Mercado Libre: the GPT-4o-powered Verdi platform — a 100x catalog expansion and nearly 99% fraud-flag accuracy

• Mercado Libre introduces Verdi, an AI developer platform powered by GPT-4o — OpenAI (customer story), 2024-09
• Mercado Libre Unveils Verdi: An AI Developer Platform Powered by GPT-4o (независимый пересказ кейса, дата 25.09.2024) — AITech365, 2024-09-25

Background

Mercado Libre is Latin America's largest e-commerce and fintech company, with a 25-year history and the status of the region's most valuable company. Its engineering scale rivals big tech: 17,000 developers maintaining more than 30,000 microservices, plus the Mercado Pago fintech arm with 50 million active users per quarter.

The company has used AI on the OpenAI API for years — and quite 'boringly', in the best sense. GPT-4 Vision tags and completes product listings: over two years this enabled cataloging 100x more products than before. GPT-4 evaluates data on millions of listings daily for fraud — accuracy on flagged items has been pushed to nearly 99%. GPT-3.5 and GPT-4 translate product titles and descriptions with attention to the dialect nuances of Spanish and Portuguese — for a market stretching from Mexico to Argentina, that is not a detail but a precondition of buyer trust. Automated review summaries on GPT-4o mini help users grasp product feedback quickly and increase orders where the feature is available; personalized push notifications lift engagement.

The competitive context is worth weighing too: Latin American e-commerce is a market where Mercado Libre withstands pressure from global players, and margin is won through operational efficiency at enormous volumes — millions of listings, disputes, deliveries, and payments daily. In such an environment AI is not an image project but a tool for holding the lead, and the company consistently applied it where the effect is measurable.

On this foundation the company built Verdi — an internal AI application development platform on GPT-4o, GPT-4o mini, and GPT-3.5 Turbo, presented in an OpenAI case study in September 2024. The Verdi story is interesting as a rare documented example of a company moving from 'AI as a set of point features' to 'AI as a platform layer' on which any of 17,000 developers can assemble applications — up to and including autonomous mediation of money disputes.

Problem

Point AI applications delivered value but hit a ceiling. Company COO Daniel Rabinovich states that ceiling undiplomatically: 'Using AI for cataloging and summarizing reviews is interesting and helpful, but not transformational. We needed a solution that could handle intricate, high-stakes tasks autonomously and more efficiently than human operators.'

The most telling such task is mediation of disputes between buyers and sellers. It is a contentious, expensive, poorly scalable process: a human operator must figure out what is broken, who is right, what the marketplace rules say, contact the parties, and make a decision with monetary consequences. Traditionally this demands lengthy human involvement, and the potential perimeter is enormous — the tasks of 9,000 operators and decisions worth $450 million a year.

The second problem is development speed and security. The company evaluated existing AI platforms and found no combination of fast iteration, scalability, and security: '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. With 17,000 developers, every home-grown AI bot with its own credentials and prompts is a separate security review; without a platform layer, mass AI development either grinds against bureaucracy or becomes unmanageable risk.

Solution

The answer was Verdi — a platform layer on GPT-4o, GPT-4o mini, and GPT-3.5 Turbo, designed, in Barrios's words, 'with a focus on lowering cognitive load and enabling the entire organization to iterate, develop, and deploy new innovative solutions.'

Verdi's main architectural difference from analogues: natural language remains the central mode of interaction throughout development — developers never see the source code. Security, guardrails, and routing logic are built into the platform itself rather than re-solved in every application. An application is assembled as a structure resembling a neural network: nodes and skills that compose with each other. A 'proof of damage' node calls GPT-4o Vision to determine from a photo what is broken or missing from a customer's item; a 'return' skill talks to the 'shipping' skill on its own — no human involved; the parameters of these interactions are set by Mercado Libre's teams within the platform's framework. Beyond language models, Verdi integrates Python nodes and external APIs, operates inside the internal developer platform (create, deploy, test, and manage applications in one environment), includes built-in unit testing, and is designed to plug in new models as they become available.

