Decagon CEO Jesse Zhang opens São Paulo office with MercadoLibre as anchor client, targets LATAM market
Sep 17, 2026 with Jesse Zhang
Key Points
- Decagon opens São Paulo office with MercadoLibre as anchor customer, replicating the expansion playbook that put Deutsche Telekom first in Europe before committing London resources.
- The AI voice agent startup identifies model capability as solved; the constraint is mapping enterprise environments from scratch, shrinking deployment timelines across airlines, banks, and e-commerce platforms.
- Decagon substitutes smaller open-source models for frontier LLMs on low-complexity tasks like routing and classification, optimizing for voice latency rather than cost per inference.
Summary
Decagon is building AI voice agents for enterprise customer support, and CEO Jesse Zhang is expanding the footprint beyond the US faster than the product rollout might suggest is warranted. He's in São Paulo this week to launch a LATAM office, with MercadoLibre already signed as an anchor client.
“I'm in Sao Paulo this week. We're out here launching our LATAM office. We're starting office here in Sao Paulo. Got a couple customers down here already like MercadoLibre. So we really felt like things are moving pretty quickly. ... Europe was the first sort of pillar for us where we launched a London office and we have customers in Germany and in the UK.”
International expansion playbook
The sequencing mirrors what Decagon did in Europe. Deutsche Telekom was the anchor there, with the company initially serving them from a US team before the volume of activity made a London office necessary. Brazil followed the same logic: get customer pull first, then commit resources. Zhang says the shortlist for LATAM expansion came down to Australia and Brazil, evaluated on tech adoption speed, market size, and serviceability. Both made the cut. Decagon now has a small pod in Australia and is launching properly in Brazil.
The São Paulo office is primarily a go-to-market build. Language is a real constraint for voice agents, and Zhang says Brazilian Portuguese fluency has to be strong for the product to hold up. The team on the ground is mostly sales and customer-facing, not engineering.
Beyond MercadoLibre, Decagon's customer base includes Deutsche Telekom, American Airlines, and Delta Airlines.
Why deployment is slow
Zhang's answer to why AI customer service is still mostly phone trees is that the models aren't the bottleneck anymore. The analogy he uses: deploying an agent is like building a self-driving car where the car's brain is already good, but the map of the surrounding environment hasn't been drawn. At a large airline or bank, that map has to be built from scratch for every deployment. Decagon's job is to shrink the time it takes to draw it.
Cost and latency
On model economics, Zhang describes a standard optimization arc. You launch on frontier models to get the product working, then identify which tasks inside the agent don't need that full capability. Detecting hallucinations, topic classification, routing decisions can all be handled by smaller, open-source models, sometimes fine-tuned. In Decagon's use case, Zhang says that substitution has worked. For voice specifically, latency matters more than cost, and the primary lever is model size, not chipset or edge infrastructure, at least for now.
Every deal, every interview. 5 minutes.
TBPN Digest delivers summaries of the latest fundraises, interviews and tech news from TBPN, every weekday.