How Mercado Livre built a decade of AI in fraud and risk
Mercado Livre has run AI in fraud, risk, and recommendation for more than a decade. The system impresses with the result. What few people read is how it was built: not in a project, but in ten years of narrow wins, each with its own criterion.
Per coverage by Seu Dinheiro and REVIIV (2026), that AI maturity reached the point where the company reduced 119 positions with the advance of automation. The number marks the depth, not the start. The start was small.
What the system does for the operation
It decides, in real time, how much to trust each transaction. It approves the legitimate payment without friction, holds the suspicious one, and does all that without sinking the conversion of honest customers.
For the operation, it is a balance: every extra point of blocking drops good sales; every point fewer opens fraud. The system lives on that rope.
What the public record supports
Nobody outside has seen the architecture from the inside, and we will not pretend we have. What the record supports is the build pattern: a decade of models in fraud, credit risk, and recommendation, compounded over time, not bought in a bundle.
AI is everywhere in there. But Mercado Livre’s advantage is not a better model. It is ten years of its own data and measured criterion, which no competitor buys as a shortcut.
The non-obvious part
The common reading is “they invested heavily in AI”. The precise reading is different. They did not do a big-bang. They compounded narrow wins, each with an acceptance number: fraud rate, approval rate, chargeback, all measured before and after.
The discipline is the advantage, not the technology. Each model entered solving a specific decision with a clear criterion, and the set became a moat. Whoever tries to copy it by buying an “AI platform” copies the form and loses the content, which is the decade of measurement.
The real cost
You cannot buy ten years of your own data. The model layer, today, anyone can turn on. The expensive part is the labeled history of real fraud, the integration with payment and risk, and the temper of operating with criterion for a decade.
It is the same hidden cost that appears in build vs buy of AI capability: the asset is not the software, it is the measured time.
The version that fits your chain
You do not have Mercado Livre’s decade, and you do not need it to start. The version at 5 to 10 percent does not chase the whole system.
Take a risk decision that already costs you a number, chargeback in a segment, loss in a payment method, and attack only that, with criterion measured before and after. Compound from there, one narrow win at a time. It is the opposite of starting from technology with no criterion, which is how most start and fail.
Think about the risk decision your operation makes by eye today that already has a known cost. That number is your starting point, not the platform the competitor bought.
Tell me the result your operation wants, the specific number: chargeback at this, approval at that, loss at this. In one hour I will send back a sketch of the architecture sized for your case, with a build or buy marker on each piece, and the first narrow win you can measure in weeks. It is the kind of thing we deliver. If the shape confirms, the two-week Diagnóstico becomes a signed spec, and that document becomes the contract for the Implementação.