The first major application was autonomous customer-service mediation: the AI conducts interactions with the parties, calls external APIs, and hands cases to human employees when necessary. Tellingly, the platform's debut task was the most inconvenient one — where multimodality (photos of a broken item), money (refunds and compensations), the emotions of conflicting parties, and marketplace rules all converge. Within a few months, Verdi began handling 10% of mediation on one of the company's major sites. Rabinovich describes the platform's philosophy: 'With Verdi, humans become co-pilots to the AI instead of the other way around. 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.'

Riding the customer-service success, the company began extending Verdi to its logistics network — for faster shipping and fewer late deliveries — and plans further expansion into search assistants and product recommendations, including for Mercado Pago's 50 million quarterly users. Rabinovich states the ambition maximally: 'Verdi is designed to solve any problem. It's a bold statement, but we're confident in its potential.'

Result

Per figures published by OpenAI: GPT-4 Vision enabled cataloging 100x more products within two years; fraud-detection accuracy reached nearly 99% for flagged items; within a few months Verdi took over 10% of customer-service mediation on one of the company's major sites; review summaries increase orders where available (with no effect size disclosed). The company estimates the potential at supporting the tasks of 9,000 operators and autonomously managing customer-service decisions worth $450 million annually.

A separate, often-skipped result is the economics of development speed: the platform was designed 'with a focus on lowering cognitive load' so that any team can assemble, test, and deploy AI applications in a unified environment with built-in unit testing. The company does not quantify this effect, but it is precisely what explains building a platform instead of hiring more operators.

Framing is mandatory. All figures are the company's self-report in a case study by OpenAI — the party selling the models used; there is no independent verification. The 100x is growth in cataloged products, not revenue or listing quality. 'Nearly 99%' refers to accuracy on flagged items — the metric says nothing about recall (how much fraud is never flagged at all). The 10% of mediations applies to one site, not the whole group; and '9,000 operators' and '$450M' are stated potential, not achieved results. Conflating these categories is the classic mistake in reading vendor case studies.

In our view, the case's key meaning lies in the sequence of admitting AI to money. Mercado Libre first spent years running models on tasks with a cheap cost of error (catalog tags, translations, summaries), accumulating data, infrastructure, and organizational trust — and only then admitted AI to disputes with monetary consequences, and in fractions: 10% on one site, with escalation to humans and a rollback path. This is the opposite of the fashionable 'autonomous agent in production within a quarter' — and probably the only way to do such things in a company where AI decisions directly move other people's money.

A second observation: 'developers never see the source code' is Verdi's most debatable and most interesting decision. We read it as a bet on radically lowering the entry barrier (any team assembles an AI app from nodes and skills) in exchange for total dependence on the quality of platform guardrails. With 17,000 developers the bet looks rational: centralized security scales, reviewing every home-grown bot does not. But we would not advise porting the pattern into an organization without a mature platform team.

Lessons learned

1. The mature path to agents runs through years of 'boring' applications: catalog, fraud, and translation gave the company data, infrastructure, and trust before AI was allowed near money disputes.
2. A platform instead of point bots: when 17,000 developers share a layer with built-in guardrails and routing, each new AI app stops being its own security review.
3. Autonomy is introduced in fractions: 10% of mediations on one site is a controlled experiment with human escalation and a rollback path, not 'switch everything to AI.'
4. Separate achieved results from potential: the 100x catalog is a fact, while '9,000 operators and $450M' is an estimate of future scale; conflating the two is the classic mistake when reading vendor case studies.
5. Multimodality resolves real disputes: a 'proof of damage' node using GPT-4o vision replaces the most contentious step of mediation — manually reviewing photos of a broken item.
6. Accuracy on flags (~99%) is not the whole truth about fraud detection: the metric is silent on recall and must be read alongside the question 'how much is never flagged at all?'
7. 'Developers never see the code' trades entry barrier for dependence on platform guardrails; the bet is rational with a mature platform team and dangerous without one.

Frequently asked questions