Digital fraud in payment methods: how to combine AI, biometrics, and behavioral analysis.

The same technology that made attacks cheaper is now the only one capable of sustaining defense. This understanding has shifted from a prediction to an operational consensus in the Brazilian financial sector.

The numbers explain the urgency. A Microsoft study with financial executives indicates that 93% believe Artificial Intelligence (AI) has made fraud more sophisticated, 90% consider the technology the next major exploitable vulnerability in the financial system, and 84% state that their institutions have already suffered attacks involving agentic AI, a percentage higher than the global average. Brazil records the highest rate of exposure to social engineering scams with deepfakes among the surveyed Latin American countries, at 64%. In response, 93% of organizations are already using or planning to implement AI-based fraud controls.

The reflection has shifted from whether AI should be included in the defense to how to make it provide true security.

Why has fraud migrated from transactions to behavior?

Because attacking the person has become cheaper than attacking the infrastructure. Tokenization, strong authentication, biometrics, and device recognition have closed off many technical vectors. But crime has not given up; it has just changed targets.

A survey by Quod recorded over 9 million indications of fraud in the first half of 2026, a 10.26% increase over the previous semester, with social engineering accounting for about 3.6 million occurrences, equivalent to 40% of the total. Mobile phones appear in 78% of cases, and Pix is the method of movement in 85%. There were 3.1 million victims, of whom about 799 thousand fell for scams more than once.

Why do traditional controls not see?

When the scam is through social engineering, the transaction reaches the antifraud engine with the correct device, correct credentials, correct geolocation, and successful authentication. Nothing seems abnormal because the authorization came from the account holder. Transactional models respond to the question, “Is this operation anomalous?” In the case of manipulation fraud, the answer is no.

What does AI start to see?

Artificial Intelligence changes the level of defense because these signals are weak, numerous, and only make sense in conjunction – precisely the type of pattern that fixed rules do not capture and models capture well:

• Behavioral biometrics: typing rhythm, pressure, touch trajectory, and how the device is held. They differ between the account holder and someone acting under external instructions.
• Session analysis: atypical hesitation, switching between the app and an active call, reading time much higher than usual on familiar screens.
• Relationship graphs: destination accounts concentrating receipts from unrelated victims reveal networks of “strawmen” that transaction-by-transaction analysis never exposes.
• Language models applied to scam narratives: recognizing patterns of urgency, authority, and fear in service channels and contest processes.

None of these signals exist in the transaction record. All exist in the session and the relationship. Those who only instrumented the transaction did not collect them, and no model compensates for data that was not captured.

How to maintain an effective model when the fraudster learns alongside?

By recognizing that this is not a common statistical drift problem, but rather adversarial adaptation. The distinction is operational. In traditional drift, the distribution changes for reasons unrelated to the model. In fraud, it changes because the model has gone live: the adversary observes what passes, infers the rule, and circumvents it. Every antifraud model starts to degrade the day it goes live, and the speed of that degradation is defined by who is on the other side.

Four practices sustain effectiveness over time.

Shorten the retraining cycle

Quarterly retraining in a vector that changes in weeks is ineffective. The cadence should be defined by the observed degradation, not by the data team’s calendar.

Preserve unbiased labels

This is the most neglected point. The model only receives outcomes from the operations it approved; it never knows if those it blocked were fraudulent. Without a small deliberately approved control group, with assumed risk and measured, the model learns only from its previous decisions and confirms its own errors.

Separate what changes quickly from what changes slowly

Deterministic rules absorb the new vector in hours. Models capture structural patterns and resist noise. Solving everything with a model takes too long to react; solving everything with a rule accumulates exceptions that nobody audits.

Do not teach the adversary

Detailed refusal messages and easily inferable limits by trial and error hand the fraudster a map. Explainability is an internal and regulatory obligation, not screen content.

False positive or loss from fraud: which error costs more?

Both are costly. What differentiates them is visibility: the loss from fraud appears in the results, has an owner, and generates a committee. The false positive does not show up anywhere because the legitimate customer is refused, does not complain, and leaves.

The Scamscope report from ACI Worldwide with GlobalData indicates that one in four scam victims say they will change financial institutions and projects losses of up to around R$ 11 billion from Pix scams in Brazil by 2028. If the experience of being a victim produces this effect on retention, being treated as a fraudster without having done anything tends to produce something similar.

The real gain of AI is not in the threshold

Here lies the point that is often lost in executive discussion. Adjusting the cutoff point merely exchanges one error for another: tightening the threshold decreases fraud but increases false positives; loosening it does the opposite. It’s a zero-sum game.

What well-applied AI does is different: it increases the power of discrimination and reduces both errors simultaneously. A model that sees behavioral and network signals separates legitimate customers from fraudsters better than one that only sees transactions, allowing for more approvals and fewer losses at the same time.

This is what differentiates a security-delivering AI project from one that merely automates the previous decision.

The three numbers that decide calibration

Once the model is defined, the cutoff point still requires the average value of the legitimate customer lost, average loss from approved fraud, and average cost of a manual review. Without them, the threshold is simply opinion.

And none of these values is constant. Refusing a customer on the first purchase costs differently than refusing someone who has been with the system for eight years. A single threshold applied across the entire operation is, in itself, an error—and it is the most common configuration in production.

What data needs to exist before any model can operate?

Most failures in antifraud projects are not due to modeling, but to data. Five conditions precede any discussion about algorithms; see below what they are: 

Reliable label

Confirmed fraud, suspected fraud, assumed loss, and unfounded contestation are different things and often occupy the same field. A model trained with dirty labels learns the noise of the operation, not the behavior of the fraudster.

Temporal correction of variables

Each attribute must reflect what was known at the moment of decision, not the current state of the base. Temporal leakage is the most frequent and treacherous defect: it generates excellent performance in validation and mediocre performance in production.

Event granularity

Aggregates serve reporting. Behavioral detection requires raw events, with timestamps, session, device, and preserved sequence—behavioral signals live in the order of occurrences, and aggregation destroys order.

Declared latency budget

A variable that takes three seconds does not exist in a decision of three hundred milliseconds. Real-time viability is a design criterion of the attribute, not discovered during implementation.

Closed feedback loop

Contestations, chargebacks, MED triggers, irregularity terminations, and decision reversals need to return as labels, automatically and with known timeframes.

The endpoint

Defense has shifted from anomaly detection to context reading. This changes what is collected, the speed at which the model learns, and which numbers need to be in place before any threshold decision.

In one of Brazil’s largest e-commerce and fintech operations, with peaks exceeding a thousand transactions per minute, market solutions covered between 85% and 90% of operations, a limitation stemming from dependence on generic databases common across the sector. Objective built a model trained on exclusive operational data, combined with the tool already in use, and the chargeback rate dropped from 2.8% to 0.3%, without negative impact on sales, which is exactly the effect of increasing discrimination rather than tightening the threshold. Read the full case.

Objective operates in building decision systems in regulated high-volume environments, integrating data engineering, applied AI, and governance. Talk to a specialist.

Frequently asked questions on the topic