AI Agents for SMBs in LATAM: Why Pygma Selected xcale for PY9

Pygma selected xcale for PY9, its ninth — and final — cohort. The thesis comes before the news, because the thesis is what got us in: AI agents for SMBs are not enterprise software at a discount. They are a different category, with a different buyer and a different channel. A Latin American small business does not buy a platform and then configure it for three months. It hires someone when the work overflows the team. That difference — buying a tool versus adding someone who does the work — defines everything we build.

This is what we argued in the application, and what we will defend at Demo Day.

What PY9 is, and why it matters to us

Pygma is a pre-seed accelerator for Latino founders in fintech and applied AI, based in New York and Bogotá. PY9 is its ninth cohort and, by its own announcement, the last one it will run before becoming a fund. The program runs six weeks and closes with a Demo Day in front of funds from the Silicon Valley ecosystem, in San Francisco. The figures Pygma publishes about its portfolio: 148 startups accelerated across 17 countries.

The badge is not the point. The point is that the program is built around the question we have been asking since day one: how do you sell AI software to a business that has never bought software?

That is not a product question. It is a question about distribution, price and time-to-first-value, and getting any of the three wrong kills the sale before the product ever matters. Six weeks with people who have watched 148 attempts at it beats six months guessing.

An SMB does not need another tool — it needs the work done

Our category is almost always described badly. "Chatbot." "WhatsApp automation." "Conversational flow." Every one of those words describes a tool that somebody has to configure, maintain, and repair the moment a customer writes something the diagram did not anticipate.

The owner of a clinic, a store or a distributor does not have that somebody. There is no operations team drawing decision trees. There is a WhatsApp inbox with 200 unread messages and two people answering between other tasks.

So we start from a different noun: the agent. Not a flow that fires, but a role with a job description — salesperson, appointment setter, support, CRM manager — that understands what it is asked, consults what the business knows, executes the action, and escalates to a person when it should. AI does not replace the team: it absorbs the work nobody has time for and hands the team back the conversations where their judgment actually changes the outcome.

The difference shows up in the only metric that matters early: how long between a business signing up and its first customer getting a useful answer. If that takes weeks, the SMB has already gone back to answering by hand. You can see how this is built in the platform, and what each role does in the five agent types.

WhatsApp is not another channel — it is the counter

In the United States, WhatsApp is a messaging app. In Colombia, Mexico, Brazil and Argentina it is where commerce happens. Customers ask for the price on WhatsApp, send the product photo on WhatsApp, book on WhatsApp and complain on WhatsApp. No contact form competes with that.

That changes the product design at the root. A CRM that lives outside the conversation will always be stale, because nobody transcribes a chat by hand. An agent that does not remember the customer from three months ago forces the person to repeat themselves — and repeating yourself is exactly what makes a customer decide to call instead. Which is why persistent memory and the CRM are not separate modules in our architecture: they are the same conversation seen from two angles.

It also changes the economics. Meta bills per conversation, not per seat, so cost scales with actual business volume rather than headcount. For an SMB that is the difference between a fixed expense you have to justify and a variable cost that pays for itself. The plans are published on pricing.

We learned it running our own companies

We did not study this problem from the outside. Before xcale I founded an aesthetic medicine clinic in Armenia, Colombia, and Nevatal, the practice-management SaaS that medical clinics run on. The clinic taught us which conversations an AI can carry end to end, and the exact point where a human has to step in. Nevatal taught us where the system of record ends and the conversation begins — which is precisely where most vertical software falls over.

The clinic today runs with an xcale agent handling patients on WhatsApp. It is our first customer and our harshest test bench: when something breaks, it breaks against real patients of a business that is also ours. That is an uncomfortable constraint, and by some distance the one that has improved the product most. The full story is in about us.

Three bets that can be proven wrong

A thesis that cannot be wrong is worth nothing. These are the three we hold, written so they can be measured against real customers.

One: time to the first useful answer beats depth of configuration. We are betting a small business prefers an agent that answers well on day one over a platform that will answer perfectly in a month. If it turns out businesses will tolerate weeks of setup in exchange for more control, we are wrong, and the product should look more like a platform than an employee.

Two: the role beats the flow. We hold that describing a job — what it sells, what it may promise, when it hands the chat to a person — produces better conversations than drawing a decision tree. If businesses end up asking for the tree anyway, because they want certainty about every single answer more than good answers on average, then the agent abstraction is surplus and we should say so.

Three: the moat is memory, not the model. Models get cheaper and converge every quarter; what an agent knows about your customers after a thousand conversations does not. If that is false, anyone with access to the same API catches us quickly.

All three are testable, and all three go under pressure during the program. We would rather publish them now, dated, than explain them afterwards.

What comes next

Six weeks of program and a Demo Day in San Francisco. In between, the usual: talk to more businesses, shorten the time to the first useful answer, and hold the line on not promising results we have not measured.

If you run an SMB in LATAM and you are evaluating AI agents for SMBs in your own WhatsApp, the best way to have an opinion on this thesis is to test it against your real conversations and tell us where it breaks.