How SuperFrete Uses AI Agents to Scale Without Scaling Headcount
SuperFrete is one of Brazil's leading shipping platforms for small and medium retailers. It enables anyone who sells — from e-commerce stores to WhatsApp and social sellers — to manage shipping in one place, with access to multiple carriers, competitive rates, tracking infrastructure, and integrated logistics support. With 130,000 active clients per month and over 1 million shipments processed monthly, scale is their daily reality.
When Lívia Rigueiral joined as CPO, the company had 35 people. Today they have 160. Growth has been rapid, but a structural question started keeping leadership up at night: what happens when you want to reach a million clients and a hundred million shipments, but growing the team at the same rate means more hierarchy, more management overhead, and more cultural dilution?
“We can't grow to a million clients if we keep growing the team at the same pace. The more the team grows, the more hierarchy grows, the more people management grows. It's all very hard.”
SuperFrete's answer was AI agents. Not as a buzzword or a press release. As actual production systems that handle real transactions, answer real customers, and catch real fraud. They started with an antifraud agent in mid-2025, then expanded to customer support by the end of the year. The results changed how the company thinks about growth.
Why start with fraud
SuperFrete's leadership had a clear principle: automate what you already understand. Fraud detection had been manual for years. The process was mapped, the pain was quantified, and the team knew exactly where it broke.
“Antifraud had been a massive hole in previous years. We said: if we can stop worrying about this, we can put our time, thinking, and intelligence into other things.”
The problem: 7.5 days of exposure
The existing process relied on daily BI reports, manual analyst review, external tools like ClearSale, and zero weekend coverage. When fraudsters found a new vulnerability, they had a free window until the team discovered the pattern and deployed a fix. That cycle took an average of 7.5 days.
Before: Manual Fraud Detection
- BI reports updated once per day (D+1 delay)
- Manual analyst review: 2-3 min per order
- No coverage on weekends or holidays
- 7.5 days average to detect new fraud patterns
- Team unable to analyze all transactions at volume
After: AI-Powered Detection
- Every transaction analyzed in real time
- Same-day detection of new fraud patterns
- 24/7 coverage with no blind spots
- Analyst focuses on strategy, not data collection
- ClearSale used only for high-risk second opinions
“When the fraudster changed behavior, we changed too. We can predict, we can prevent much more fraud, and we can recover fraud that we do take, because the moment we detect it, we stop the fraudster from completing it.”
How the agent works
New payment triggers the workflow
Every transaction on the platform fires a webhook that starts the analysis automatically. No human intervention, no batch delays.
Rule-based scoring + LLM analysis
Deterministic code checks order velocity, email patterns, payment history, and device fingerprints. An LLM handles ambiguous edge cases. This hybrid approach keeps costs low and accuracy high.
ClearSale as second opinion (not primary)
Only high-risk cases go to ClearSale for verification. The AI does the first filter using SuperFrete's own data, flipping ClearSale from primary decision-maker to a targeted checkpoint.
Slack alert → Human review
Flagged cases arrive in Slack with the agent's reasoning. The analyst reviews only these cases and provides feedback that continuously improves accuracy.
The results
Beyond the numbers, the team transformed. The fraud analyst who previously spent his days on data collection moved to strategic work: identifying new patterns and proposing system improvements. Fraud stopped being a topic in leadership meetings. The team now spends its energy on growth, not damage control.
Agent 2: Customer support under fire
The second agent didn't start with a strategy deck. It started with a crisis.
SuperFrete had been gradually rolling out an AI sales agent named Sofia, which was converting new users at 35%, matching human performance. The plan was to eventually expand Sofia to handle customer support too. Eventually. Slowly. With careful testing.
Then December 2025 happened.
The crisis
Support ticket volume had already been climbing: 1.5x from September to November. Then Correios, Brazil's national postal service, went on strike. Private carriers started delaying packages too. Ticket volume tripled. NPS dropped. The majority of detractors cited support as the reason. And SuperFrete was about to enter collective holiday leave on December 19th, with half the team going offline.
