Mercado Libre's Verdi Puts GPT-4o in Charge of $450M in Disputes

Mercado Libre's new AI development platform Verdi is already autonomously resolving 10% of customer service mediation on one of the company's major sites — disputes that involve decisions worth $450 million annually and work currently performed by 9,000 human operators.

Verdi runs on OpenAI's GPT-4o, GPT-4o mini, and GPT-3.5 Turbo, and Mercado Libre, Latin America's largest e-commerce and fintech company, built it to let any of its 17,000 developers work with large language models without ever seeing source code. The platform serves more than 30,000 microservices across the company's operations.

The scale of what is at stake here is unusual for an enterprise AI deployment. Mercado Libre is the most valuable company in Latin America after 25 years of growth, and it operates in a crowded competitive market where customer service quality on high-stakes disputes — damaged goods, fraudulent listings, refund claims — directly affects buyer and seller trust. Handing those decisions to an autonomous AI system, rather than using AI to assist human agents, marks a notable shift in how large platforms deploy language models in production.

From cataloging to autonomy

Mercado Libre has used OpenAI's API for years across its operations. GPT-4 Vision tags and completes product listings, letting the company catalog 100x more products than before over two years. GPT-4 evaluates millions of product listings daily for fraud, pushing detection accuracy to nearly 99% for flagged items. GPT-3.5 and GPT-4 translate product titles and descriptions, adapting to nuanced Spanish and Portuguese dialects. GPT-4o mini powers automated review summaries that increase orders where the feature is available, and personalized push notifications drive engagement through product recommendations.

Those results were useful, but the company's leadership did not consider them transformative. "Using AI for cataloging and summarizing reviews is interesting and helpful, but not transformational," says Daniel Rabinovich, Mercado Libre's COO. "We needed a solution that could handle intricate, high-stakes tasks autonomously and more efficiently than human operators."

Verdi is the answer to that requirement. Its first major application is autonomous customer service mediation — the dispute resolution process between buyers and sellers that traditionally requires time-consuming human intervention. Verdi manages these interactions end to end, interfacing with external APIs and internal employees when necessary.

"With Verdi, humans become co-pilots to the AI instead of the other way around," Rabinovich explains. "This platform can autonomously decide on cases, manage complex interactions, and even handle tasks requiring a human touch, informed by security protocols and guardrails built into the platform."

That inversion of the widely used "AI as co-pilot" framing is the core of the story. Most enterprise deployments position AI as an assistant that drafts responses or surfaces information for human approval. Mercado Libre has built the opposite arrangement for a defined set of decisions, with humans stepping in on the edge cases.

Natural language as the entire development surface

Verdi's most distinctive design choice is that natural language remains the central mode of interaction throughout the development process. Developers never see the source code. Security, guardrails, and routing logic are built into the platform itself.

"When looking for a development platform, we didn't see one that would solve our need for fast iteration while guaranteeing scalability and security," says Sebastian Barrios, SVP of Technology. "So we designed our ideal AI platform using GPT-4, with a focus on lowering cognitive load and enabling the entire organization to iterate, develop, and deploy new innovative solutions."

The platform integrates language models, Python nodes, and APIs into a structured framework that Mercado Libre describes as resembling a neural network. GPT-4o's context understanding and multimodal capabilities let Verdi process diverse data types — a "proof of damage" node, for instance, can call GPT-4o with Vision to interpret what is broken or missing from a customer's item.

Skills can also connect to each other autonomously. In a product return scenario, a "return" skill can talk to the "shipping" skill without a human getting involved, with parameters determined by Mercado Libre's teams inside Verdi's framework.

Verdi operates within Mercado Libre's internal developer platform, giving developers a unified environment to create, deploy, test, and manage applications. Built-in unit testing capabilities let teams ensure quality from the start, and the architecture is designed to incorporate new AI models and capabilities as they become available — a hedge against the fast pace of model releases.

Next: logistics and Mercado Pago

The customer service results have already pushed the company to expand. Mercado Libre has started deploying Verdi to its logistics network to speed up shipping and reduce late deliveries. The company also plans to add more advanced AI capabilities, including search assistants and product recommendations, to serve Mercado Pago's 50 million active users each quarter.

"Verdi is designed to solve any problem," says Rabinovich. "It's a bold statement, but we're confident in its potential to transform our operations. From handling routine customer queries to managing high-stakes disputes, Verdi is setting a new standard."

The stakes extend beyond one company. If Verdi's autonomous mediation model proves durable at Mercado Libre's scale — millions of listings, tens of millions of fintech users, billions in transaction volume — it will provide one of the clearest production examples yet that LLM-based systems can make consequential operational decisions, not just assist the humans who make them.