Why Brazilian cases matter
It's easy to cite Google, Amazon, or Tesla when it comes to applied AI. But American or European cases carry contexts very different from the Brazilian one: distinct infrastructure, different regulation, different labor costs, and, especially, a consumer market with its own characteristics. The six cases below happened in Brazil, with Brazilian companies, facing Brazilian challenges.
Case 1: Itaú Unibanco — AI that prevents fraud in milliseconds
Challenge: Itaú processes more than 2 billion transactions per year. Credit card fraud represented losses of hundreds of millions of reais annually, with traditional rule-based detection systems generating many false positives — blocking legitimate customer transactions.
Solution implemented: In 2023, the bank deployed real-time anomaly detection machine learning models, trained on each customer's transactional behavior data. The system analyzes more than 200 variables per transaction in under 50 milliseconds — before authorization.
Result: A 35% reduction in credit card fraud losses and a 28% drop in false positives, according to data released by the bank in its 2024 sustainability report. The system processes 99.8% of transactions without human intervention.
Lesson learned: "The biggest obstacle wasn't technical — it was cultural. Risk managers with 20 years of experience had a hard time trusting a model they couldn't completely interpret." (Itaú CTO, Valor Econômico, 2024)
Case 2: Embrapa — computer vision to defend Brazilian soy
Challenge: Brazil is the world's largest soy producer. Pests and diseases account for 15-30% of production losses in bad years. Early diagnosis requires specialized agronomists — scarce and expensive for most rural producers.
Solution implemented: Embrapa developed AgroIA, a computer vision app that allows the producer to photograph a leaf with their cell phone and receive a pest and disease diagnosis in seconds, along with management recommendations. The model was trained on 2 million images labeled by Embrapa experts over three years.
Result: Over 180,000 downloads in 14 months of operation. 91% accuracy for the top 40 soy pests. Producers report diagnosis 3-5 days faster than the traditional method.
Lesson learned: Rural connectivity is a critical bottleneck. Embrapa had to develop an offline mode to function in areas without a signal — a feature that was not in the original plan.
Case 3: Magazine Luiza — AI that humanized customer service
Challenge: With more than 40 million active customers and demand spikes on dates like Black Friday, Magalu faced customer service queues of hours and falling satisfaction indices.
Solution implemented: Lu — Magalu's virtual assistant — was repowered in 2024 with large language models capable of resolving 87% of post-sale demands without transferring to a human. The system recognizes emotional tone and automatically escalates to a human agent in cases of detected frustration.
Result: 42% reduction in average demand resolution time. Customer service NPS rose 18 points. Cost per service dropped 31%. Lu now responds to more than 50 million interactions per month.
Lesson learned: The decision not to hide that Lu is an AI — and to invest in making it transparent about its limitations — increased customer trust and reduced complaints about the digital channel.
Case 4: Rede D'Or — AI that reduces imaging report time
Challenge: Rede D'Or, Brazil's largest private hospital network, faced a critical bottleneck in radiology. CT and MRI reports took 6-24 hours on average. In urgent cases, this time has a direct clinical impact.
Solution implemented: Partnership with Brazilian startup Nuveo to deploy radiological diagnosis assistance AI. The system analyzes CT and MRI images, flags suspicious findings, and automatically prioritizes urgent cases in the radiologist's queue. The AI does not issue reports — it flags and prioritizes; the doctor writes the report.
Result: Average report time for urgent chest CTs fell from 4.2 hours to 38 minutes in pilot units. Pulmonary nodule detection increased by 23%. Radiologist satisfaction: 78% positive in an internal survey.
Lesson learned: Initial resistance was high among radiologists. The turning point happened when a case of pulmonary embolism was detected by the AI 2 hours earlier than it would have been detected in the normal queue — the patient survived. "One concrete case is worth a thousand training slides," said the unit's medical director.
Case 5: iFood — AI that predicts demand and reduces waste
Challenge: iFood partner restaurants lost an average of 12% of ingredients due to overproduction. For tight-margin restaurants, this is critical.
Solution implemented: iFood launched the Smart Prediction tool in 2025. The model analyzes order history, local weather, regional events, day of the week, and seasonality to project demand per menu item 6 hours in advance.
Result: Average waste fell from 12% to 7.3% in 90 days. Operating margin improved by 2.1 percentage points. Voluntary adoption rate: 67% of eligible partner restaurants in 6 months.
Lesson learned: A simple interface was decisive. iFood tried an initial version with complex dashboards — adoption was 12%. The simplified version with direct recommendations reached 67%.
Case 6: Totvs — AI that democratized financial analysis for SMBs
Challenge: Brazilian SMBs — 99% of the country's companies — rarely have financial analysts. Cash flow, pricing, and investment decisions are made with little or no data support.
Solution implemented: Totvs integrated an AI assistant into its ERP for SMBs in 2024, Totvs Lia. The assistant answers natural language questions about the company's financial data. The model only has access to the client's own ERP data.
Result: 23% of Lia users report having made at least one different financial decision after consulting the tool within 90 days. ERP NPS rose 14 points. Totvs reported a 19% increase in SMB customer retention.
Lesson learned: The greatest value wasn't the analysis itself, but the confidence generated. Entrepreneurs who had never had access to professional financial analysis began making decisions more securely.
The pattern that emerges
Six cases, six different sectors — but three patterns appear in all of them: a specific problem before a generic solution; a real pilot before scale; and human resistance as a greater obstacle than technical limitation. AI did not fail in any of these cases for technical reasons. Where there was friction, it was for human reasons — and where it succeeded, it was because the teams understood this and worked on the human aspect as carefully as the technical one.
