On September 1, 2026, Visa announced an enhanced version of A2A Protect, its fraud-detection solution for account-to-account payments. The update matters for two reasons: Visa is introducing a new unified real-time fraud score, and for the first time it is bringing Featurespace technology into a commercial Visa offering following the acquisition completed in December 2024.
The objective is straightforward to describe, but difficult to achieve: give financial institutions enough risk intelligence to identify suspicious payments before money leaves a customer’s account.
The enhanced solution combines artificial intelligence, transfer learning, continuous model improvement and, for institutions that opt in, signals drawn from a broader ecosystem.
Visa announcement — September 1, 2026
What exactly is A2A Protect?
A2A stands for Account-to-Account: a payment or transfer sent directly from one account to another without necessarily relying on a card network.
A2A Protect is therefore not a new payment rail or an instant-payment system. It is a risk-detection layer that financial institutions can integrate into their existing fraud controls.
Visa says the solution can be connected through a single API.
When a transaction is assessed, A2A Protect returns a unified real-time risk score. Each alert also includes a plain-language explanation of why the transaction was flagged, giving fraud teams clearer information to support their decision-making.
The solution enriches the fraud decision rather than replacing the financial institution’s own policies and controls.
What Featurespace brings to the new version
Visa completed its acquisition of Featurespace on December 19, 2024. The company specialises in real-time artificial intelligence and behavioural analytics for fraud and financial-crime detection.
Visa describes the September 2026 release as the first in-market offering to combine Featurespace technology directly with Visa capabilities.
One of the key elements is transfer learning.
Traditional fraud models often need a substantial amount of institution-specific history before they can establish reliable patterns of normal and abnormal behaviour.
Transfer learning allows knowledge developed by a model in one context to be reused as a starting point in another. Visa says this enables A2A Protect to provide useful risk insights from day one rather than waiting months for sufficient local transaction history.
That does not mean the model immediately understands every local behaviour. Visa also stresses that detection performance improves continuously as the model learns within the institution’s own environment.
In practical terms, the approach combines two stages: start with useful existing intelligence, then keep learning locally.
Visa — completion of the Featurespace acquisition
A broader risk score — when institutions choose to opt in
The second important development concerns the breadth of information available to the risk model.
An institution can use A2A Protect without waiting for other banks to join a consortium. However, institutions that opt into network-level intelligence sharing can incorporate additional signals from across the network or ecosystem.
Visa says this broader view can help identify emerging scam hotspots and coordinated fraud activity that may be difficult for a single institution to detect from its own transaction data.
This matters when multiple payments that appear relatively ordinary in isolation are actually part of the same fraud campaign.
One bank may see an unusual payment to a new beneficiary. Another may observe several transfers to a different account. Individually, the signals may not be strong enough to trigger action. But if those beneficiary accounts are connected to the same fraud infrastructure or to similar money-movement patterns, a broader view can reveal relationships that are not visible locally.
The value is therefore not only in calculating a score faster. It is also about giving that score more context.
Why instant payments make fraud detection harder
Speed fundamentally changes the fraud-control problem.
When a transfer is executed and the recipient is credited within seconds, the time available to detect an anomaly, apply additional controls or attempt to stop the movement of funds becomes much shorter.
Visa places particular emphasis on Authorized Push Payment — APP — scams. In these cases, the customer is often properly authenticated and initiates the payment personally, but does so because a fraudster has manipulated them through a fake supplier, an urgent family request, impersonation, an investment scam or another social-engineering technique.
Fraud controls therefore need to answer more than one question: “Is this really our customer?”
They also need to consider: “Is this genuine customer being tricked into sending money to a fraudulent destination?”
That is one of the reasons Visa is combining behavioural analytics, continuous learning and broader beneficiary or ecosystem signals.
What Visa says the solution can deliver
Visa has published several performance figures for A2A Protect.
The company says the solution has been shown to increase fraud detection by 75% during the first six months of deployment. Visa also says it can help reduce unnecessary fraud alerts by more than 40%.
On its product page, Visa also cites its work with Pay.UK, where incorporating Visa data reportedly identified an additional 54% of fraud and APP-scam value beyond what the participating banks’ own systems had already detected, while reducing false positives by 40%.
