Financial services have never been better at detecting fraud. Models are accurate, scores arrive in milliseconds, and AI investment keeps climbing. Yet losses climb with it. That is not a detection problem. It is a decision problem. As AI models arrive inside these complex, data-driven infrastructures, high-performance decision engines are being paired with fragmented, often stale data foundations, and the results show.
A fraud score computed after the authorization window closes is an audit record, not prevention. The model was right. The alert fired. The money moved anyway. Between detection and enforcement, most FS architectures have a gap: nothing turns a score into an enforceable, auditable decision while the transaction is still waiting.
What the numbers say
In testimony to the US Congress Joint Economic Committee, the FTC reported $15.9 billion in consumer fraud losses for 2025, up roughly 430% since 2020. Accounting for underreporting, the true annual cost may be closer to $196 billion. Nearly two thirds of victims send money within 24 hours of first contact, increasingly over rails where settlement is instant and irreversible.
And a late decision costs you twice. Alongside the fraud that gets through, false declines cost North American merchants and issuers an estimated $81 billion a year and climbing: legitimate customers turned away by decisions made on stale context. The loss is not only the transaction. It is often the customer.
The architecture problem
Three things must happen simultaneously for a real-time decision to hold: the current state must be read, the logic must run against it, and the outcome must be enforced. In most stacks those three live in three different systems: one holds the state, an ML service produces the score, and a message queue carries the block instruction.
Concurrent checks read the same velocity counters before any of them are updated. That race condition is exactly what fraudsters exploit with rapid parallel transactions. And when the pieces are separate, there is no single audit trail of the decision sequence: what was queried, how current it was, what was scored, and what was decided.
The regulatory clock is running
- UK APP reimbursement rules put prevention failure directly on bank P&L.
- SEPA Instant is mandatory; FedNow adoption has passed 1,400 institutions.
- ISO 20022 migration is complete on FedWire and CHIPS: richer data, same shrinking window.
- RBI now requires real-time screening across UPI, at 18B+ transactions a month.
Instant rails settle in seconds and do not reverse. Once funds release, they are unrecoverable. Every step, validating, assessing, approving or declining, settling, releasing, has to complete inside that sub-second window. Batch decisioning is now both a performance drag and a compliance risk, and the risk compounds as high-throughput AI models are bolted on top of it. The issue is not the AI. It is the data infrastructure underneath it.
The decisioning layer, defined
A decisioning layer holds an authoritative state, runs the decision logic against it, enforces the outcome, and records the full decision trail, all within one transaction, while the payment is still waiting. Not faster plumbing. A layer with decision authority.
It is also the foundation any AI ambition depends on: as our Field CTO Dheeraj Remella argues, most of what makes AI trustworthy in fraud prevention is the deterministic infrastructure underneath it, not the model itself.
Where to go next
For the business case, read Ken Ballou’s executive brief, Why False Declines Cost More Than Fraud in AI-Driven Payments. For the hands-on view, the Volt Dev Hub lets your architects see a real-time decision made in practice. Our CEO, David Flower, takes the FTC numbers head-on in a new piece publishing soon!
Originally Posted At : The Decisioning Layer in Financial Services: Why a Late Decision Is a Wrong Decision
Sign in to leave a comment.