" makes it sound like a government form.

The brand voice on your AI is not a translation problem. It's a register problem. Get the register wrong and you sound like a corporate chatbot in a country that prefers messaging the owner directly.

3. Dollar pricing vs. local pricing is a real strategy decision

The default founder instinct is to price in USD and let the customer convert. This is wrong for a non-trivial subset of LATAM customers.

Three reasons:

Tax friction. Local invoices need local currency for clean accounting. A USD-only invoice introduces a manual step every billing cycle.
Psychological cost. "39 dólares" lands differently than "300 quetzales" or "700 pesos." Even when the numbers are roughly equivalent, the local-currency framing reduces sticker shock by ~20% in our split tests.
Volatility. USD-pricing exposes the customer to currency fluctuation they can't hedge. They notice. It poisons the relationship.

What works in our experience: USD-denominated B2B contracts at the enterprise tier (where the buyer is sophisticated and probably already has dollar exposure), local-currency pricing at the SMB tier with a quarterly USD reference rate baked in. The split is roughly 80% local-currency at the SMB tier we serve.

This is operational engineering, not finance. If your billing system can't handle multi-currency cleanly, fix it before you scale.

4. The sales cycle is measured in WhatsApp voice notes, not Google Meet calls

The default US-startup sales motion is: cold email → discovery call → demo → proposal → close. Average cycle: six to twelve weeks for SMB.

The LATAM SMB motion is: a voice note from a friend-of-a-friend → a WhatsApp demo (a recorded screen-share, not a live call) → another voice note with three more questions → a quote shared as an image of a PDF → a deposit sent through a local bank transfer. Average cycle: four to ten days.

The trade-off: the LATAM motion needs personal trust to start. You can't blast cold emails. But once trust exists, the cycle compresses by an order of magnitude. The currency is reputation, and reputation propagates through the same channels the work happens on.

Practical implication: hire (or be) someone who can voice-note in the customer's register, on weekends, fluently. Your SDRs from a US playbook don't translate. Your founder might.

5. The legal and banking constraints are real and unevenly distributed

Stripe doesn't reach Guatemala the way it reaches the US. Mercury doesn't either. The default YC stack — Stripe + Mercury + Modern Treasury + a US LLC — works for the US-targeted side of your business and breaks for the local side.

What we ended up with:

A US Delaware LLC for global contracts and the dollar-denominated tier.
A Guatemalan SA for local invoicing, local taxes, and the local-currency tier.
Two parallel banking stacks. One in the US (Mercury), one local (a regional bank).
A reconciliation pipeline (boring spreadsheet plus a Claude agent) that maps between them.

Other LATAM operators run similar splits with different country flags depending on where their customers are. The shape is the same: dual entity, dual banking, parallel stacks. Build it on day one. Trying to retrofit it after you've signed your first big enterprise contract is a tax and accounting nightmare.

6. Banks and consultancies are the channel, not VCs

Santander, Banorte, BBVA, Citibanamex, Itaú. These are your channel partners in LATAM. They have the customer relationships, the brand trust, the regulatory cover, and the existing footprint. The default US channel-partner playbook (system integrators, marketing agencies, App Store / Cloud marketplaces) doesn't work cleanly here.

In March 2026, Santander partnered with Visa to ship LATAM's first end-to-end agentic payments system. NTT Data is partnering with AWS for managed AI services across the region. BDO Colombia is shipping internal agents on Microsoft Copilot Studio.

If you are building AI products for LATAM B2B, find the partnership team at the regional bank, accounting firm, or systems integrator before you find the venture firm. The capital is downstream of the channel here.

7. The talent argument cuts both ways

The standard objection is "you can't build AI from LATAM, the talent isn't there."

Half right.

The volume of senior AI engineers per capita is lower than the Bay Area. That's true. The regional AI-readiness number is around 14% according to recent industry data, versus much higher in NA / EU. The talent pipeline is narrower.

What's missed: the operational talent that AI for B2B actually requires is denser here than the engineering talent is sparser. Industrial engineers, operations managers, process designers, restaurant chain operators, healthcare administrators. The people who understand how the work actually runs in a clinic or restaurant chain are everywhere. They're cheap to hire by SF standards. They speak the customer's language. They know the WhatsApp register.

The 2026 reframe: the bottleneck of LATAM AI is not "we need more engineers." It's "we need to pair the engineers we have with operations talent that the US doesn't have to the same density." The combination is the wedge.

8. The thing that changes everything is being here

Every advice list ends with a meta point and this one is mine.

If you're building for LATAM operations from anywhere else, you are guessing at the operations layer. You are reading screenshots. You are asking customers for product feedback in a structured format that flattens the texture of how they actually work.

If you are here, you are observing. The clinic is a Toyota or a tuk-tuk away. The restaurant chain's CFO might be in your church group. The receptionist whose Saturday you saved by automating WhatsApp triage will tell her cousin who runs the dental clinic on the next block, and your CAC for the next deal just dropped to zero.

That's the unfakeable advantage. The customer is not a screenshot in a product requirements document. She is the woman who runs the front desk three blocks from your apartment.

Build from here. Or hire someone who lives here. There is no third path that ships products that work.

If this post was useful and you're thinking about LATAM, find me on LinkedIn or Instagram. I read every message. If you forward this to a US founder you know who's about to ship something to LATAM with a translated copy of their US dashboard, you might save them six months.