The Asymmetry at the Centre of Payments
When an issuing bank evaluates a card transaction, it sees what it has always seen: the cardholder’s history with that bank, the merchant’s category, the transaction amount, and the behavioural baseline the bank has built from its own portfolio. That is a rich dataset. It is also an incomplete one — bounded by the walls of a single institution and blind to everything happening at the same merchant, the same device, and the same fraud ring across every other bank on the network.
The payment network sees all of it.
With an important caveat: the network’s visibility is entirely transactional. It holds no KYC data, has no direct relationship with the cardholder, and carries no identity or account information of its own. It does not know who the cardholder is — only what the card has done, across every institution on the network. That constraint is also the source of the network’s analytical power. Because its intelligence is derived purely from behavioural and transactional patterns rather than from any individual’s identity, it is the one participant in the payment chain that can legitimately aggregate and act on signals across all institutions simultaneously.
This is the defining structural asymmetry in financial services AI. The entity that processes every transaction across every participant — that sees the merchant whose fraud rate has spiked across six issuing banks in the past 72 hours, the device that has been used to test cards at forty different issuers, the money mule account receiving funds from APP scam victims at banks it has never interacted with directly — does not make the final authorisation decision. That decision belongs to the issuing bank, which holds the cardholder relationship, the KYC data, and the fraud liability, and which makes a call in milliseconds.
The conventional response to this asymmetry has been to close the authority gap — to get the network’s intelligence as close as possible to the bank’s decision through real-time scoring, risk signals embedded in authorisation flows, and model outputs that influence the approve/decline decision without formally making it. This is the right instinct partially executed. It addresses the transaction-level decision. It does not address the deeper question: what should a network do with cross-institutional visibility that no individual participant can replicate?
What the Network Sees That No Bank Can
The answer begins with being precise about what cross-institutional visibility actually produces — and why it is qualitatively different from better transaction scoring.
A fraud ring executing a card-not-present scheme does not concentrate its activity at a single issuer. It distributes transactions across dozens of banks, keeping volumes at each institution below the thresholds that trigger manual review. From each issuer’s perspective, the pattern is noise — a slightly elevated anomaly rate on a handful of accounts, well within normal variance. From the network’s perspective, the same accounts, devices, and merchant endpoints are lighting up simultaneously across the entire participant population. The ring is invisible at the bank level and obvious at the network level. No amount of investment in single-issuer AI closes this gap. It is a structural property of the data, not a modelling problem.
The same logic applies to APP scam detection. An authorised push payment scam works because the sending account and the receiving mule account are typically at different banks. The sending bank sees a customer making what looks like a legitimate payment; their fraud models, trained on their own transaction history, see no anomaly. The receiving bank sees an account receiving a payment — unremarkable in isolation. The network sees both sides of the transaction simultaneously, and it sees the mule account receiving funds from thirty different senders across twenty different issuing banks over the past 96 hours. That pattern does not exist in any single institution’s data. It only exists in the network.
Or consider false decline rates. An issuer operating with a false decline rate of 3% on e-commerce transactions believes it is managing fraud conservatively. It has no way of knowing that its false decline rate is 2.5 times the network average for its peer group, or that three comparable issuers have achieved a lower fraud loss rate with half the false decline rate, because the data needed to make that comparison does not exist within any single institution. The network has it. Most networks are not delivering it.
This is the distinction that matters strategically: the network’s advantage is not in building a better transaction score than the issuing bank. Any bank with sufficient data science capability can build a competitive transaction score on its own portfolio. The network’s irreplaceable advantage is in generating intelligence that requires cross-institutional data to exist at all.
The Authority Gap Is Not the Problem
The authority gap — the network provides intelligence, the bank makes the decision — is frequently described by network executives as the central constraint on their AI strategy. It is worth examining whether this framing is accurate or whether it is a choice dressed up as a constraint.
The issuing bank controls the approval decision because it carries the cardholder relationship, the KYC obligations, and the fraud and credit liability. That allocation is not arbitrary, and it is not changing. A network that pursued formal authorisation authority would face regulatory and contractual barriers, lose the neutrality that makes it acceptable to hundreds of competing issuers, and take on liability it is not structured to absorb. The authority gap is not a constraint to be overcome. It is the correct structural arrangement for an infrastructure provider operating across a competitive ecosystem.
The mechanism through which networks currently exercise influence within this structure is transaction enrichment. As an authorisation request travels from the merchant’s acquirer to the issuing bank, the network appends a real-time risk score — derived from its cross-institutional transaction intelligence — to the message. The issuer receives both its own models’ outputs and the network’s risk signal, and combines them in its approval decision. The network does not decide; it informs. This enrichment model is the right architecture, and it is where most network AI investment is currently concentrated.
Where the current model is incomplete is in the resolution of ambiguous cases. An issuer presented with a high-risk signal faces a binary choice — approve or decline — and the cost of a wrong decision in either direction is real. The network’s risk scoring infrastructure, used in conjunction with 3D Secure and step-up authentication protocols, creates a third path. Rather than forcing a decline on an uncertain transaction, the issuer can trigger a step-up authentication challenge — a one-time passcode, a biometric prompt, an out-of-band confirmation — that resolves the ambiguity by asking the cardholder to prove they are present. The legitimate cardholder authenticates and the transaction completes. The fraudster, lacking access to the authentication channel, fails. The result is lower fraud losses and lower false decline rates simultaneously — the precision improvement that matters most, achieved through a mechanism the network is positioned to orchestrate across every issuer on its rails.