“It was almost a contingency scenario. Massive ticket volume, Correios on strike delaying a ton of packages, and we were about to go on collective vacation. I told the engineering team: we need to figure this out, we have to put her live and just go. The way things are, it can't get worse.”
Sofia went into full customer support mode. First, focused only on shipping delay queries (the highest volume topic). Then, as the results proved out, opened to all support categories. By February 2026, Sofia was handling 100% of incoming conversations as the first responder across all topics.
How Sofia works
Sofia is a conversational agent that operates on both WhatsApp and the in-platform chat, integrated with HubSpot. Each message triggers a workflow that gathers conversation history and client data, then generates a response using a knowledge base, operational tools (tracking, tickets, database queries), and strict behavioral guardrails. When she can't resolve something, she escalates to a human with full context already gathered.
One key design choice: SuperFrete's team edits Sofia's prompts and knowledge base directly, without engineering involvement. When Correios went on strike, they simply updated her instructions for handling delay-related questions. That operational autonomy is what makes the system sustainable.
The guardrail lesson
Deploying under pressure came with a hard lesson. Sofia started promising refunds and financial compensation to customers. For a logistics company processing over a million shipments per month, that's a serious liability.
The fix was elegant: a second, cheaper LLM acts as a guardrail, reviewing every message before it's sent to the customer. It checks for financial promises (refunds, chargebacks, compensation) and blocks them. The cost of running this guardrail across 5,000 messages: $0.30.
AI agents will make mistakes. The question is whether your architecture lets you catch and correct them quickly. SuperFrete's guardrail was built and deployed within hours of discovering the problem. The system improved permanently because of a real-world failure.
The results
When Sofia was handling sales conversations, her conversion rate matched humans at 35%. But something unexpected happened: the human agents' conversion rate jumped from 35% to 48%. Because Sofia pre-qualified leads and handled the routine questions, humans had more time per customer. The AI didn't just match human performance. It made human performance better.
“Nobody thinks she's AI. Everyone thinks she's human. And I need our customers to keep thinking that, because they really like her.”
What SuperFrete learned about AI adoption
SuperFrete's experience reveals a pattern that applies beyond logistics. There are a few principles that shaped their success.
Start with what you know
SuperFrete chose antifraud because they already understood the process deeply. AI works best when the problem is well-mapped and the team can evaluate results. Don't start with your most ambiguous problem.
Measure impact, not automation
Every project had a dual mandate: reduce manual work AND improve outcomes. The antifraud agent didn't just replace analysts, it caught 70% more fraud. Sofia didn't just answer faster, she improved human conversion rates.
Give the team ownership
SuperFrete's team edits prompts and knowledge bases directly. When new carriers join or policies change, they update the AI themselves. This makes the system sustainable, not dependent on external engineering.
Accept that it won't be perfect
Sofia promised refunds she shouldn't have. The antifraud agent had false positives. Both improved through rapid iteration. Starting imperfect and improving beat waiting for perfection.
“It wasn't about reducing the number of people. We just want to grow much more than we grow the team. And once they started seeing the opportunities, the fear went away.”
What comes next
Both agents were built on deco's platform, which provided the workflow engine, MCP integrations, AI gateway, and the interface that lets SuperFrete's team manage prompts and knowledge bases without engineering involvement. SuperFrete has since started building additional agents on the platform independently. The next one is an audit agent that will analyze the data produced by both existing agents, surfacing patterns and opportunities the team wouldn't have time to find manually.
“The question is not whether to implement AI. It's how. Pick the thing that will have the most impact, figure out what risks you're willing to take, and just start. Because as long as we're just listening to other people's stories instead of living it ourselves, we won't understand the real scale of what's possible.”
Building AI agents for your operation?
We're showing live demos and discussing real implementation patterns at VTEX Day 2026. Come talk to us about your use case.
