Specialised AI That Stops All-Cause Fraud in Finance

All-cause fraud is eating financial services from the inside out. Not just card fraud. Not just account takeover. Not just scams. Everything, everywhere, all at once—often in the same customer journey.

That’s why John Filby, CEO of Outseer, made a point in a recent FinextraTV interview that I strongly agree with: specialisation matters. General-purpose fraud tooling struggles when fraudsters constantly switch channels, tactics, and identities. The practical answer isn’t “more alerts.” It’s building a fraud posture that assumes attacks will come from multiple directions—and stopping them with systems designed for that messy reality.

This post sits within our AI in Finance and FinTech series, where we look at what AI is actually doing inside banks and fintechs (beyond hype). Here, the theme is clear: AI-driven fraud detection has to move from product-by-product controls to all-cause fraud prevention platforms, backed by specialist models, integrated data, and bank–vendor partnerships that treat customer safety as a core outcome.

All-cause fraud prevention is the only model that matches reality

All-cause fraud prevention means one thing: you don’t fight fraud in silos. You treat scams, authorised push payment (APP) fraud, account takeover (ATO), mule activity, synthetic identity, and first-party misuse as connected problems.

Fraud teams learned the hard way that channel-based defences create gaps. A scam might begin with social engineering on a phone call, move to a password reset on the web, then end with a high-risk transfer through a mobile app. If your controls are separated by product, channel, or business line, you’re effectively giving criminals a map of where your blind spots are.

Why the “authorised” part is the hardest part

The rise in scams and authorised fraud is especially painful because the customer is often doing the action “correctly” from a pure authentication standpoint. The login is valid. The device is known. The payment is initiated by the user.

The signal isn’t “is this customer authenticated?” It’s “does this behaviour look like harm?” That’s a different question, and it forces banks to combine:

• Behavioural analytics (typing patterns, navigation flows, hesitation, copy/paste indicators)
• Transaction context (new payee, unusual amount, first-time corridor, time-of-day anomalies)
• Network intelligence (mule accounts, beneficiary risk, shared device/email/phone relationships)
• Customer history (typical patterns, past scam contact, known life events)

When Filby talks about “all-cause” platforms, this is the heart of it: a unified view of risk across the customer journey, not just at the payment screen.

The risk metric that matters: harm prevented per decision

Most organisations still measure fraud tooling with internal metrics like detection rate or false positives. Those matter, but they’re incomplete.

A better north star is: harm prevented per decision, with minimal customer friction.

If a control stops $1M in losses but triggers 50,000 unnecessary step-ups, your contact centre pays the price, your NPS drops, and customers start finding ways around security. All-cause platforms matter because they can make fewer, smarter interventions—only when the combined signals justify it.

Why AI specialisation beats “one model to rule them all”

The core argument behind specialisation is simple: fraud is not one problem. It’s a family of problems with different data, different adversaries, and different success measures.

A model that’s great at card-not-present fraud won’t automatically be great at detecting mule networks. A model trained to identify bot-driven credential stuffing won’t necessarily catch coercion scams where the customer is on the phone with a criminal.

Specialised AI = specialist models + shared orchestration

The sweet spot banks are moving toward looks like this:

1. Specialist models for specific fraud types and points in the journey
2. A shared decision layer (or orchestration engine) that combines model outputs with policy, business rules, and operational constraints
3. A consistent feedback loop from outcomes (confirmed fraud, scam typology, customer disputes, SAR/SMR indicators) back into model improvement

That combination gives you the best of both worlds: deep expertise at the model level, and coherent action at the platform level.

Using “all forms of AI” without creating a black box

Filby’s point about leveraging all forms of AI is practical if it’s done with discipline. In fraud stacks today, you’ll typically see:

• Supervised machine learning for classification (fraud/not fraud), tuned per channel
• Unsupervised learning for anomaly detection, especially useful for new typologies
• Graph AI / network analytics to identify mule rings and hidden relationships
• NLP to analyse scam narratives (from call notes, chat transcripts, complaint text)
• Generative AI for analyst support (case summarisation, investigation assistance, consistent customer communications)

Here’s the stance I’ll take: generative AI should not be the decision-maker for high-risk declines. Use it to speed investigation and standardise operations, yes. But the decision engine should remain a combination of robust models, well-governed features, and clear policies.

What banks and fintechs should demand from an AI fraud platform

If you’re evaluating an all-cause fraud prevention platform (or re-architecting your existing stack), the difference between “demo magic” and real protection comes down to operational reality.

1) A single view across channels (not just a single dashboard)

A unified dashboard is cosmetic. A unified view means:

• Shared identity resolution across devices, accounts, and sessions
• Consistent risk scoring that updates in near real time
• The ability to carry context from login → payee creation → payment execution

If your fraud teams still swivel-chair between tools to piece together a narrative, your platform isn’t all-cause. It’s just bundled.