What follows from accepting the enrichment-and-orchestration model is a different framing of the network’s AI role. The network is not a participant in the transaction decision. It is the provider of the intelligence layer that makes every participant’s decisions better. That is not a lesser role. It is a more scalable one — every improvement to the network’s cross-institutional intelligence compounds across every transaction made by every participant, rather than being contained within a single institution’s approval flow.
The most important AI question for a payment network is therefore not “how do we make our transaction score more accurate?” It is “what intelligence can we produce from cross-institutional data that participants cannot produce for themselves — and are we building the capability to generate and deliver it?”
Most networks, evaluated against that question, are underinvested in the second part.
Where the Intelligence Layer Creates Value
Three domains illustrate what the cross-institutional intelligence layer produces when it is built deliberately.
Cross-network fraud intelligence is the clearest case. The fraud patterns that are hardest for individual banks to detect — distributed card testing, mule account networks, coordinated merchant fraud — are the patterns most visible at network scale. A network that aggregates these signals across its participant population and delivers them as actionable intelligence — not just as a higher fraud score on an individual transaction, but as a characterisation of an active fraud campaign, the accounts and merchants involved, and the issuing banks being targeted — gives each participant a capability they cannot replicate internally. The fraud ring that is invisible to fifteen separate banks becomes actionable intelligence delivered to all fifteen simultaneously.
Participant performance benchmarking is the intelligence application most consistently underexploited relative to its value. Networks possess the data to tell every issuing bank exactly where their approval rates, false decline rates, and fraud loss rates stand relative to comparable institutions — by merchant category, transaction type, geography, and cardholder segment. This is not generic industry data. It is precise, current, and based on actual transaction outcomes across the network’s full participant population. An issuer that discovers its false decline rate on travel e-commerce is 3x the network peer median has a specific, actionable improvement opportunity it would never have identified from its own data. A network that delivers this intelligence systematically is providing something that creates direct economic value for participants — and that strengthens the participant relationship in ways that a fraud score improvement does not.
Systemic risk monitoring is the domain where cross-institutional visibility has consequences beyond individual participant economics. A single large acquirer whose merchant portfolio is deteriorating creates risk not just for itself but for every issuer whose cardholders transact with those merchants. Settlement exposure concentrations that look manageable at the individual participant level can represent systemic stress when viewed in aggregate across the network. The network is the only entity positioned to see these accumulations as they build — and the only entity that can act before they reach a scale where individual participant risk management, however sophisticated, cannot contain the contagion.
The False Decline Number Nobody Is Tracking
There is a specific number in payments that deserves more attention than it receives: the estimated cost of false declines across major card networks is $50–100 billion annually in lost legitimate sales — roughly two to three times the value of actual fraud losses the same models are preventing.
This ratio should change how networks think about the value of their AI investment. The current framing — “our fraud models prevent X billion in fraud losses” — captures roughly one third of the economic value at stake. The other two thirds, the legitimate sales that were declined, the customers who abandoned their cards, the merchants who lost revenue, does not appear in the fraud model’s performance metrics because it represents revenue that never existed in anyone’s ledger.
The network is the only entity that can measure this at full scale. An individual issuer can estimate its own false decline rate with effort. It cannot know how its false decline rate compares to peers, which merchant categories are generating the most false positive friction, or how much legitimate volume is being lost across the network as a whole. That visibility sits at the network level and almost nowhere else.
A network that measures total decision cost — fraud losses plus false decline costs — and delivers that measurement to participants as a basis for model calibration is doing something no single bank can do for itself. It is also making the argument for AI investment in precision improvement, rather than detection improvement, in terms that speak directly to the volume growth objectives that network leaders own.
The Recommendation
Define cross-institutional intelligence as a distinct strategic layer, separate from transaction scoring, and invest in it with the same seriousness as fraud model development.
In practice, this means three commitments.
First, build the signal aggregation infrastructure that makes cross-network patterns visible and actionable in near-real time — fraud campaign detection, mule account network identification, merchant risk signals that span multiple acquirers and issuers. These signals do not emerge automatically from transaction data. They require analytical investment in the graph structures, temporal patterns, and entity resolution that make cross-institutional patterns legible. That investment is what separates a network intelligence layer from a better version of the transaction score every bank already has.
Second, deliver participant performance benchmarking as a systematic intelligence product rather than an occasional data service. Every participant on the network should have continuous visibility into how their approval rates, false decline rates, and fraud outcomes compare to anonymised peer benchmarks — by segment, by merchant category, by transaction type. This benchmarking is only possible because the network sees across all participants. It creates direct economic value for participants. And it generates the data that makes the network’s AI value proposition concrete and measurable in terms participants can act on.
Third, measure the full decision cost, not just the fraud cost. Build the measurement infrastructure that tracks false decline rates across the network alongside fraud loss rates, and report both to participants. A network that helps its issuers understand that they are leaving $200 million in legitimate volume on the table through excessive false declines is not just a fraud prevention service. It is a growth partner. That is a different relationship, and a more durable one.
The banks will continue to make the final decision on every transaction. That is as it should be. The network’s job is to ensure that decision is made with the best possible intelligence — intelligence that only the entity seeing across all banks can produce. Most networks have not fully built that capability. The ones that do will hold a position that no individual participant, and no competing infrastructure provider, can replicate.
