Problem
Banco Daycoval says an AI agent multiplied sales of payroll-deducted loans by 12 from the operation’s starting level. The number has an identifiable source. It does not have enough context to establish better underwriting.
Folha de S.Paulo attributes the claim to Alessandro Segura, Daycoval’s superintendent of planning and control. The bank initially recorded a fivefold increase in the first week, according to Segura. Sales then reached 12 times the initial level after further changes, including the addition of channels beyond WhatsApp.
A separate TI Inside interview with Segura reports the same sequence: five times after one week and 12 times in subsequent months. The claim came from named bank executive interviews, not unattributed marketing copy.
The denominator remains hidden. Neither account gives initial or subsequent loan counts, disbursed value, exact measurement dates, approval rates, or a control group. The later result also followed channel changes. The public evidence therefore cannot isolate how much of the increase came from the agent, more traffic, faster contact, different pricing, or wider distribution.
That distinction is not statistical fussiness. Crédito do Trabalhador lets a worker request and compare proposals from multiple institutions. A bank that replies in seconds can win business that a slower bank would have approved on identical terms. Faster completion can also prevent an accepted borrower from abandoning biometric verification. Both effects increase booked loans without changing the credit rule at all.
The fivefold and 12-fold figures are not a clean before-and-after pair. The first covers one week. The second covers unspecified later months after product adjustments and a channel expansion. A credible test would hold traffic sources and eligibility rules constant, then compare conversion and losses across matched periods.
Analysis
The system is more than a scripted lead form in Daycoval’s account. Segura told Folha that peak demand can reach 300,000 credit simulations in a day. He described a system that recognises identity data, calculates a credit score, applies Daycoval’s policy using factors including age, employer and income band, and generates an offer.
In TI Inside’s interview with Segura, the customer receives the proposal on WhatsApp and the agent can return a counteroffer when the customer rejects the rate, within parameters set by the bank. Segura said the system then coordinates digital signing, selfie verification, status updates and disbursement through Pix. He also described a second AI layer that monitors conversations and sends unresolved cases to a human. Daycoval has not disclosed how often that escalation happens.
The public descriptions do not separate the generative agent from the predictive score and the bank’s rules engine. TI Inside says Daycoval integrated Salesforce Agentforce with APIs supporting the customer journey. It does not show whether Agentforce calculates the score, calls an existing model, selects a policy-approved price, or merely presents the result. Those are different forms of autonomy with different model risks.
The software is doing more than collecting a phone number. The evidence still measures selling, not lending.
Brazil’s programme makes that distinction consequential. The Labour Ministry’s consolidated 2025 report records 15.1 million contracts worth about R$94.2 billion across 8.2 million workers. The total includes loans migrated from the previous system. It also says 53% of beneficiaries earned no more than four minimum salaries.
That is scale, not a performance benchmark. A fast sales system is reaching borrowers for whom a pricing or approval error can materially change monthly cash flow.
The national totals cannot fill Daycoval’s disclosure gap. They combine 99 active financial institutions and migrated legacy contracts. They reveal neither Daycoval’s share nor the performance of loans sourced by its agent. Using the R$94.2 billion programme total to validate one bank’s sales ratio would replace one missing denominator with another.
The legal ownership does not migrate to the software vendor. Portaria MTE 435 defines the loan as a contract between the borrower and the authorised lending institution. It requires the proposal to show the net amount, instalment, total repayment, interest rate and total effective cost. It also makes the institution responsible for acts performed in its name, including through a correspondent. Dataprev operates the payroll-registration infrastructure. Daycoval owns the credit decision and the customer outcome.
Implications
The missing comparison starts with a lending funnel. Daycoval would need to disclose requests, offers, approvals, acceptances and completed contracts for comparable periods. It would also need the offered and accepted interest rate and total effective cost. Without those fields, 12 times sales could reflect better response speed, lower prices, a looser approval threshold, or all three.
Credit quality arrives later. The useful measures are 30, 60 and 90-day delinquency, default, loss after payroll and FGTS recoveries, early repayment, cancellation and complaints by origination vintage. Daycoval has published none of those measures for the agent-originated cohort in the materials reviewed.
Fairness needs the same discipline. Approval, price, completion, human referral and arrears should be compared across relevant borrower groups. The disclosed use of age, employer and income band does not prove discriminatory treatment. It creates an audit question that aggregate sales cannot answer.
The strongest current claim is operational. Daycoval appears to have compressed a high-volume, low-ticket sales journey that human correspondents could not process at the same speed. That may be commercially valuable. It may even produce a better customer experience.
It is not yet evidence that the bank approves better risks, charges fairer prices, or treats similar applicants consistently. A 12-fold sales result belongs on a distribution dashboard. Underwriting needs a vintage table with losses by cohort.
Discussion
Sign in to join the discussion.
No comments yet. Be the first to share your thoughts.