2) Built-in scam detection, not only payment fraud controls

Scams often require intent detection: is the customer acting under manipulation?

Strong platforms support:

• Payee risk scoring and beneficiary intelligence
• Behavioural biometrics and journey analytics
• Step-up strategies tailored to scam scenarios (cooling-off periods, confirmation delays, payee warnings)
• Decisioning that considers customer vulnerability signals

And importantly: they support graduated responses, not only “approve/decline.”

3) Measurable outcomes, with an experimentation mindset

Fraud teams need to run controlled tests. If your platform can’t support champion/challenger strategies, it slows learning.

A practical measurement framework includes:

• Scam loss rate (losses per $1M transferred)
• False positive cost (customer friction + operational handling)
• Time-to-detect for emerging typologies
• Case handling time (minutes per investigation)
• Step-up acceptance vs abandonment (customer experience impact)

If the vendor can’t talk about experimentation and measurement, it’s a red flag.

4) Governance that won’t collapse under regulators and auditors

AI in finance needs governance. For fraud specifically, that means:

• Model explainability appropriate to the decision (reason codes, top drivers)
• Monitoring for drift (new scam typologies, seasonality spikes, new channels)
• Clear human-in-the-loop workflows for edge cases
• Documented controls for data lineage, feature use, and decision policies

In late 2025, regulators and boards are far less patient with “the model said so.” Build governance in from day one.

How to move from reactive fraud to predictive fraud prevention

Reactive fraud fights the last war. Predictive fraud prevention assumes criminals adapt—and builds systems that adapt faster.

Start with the journey map, not the product map

Banks often organise controls around products: cards, deposits, payments, digital banking. Criminals organise around people.

A better approach is to map:

1. Entry points (phishing, SIM swap, credential stuffing, remote access tools)
2. Account control (session hijack, device change, new payee)
3. Value extraction (fast payments, international transfers, cash-out via mules)
4. Cover tracks (limit changes, address changes, notification suppression)

Then place detection and friction where it’s most effective—often before the payment event.

Use graph intelligence to shrink the unknown unknowns

Scam and mule activity is highly networked. A single “clean” transaction can be part of a dirty network.

Graph-based approaches help you spot:

• Beneficiaries receiving funds from many unrelated victims
• Devices reused across “different” customers
• Phone/email/ID elements shared across synthetic identities
• Rapid formation of new account clusters after takedowns

This is where specialisation shines: graph risk isn’t a bolt-on feature; it’s its own discipline.

Design interventions that customers will actually follow

The best fraud decision is the one that prevents harm and keeps the customer on your side.

Effective scam interventions are usually:

• Timed (right at payee creation or first transfer, not after)
• Specific (plain-language reason: “This payee is strongly linked to known scam activity”)
• Actionable (offer a call-back, a short delay option, or a guided verification step)
• Respectful (avoid blame; scams thrive on shame)

If you treat scam victims as “careless,” you’ll lose them twice: once to the criminal, and again to churn.

Partnerships: why “purpose” is a practical strategy, not a slogan

Filby also emphasised anchoring to a higher purpose—partnering with banks to “make the world safer.” That can sound like marketing unless you translate it into operating behaviour.

Here’s the practical interpretation: fraud prevention only works when vendors and institutions share outcomes, not just deliverables.

In strong bank–fintech partnerships, you see:

• Joint typology reviews (what’s changing this month, what’s emerging next)
• Shared playbooks for scam spikes (holiday shopping season, tax-time, major data breaches)
• Faster model iteration cycles with controlled rollout
• Clear lines between automation and analyst review
• Mutual investment in customer education and safer UX

And yes—December matters. Scam patterns typically surge around holiday spending, delivery scams, and last-minute invoice fraud. If your platform can’t adapt to seasonal pressure quickly, it won’t protect customers when they’re most exposed.

A good fraud platform reduces losses. A great one reduces harm without making customers feel punished for being targeted.

Practical next steps for 2026 planning

If you’re a bank or fintech leader mapping 2026 priorities, here’s what I’d do next week:

1. Run an “all-cause” gap assessment: pick 20 confirmed scam/fraud cases and trace the full journey. Where did signals exist but weren’t connected?
2. Unify identity and session telemetry across channels: device, behaviour, login, payee, payment.
3. Add scam-specific controls: beneficiary intelligence, payee risk, behavioural signals, graduated interventions.
4. Stand up measurement that includes customer friction: quantify step-ups, abandonment, complaints, and contact centre cost.
5. Formalise vendor partnership rhythms: monthly typology reviews, quarterly model governance, and rapid response procedures for spikes.

Fraud isn’t slowing down. The organisations that win won’t be the ones with the most tools—they’ll be the ones with specialised AI working as one system, aligned with a clear purpose: keeping customers safe while keeping banking usable.

If you’re building or buying an all-cause fraud prevention platform, ask yourself one hard question: when scams shift tactics the week after you deploy, can your controls adapt fast enough to stay relevant?