Instant payment schemes, such as Faster Payments, SEPA Instant Credit Transfer, and real-time A2A rails, have changed the economics of payment fraud. Once a payment settles, recovery is unlikely, which means fraud controls must operate before settlement, not after.
Many traditional fraud platforms were designed for cards, chargebacks, and post-event investigation. They assume time for review, reversals, or manual intervention. Instant payments remove that buffer. Decisions must be made in milliseconds, often under scheme-level SLA and regulatory constraints. This has led to the emergence of AI-driven payment fraud solutions that can operate inside, or directly adjacent to, the transaction processing flow.
These systems are not just scoring engines. They are designed to return enforceable decisions (allow, block, or step-up) within the execution window of a payment.
This article examines how four vendors, Vyntra, Feedzai, DataVisor, and Hawk, approach in-flow (or near in-flow) AI fraud decisioning, and what you should bear in mind when choosing between them.
In this article
What does “in-flow” fraud detection mean?
In-flow fraud detection refers to the process of identifying and flagging fraudulent activity in real time as transactions or events occur, rather than analyzing them after the fact. It can describe very different architectures. At one end of the spectrum, fraud logic is a hard gate:
- A payment cannot proceed unless a fraud decision is returned
- If the fraud system fails, the payment flow is explicitly handled (fail-open or fail-closed)
- Latency is deterministic and isolated from other workloads
At the other end of the spectrum, fraud systems operate in an advisory capacity:
- A score or recommendation is returned quickly
- Another system (payment engine, orchestration layer) decides whether to act
- Enforcement depends on integration quality, not the fraud platform alone
Why transaction observability matters for in-flow fraud
In instant payments, fraud decisions are constrained not only by risk but also by time. Scheme SLAs apply end-to-end, across multiple internal systems, rails, and handoffs. A fraud engine that does not understand where a payment is in its execution state may correctly identify risk, but too late to act.
Effective in-flow fraud detection depends on real-time transaction observability. This includes visibility into timing, sequencing, and execution state, not just the final payment message. Latency anomalies, unexpected sequencing, or deviations from normal volume and value patterns can signal both fraud risk and whether a payment is still interceptable before settlement.
Without this context, fraud systems are forced to make binary decisions in isolation. With it, they can determine not only whether a payment is risky, but also whether enforcement is still possible within the scheme’s constraints. In instant payments, that distinction is critical.
Vyntra: Fraud enforcement embedded in the payment execution path
Vyntra’s approach to fraud prevention starts from the premise that instant payments are irrevocable, so fraud decisions must be enforced by design and made with full awareness of the payment’s execution state.
Vyntra operates inside the transaction processing flow:
- Payments do not progress unless a fraud decision is returned
- There are no asynchronous callbacks or “block later” patterns
- Instant payment execution paths are isolated from batch analytics and AML workloads to preserve deterministic latency
Fraud decisions are informed by more than the final payment instruction. The platform has real-time visibility into how a payment is behaving as it moves through the execution path, including:
- Behavioral signals from initiation and authentication
- Sequencing and timing context across the payment lifecycle
- Transaction observability data is typically owned by payments and operations teams
This allows the system to assess whether a payment is risky and whether it can be intercepted within scheme-level SLAs. Latency anomalies, unexpected sequencing, or deviations from normal processing patterns can therefore inform enforcement decisions before settlement occurs. Explainability is generated at decision time, using the same execution and behavioral context that drove enforcement.
Feedzai: Real-time fraud decisions via embedded risk services
Feedzai is widely deployed across cards, account-to-account payments, and digital banking channels. Its platform uses machine learning models trained on behavioral, transactional, and device-level signals, with a strong focus on explainability and governance.
In most implementations, Feedzai operates as a real-time decision service:
- Transactions are evaluated via APIs during payment processing.
- The platform returns allow, block, or step-up recommendations.
- Financial institutions configure how those decisions are enforced.
Feedzai can meet low-latency requirements, including for instant payments. However, whether it is truly “in-flow” depends on architecture:
- Enforcement is usually handled by the financial institution’s orchestration or payment engine.
- Determinism depends on how tightly Feedzai is coupled to execution.
For instant payment use cases, Feedzai can operate close to the flow, but enforcement guarantees are implementation-dependent, not intrinsic.
