Building agentic infrastructure for Latin American commerce

In AI-enabled commerce, market position is rarely a feature problem. It is a data and distribution problem—and both compound exponentially over time.

Shipping an agent layer twelve months late does not mean taking twelve months to close the technical gap. It means missing twelve months of unrecorded seller behavior, twelve months of exclusive platform partnerships, and twelve months of presence inside the interfaces where merchants make decisions.

On July 28, 2026, we launched JoomAI in Brazil. We chose to go to market early rather than wait for feature parity in 2027. In software infrastructure, features can be copied; lost compounding time cannot be recovered.

The operational bottleneck in marketplace commerce

For the past two years, AI in e-commerce was restricted to operational surface work: copywriting, basic image editing, and customer replies.

However, on high-leverage decisions—which niche to enter, what to sell, and how competitors are priced—models hit a wall: a complete lack of access to real platform data.

Today, the average marketplace seller spends only 30% of their time actually selling. The remaining 70% is consumed by manual work: researching products, checking competitor prices, recalculating margins, and monitoring catalog listings.

JoomAI collapses this daily routine into automated background tasks.

An assistant with market context pre-installed

Frontier LLMs are now available to everyone. Model choice is a UX and cost decision, not a defensive moat. The moat lies entirely in the underlying data:

• Cross-marketplace demand: Real-time pricing, sales estimates, niche saturation, and seasonality across Mercado Livre and Shopee.
• Supply economics: Sourcing data from the Joom ecosystem, including supplier costs, landed DDP economics, delivery timelines, and factory reorder rates.
• First-party store data: Direct connection to a seller's store, so the system evaluates real inventory and active listings rather than abstract scenarios.

Consider a single query: "Analyze my sales, identify my best-selling product, and find comparable products in the supply catalog at a workable margin."

Executing this requires bridging three distinct data sources simultaneously. Very few players hold both sides of the transaction; almost none bridge them directly into the seller's live store.

Control loops over chat wrappers

The term "agent" is frequently overloaded. Our definition is specific:

An agent is not a single model call. It is a loop: it evaluates context, decides on an action, invokes a tool, checks the result, and repeats until the task is complete.

JoomAI runs more than 100 task-specific agents covering pricing, listings, competitor tracking, inventory, and demand opportunities 24 hours a day. They run on a schedule and alert the seller automatically.

A chat interface is a surface feature that can be copied in a quarter. A catalog of loop-based agents wired into proprietary supply data and live store connections functions as an operating system for a business—and it accrues switching cost with every run.

Stress-testing in high-friction markets

We intentionally launched JoomAI first in Brazil. Brazil is the largest e-commerce market in Latin America (R$351 billion in 2024, growing ~20% a year), but it is also one of the hardest markets in the world to operate in:

• A cascading tax system that can double landed costs.
• Complex customs processes that stop businesses at the border (64% of local sellers report import complexity as their biggest barrier to growth).
• SME credit priced at 30% to 45% a year.

Our strategy is simple: an AI operator that masters the market that punishes mistakes hardest can work anywhere.

Distribution at the workflow layer

Simultaneously with our platform launch, JoomAI published an official e-commerce data connector listed directly in Anthropic's integration directory—the first such connector for an e-commerce data platform in Brazil.

Instead of forcing sellers to open a new application, JoomAI pipes live market data directly into Claude and ChatGPT. Sellers can run market queries in natural language without leaving their existing chat interface.

This distribution position enabled our exclusive analytics partnership for Shopee sellers in Brazil (covering ~325 million monthly visits and 80,000 to 100,000 professional sellers). Combined with Mercado Livre coverage, sellers can now analyze both of Brazil's major marketplaces in one place.

Measurable seller outcomes

We evaluate JoomAI on seller-level economic outcomes rather than model benchmarks:

• 31% faster annualized growth for active sellers.
• Around one month to a measurable increase in profit.
• 3x faster to the first ten sales.
• 10+ hours a week returned from manual analysis to the operator.
• 70% less risk on product investment decisions.

More than 100,000 sellers use the platform today.

The comfortable strategy in AI is to wait for the category to define itself, then enter with a better product. It fails for a specific reason: by the time the category is clear, the data and distribution have already been allocated. We prefer an evolving infrastructure lead over a complete second place.

Here’s some test content for display.