These figures are relevant, but they need to be presented with the right qualification: they are results reported by Visa, not an independent assessment or a guarantee that every institution and market will achieve the same outcome.
Actual performance will depend on the available data, each institution’s decision rules, the fraud patterns present in the market and how the solution is integrated into existing controls.
What this can change for a fraud team
For a bank or payment provider, the operational proposition can be summarised in four areas.
First, reduce the amount of time needed for a model to become useful through transfer learning.
Second, provide a single real-time score that can be incorporated into the fraud decision process.
Third, make alerts easier to investigate through plain-language explanations.
And fourth, when the institution chooses to use network intelligence, enrich its local view with signals that may expose coordinated patterns.
The operational challenge remains unchanged: improve fraud detection without creating an excessive number of false positives that slow legitimate payments or overwhelm fraud teams.
In Africa, some building blocks for shared fraud intelligence already exist
Visa presents the enhanced A2A Protect as a global solution, but its announcement does not provide a detailed country-by-country list of current availability.
It would therefore be premature to describe the new version as already deployed at scale across Africa.
The broader idea of centralising or sharing selected fraud signals, however, is not new on the continent.
🇿🇦 In South Africa, SABRIC — the South African Banking Risk Information Centre works with banks and other stakeholders on information sharing, threat intelligence and coordinated responses to financial crime. SABRIC describes its role as strengthening the collective defence of the financial system through better ecosystem-wide intelligence.
🇳🇬 In Nigeria, NIBSS operates a Central Fraud Management System — CFMS and an anti-fraud reporting portal. NIBSS material also describes fraud management as a core industry service, supported by a central fraud-monitoring and mitigation platform.
NIBSS — Central Fraud Management System
NIBSS — Anti-Fraud Reporting Portal
AfricaNenda’s SIIPS 2025 report also recommends that instant-payment operators prioritise shared fraud-prevention infrastructure and intelligence platforms capable of serving the broader network.
AfricaNenda — State of Inclusive Instant Payment Systems in Africa 2025
These initiatives should not be presented as technical equivalents of A2A Protect. They simply show that part of the problem Visa is addressing — obtaining better fraud visibility beyond a single transaction or institution — is already being tackled in different ways in African markets.
Technology alone does not solve the information-sharing problem
Using signals from multiple institutions immediately raises governance questions.
Which data points can be shared? Who is allowed to flag an account or beneficiary? How long should a signal remain valid? How can institutions prevent a false positive from spreading across the ecosystem? Which information should remain local, and which elements can be pseudonymised or transformed into risk indicators?
These issues become particularly sensitive in the case of mule accounts, which are used to receive or redistribute fraud proceeds.
An effective shared framework cannot simply become a common blacklist. It also needs rules for creating, reviewing, challenging, updating and removing fraud signals.
Models will still need to prove themselves locally
Transfer learning helps solve the cold-start problem, but it does not remove the need for local learning.
Payment behaviour varies significantly from one market to another. Across Africa, the combination of banks, mobile-money operators, fintechs, agent networks and different account types creates transaction patterns that may look very different from those seen elsewhere.
A global model can provide a useful starting point. It still needs to be measured against local outcomes: detected fraud, missed fraud, false positives, performance by channel and changes over time.
That may become one of the most important tests for A2A Protect as it enters new environments.
A first concrete look at what Visa plans to do with Featurespace
The significance of this announcement is as much strategic as it is technical.
Less than two years after completing the Featurespace acquisition, Visa is beginning to embed the company’s technology into its own risk products. A2A Protect is therefore an early concrete example of how Visa intends to combine its payments reach and risk data with Featurespace’s artificial-intelligence capabilities.
The product is trying to address three very current constraints: make a risk decision in real time, make the model useful quickly and improve visibility when fraud extends beyond a single institution’s view.
For African payment providers, the point is not to replicate Visa’s model automatically. It is to observe what the approach actually delivers, how it performs in different environments and which forms of shared intelligence could strengthen controls already in place.
In that respect, the next real-world deployments of A2A Protect may be more informative than the launch metrics themselves.