DataVisor: Real-time scoring with flexible orchestration
DataVisor positions itself as a broad fraud and financial crime platform combining:
- Supervised and unsupervised machine learning
- Rules and decision logic
- Graph-based and network analytics
Its Decision Flow tooling allows teams to:
- Route transactions dynamically
- Apply multiple models in sequence
- Invoke third-party data sources in real time
DataVisor can support low-latency environments, but it is typically adjacent to the transaction flow rather than embedded within it:
- Decisions are returned quickly
- Downstream systems are responsible for enforcement
This architecture offers flexibility, including the ability to experiment and run A/B tests. The trade-off is operational complexity:
- More internal orchestration to manage
- Greater risk of latency creep
- Less deterministic enforcement under failure conditions
For instant payments, financial institutions need to ensure DataVisor’s decisioning occurs early enough to comply with scheme rules, without relying on post-processing.
Hawk: API-first, pre-settlement fraud interdiction
Hawk positions itself as a real-time, explainable AI platform for transaction fraud and payment screening. It is designed to be rail-agnostic and API-driven.
In most deployments, Hawk functions as a pre-settlement interdiction layer:
- Transactions are assessed before execution completes
- Responses are returned within tight latency budgets
- Financial institutions can block, hold, or release payments based on policy
The platform emphasises:
- Explainable AI outputs suitable for customer and regulator queries
- Self-service rule management
- Rapid model iteration and deployment
As with Feedzai, Hawk’s ability to operate “in-flow” depends on integration depth:
- Enforcement logic usually sits with the financial institution
- Fallback behavior must be carefully designed
How should financial institutions evaluate in-flow AI fraud solutions?
You can pressure-test prospective vendors with the following questions:
1. Is enforcement mandatory or optional?
Does the payment physically stop if no decision is returned, or can it proceed without one?
2. Are latency guarantees deterministic?
Ask for p95 and p99 response times under peak load. Confirm whether instant payment scoring is isolated from other workloads.
3. What happens on failure?
Clarify fail-open versus fail-closed behavior, fallback rules, and incident handling during degradation.
4. How much context feeds the model?
Does the system see only the final payment, or behavioral and sequencing data from across the lifecycle?
5. Is explainability available immediately?
Can you explain a block in real time to a customer or regulator, without post-event reconstruction?
6. Where does operational complexity sit?
How much orchestration, tuning, and incident management do you own, and what does the platform handle natively?
How to choose the right in-flow fraud prevention solution
In-flow AI fraud prevention is not a single product category. It is a set of architectural choices about where decisions are made, how they are enforced, and what happens when things go wrong.
Some platforms are designed to act as execution gates by default. Others provide high-quality real-time intelligence that can be placed close to the flow with sufficient engineering effort.
As instant payments continue to grow, financial institutions must be clear about what they are buying (advice at speed, or enforcement by design) and choose accordingly based on their risk appetite, regulatory exposure, payment volumes, and operational maturity.
Why enforcement, not scoring, defines in-flow fraud
As instant payments continue to grow, financial institutions must be clear about what they are buying (e.g., risk advice at speed, or enforcement by design) and choose accordingly based on their risk appetite, regulatory exposure, payment volumes, and operational maturity.
For instant payments, fraud prevention only works if decisions are enforced before settlement. The real distinction between “in-flow” solutions is whether a payment is structurally prevented from proceeding without a fraud decision.
Some platforms are designed as execution gates (no decision, no payment). Others deliver fast, high-quality risk signals but depend on downstream systems to enforce outcomes, manage latency, and handle degradation.
In practice, this determines where risk and operational responsibility reside. Architectures that make enforcement intrinsic reduce reliance on orchestration logic, minimise ambiguity during failure, and provide clearer accountability at the point of execution.
Sources:
https://www.datavisor.com/intelligence-center/product-sheets/decision-flow
https://www.datavisor.com/products/fraud-platform
https://www.datavisor.com/intelligence-center/product-sheets/fraud-and-risk-platform
https://www.datavisor.com/news/datavisors-innovation-in-advanced-decision-flow-capabilities
https://www.feedzai.com/solutions/transaction-fraud
https://www.feedzai.com/fraud/
https://www.feedzai.com/resource/inbound-payment-fraud-detection-and-mule-risk-modelling
https://www.feedzai.com/solutions/risk-management-for-acquirers-and-payment-service-providers/
https://hawk.ai/our-products/transaction-fraud
https://hawk.ai/our-products/payment-screening
