Blocking fraudulent transactions is an obvious measure of fraud-prevention performance. Blocking legitimate customers is another matter. As explored in our earlier AI Fraud Detection Software: What Fraud Teams Should Compare buyer’s guide, effective fraud detection increasingly combines rules, machine learning, contextual intelligence and human expertise to identify suspicious activity. Once those systems are operational, however, fraud teams face a difficult balancing act: How do you stop more fraud without creating unnecessary friction for genuine customers?
An overly permissive system creates losses. An overly cautious one can decline good transactions, frustrate customers and potentially send them elsewhere. Effective AI fraud prevention therefore isn’t simply about detecting more risk. It is about making better decisions about risk.
Understand the Cost of False Positives
A false positive occurs when legitimate activity is incorrectly identified as suspicious or fraudulent.
The immediate consequence might be:
Transaction declined.
But the wider impact can include:
- Lost revenue
- Abandoned purchases
- Customer frustration
- Additional verification
- Contact centre enquiries
- Manual-review costs
- Reduced customer trust
This means fraud teams should consider two forms of loss:
Fraud allowed through
and
Good business incorrectly blocked
Fraud Prevention Principle
A fraud system shouldn’t be judged only by how much bad activity it stops. It should also be judged by how much good activity it lets through.
Move Beyond Simple Rules
Rules remain useful fraud controls.
For example:
Transaction value above £X → review
Five failed login attempts → block
New device + unusual location → challenge
The problem arises when individual signals are treated as proof of fraud.
A legitimate customer might make an unusually expensive purchase while travelling overseas from a new phone.
Each characteristic could appear suspicious.
Together with wider context, however, the transaction may still be legitimate.
AI and machine-learning systems can help organisations evaluate combinations of signals and patterns rather than relying entirely on isolated rules.
Use Risk Scores Rather Than Binary Decisions
Not every transaction needs to be classified immediately as simply:
Fraud | Not fraud
A more nuanced approach can assign different levels of risk.
For example:
Low risk → approve
Medium risk → additional verification
High risk → review or decline
This creates space between approving everything and blocking anything unusual.
Providers including SEON, Fingerprint, Mitek and Mastercard offer different forms of fraud, identity, device and risk intelligence that can contribute to these decisions.
Add More Context to Decisions
The more relevant context available, the easier it becomes to distinguish unusual behaviour from genuinely suspicious behaviour.
Signals can include:
- Device
- IP address
- Location
- Account history
- Transaction history
- Email characteristics
- Phone information
- Behavioural patterns
- Velocity
- Identity information
Imagine two customers attempting a £1,000 transaction.
For Customer A, it is their first interaction with the business from an unfamiliar device.
For Customer B, it is consistent with several years of purchasing history.
Transaction value alone says £1,000. Context says considerably more.
Use Device Intelligence
Devices can provide particularly valuable fraud context.
Fingerprint, for example, provides device intelligence designed to help organisations identify visitors and detect suspicious activity across web and mobile environments.
Fingerprint – https://fingerprint.com/
Device signals can help identify patterns such as:
- Multiple accounts linked to one device
- Rapid account switching
- Repeat fraudulent activity
- Unusual device configurations
- Attempts to disguise identity
But device information should still form part of a wider risk assessment rather than automatically determining the outcome.
Recognise Returning Good Customers
Fraud systems naturally focus on detecting suspicious behaviour.
They should also become better at recognising trusted behaviour.
A customer with a long history of legitimate transactions may present a very different risk profile from a newly created account displaying similar activity.
Historical behaviour can therefore help organisations distinguish:
unusual for everybody
from
unusual for this particular customer.
That distinction can reduce unnecessary friction for established customers while maintaining scrutiny where risk is genuinely elevated.
Apply Friction According to Risk
Security does not always have to mean stopping a transaction.
For activity that is suspicious but not clearly fraudulent, organisations can introduce step-up verification.
This might involve:
- Additional authentication
- Identity verification
- One-time passcodes
- Biometric checks
- Manual review
Mitek, for example, provides identity-verification technologies incorporating document and biometric verification alongside fraud-prevention capabilities.
Mitek – https://www.miteksystems.com/
The principle is straightforward: Low risk should encounter less friction. Higher risk should justify more assurance.
This allows fraud teams to concentrate stronger controls where they are actually needed.
Segment Fraud Strategies
One fraud policy may not suit every customer, product or transaction.
Risk can differ according to:
- Customer type
- Geography
- Product
- Payment method
- Transaction value
- Channel
- Account age
Fraud teams can therefore develop strategies around meaningful segments rather than applying identical thresholds everywhere.
A high-value transaction from a new customer may warrant different controls from a low-value repeat purchase by a long-established customer.
Segmentation helps organisations become more precise rather than simply more restrictive.
Combine Machine Learning with Rules
AI does not necessarily replace fraud rules.
The two can work together.
Rules can remain valuable where organisations have clear policies or known fraud patterns, while machine learning can identify relationships and behaviours that would be difficult to encode manually.
SEON, for example, combines digital footprint and device intelligence with fraud-detection capabilities designed to help organisations evaluate transaction and customer risk.
SEON – https://seon.io/
The strongest approach may therefore be:
rules + models + contextual intelligence + human judgement
rather than treating any one element as the complete solution.
Feed Outcomes Back into the System
Fraud models need useful feedback.
Organisations should capture what happened after a decision.
For example:
Transaction approved → legitimate
Transaction approved → later confirmed fraud
Transaction reviewed → legitimate
Transaction declined → subsequently verified as genuine
These outcomes can help fraud teams understand where controls are performing well and where adjustments may be required.
AI Fraud Insight
A fraud decision is not only an outcome. It is also data that can improve the next decision.
Poor feedback, however, can reinforce poor decisions.
Data quality and governance therefore remain essential.
Keep Human Review Focused
Manual review is valuable but expensive and difficult to scale.
The objective should not necessarily be to remove people from fraud decisions entirely.
Instead, automation can help direct human expertise towards the cases where it adds most value. For example:
Clear low risk → automated approval
Clear high risk → automated action where appropriate
Ambiguous risk → human review
This can reduce review volumes while allowing analysts to concentrate on genuinely difficult cases.
Monitor Model Performance Over Time
Fraud patterns change.
Customer behaviour changes.
Businesses launch new products, enter new markets and introduce new payment methods.
A model that performs well today may therefore behave differently in future.
Fraud teams should monitor:
- False-positive rate
- Fraud losses
- Approval rate
- Manual-review rate
- Challenge rate
- Customer abandonment
- Model drift
- Performance by customer segment
Unexpected changes can indicate that thresholds, rules or models need investigation.
Don’t Optimise Fraud Loss in Isolation
Imagine a new fraud control reduces losses by £50,000.
That sounds successful.
But what if it also incorrectly blocks £200,000 of legitimate sales?
Fraud teams therefore need to evaluate performance alongside commercial outcomes.
Mastercard, for example, operates fraud and decisioning technologies across the payments ecosystem, where assessing transaction risk while supporting legitimate approvals is central to the challenge.
Mastercard – https://www.mastercard.co.uk/
Measurement Principle
The objective isn’t the lowest possible fraud rate. It is the best possible balance between fraud prevention, customer experience and legitimate revenue.
A Practical False-Positive Reduction Framework
Fraud teams can work through a simple cycle:
1. Measure
Establish current fraud, approval and false-positive rates.
2. Analyse
Identify where legitimate customers are being blocked.
3. Add context
Determine whether additional device, behavioural or identity signals improve decisions.
4. Segment
Apply controls according to meaningful differences in risk.
5. Step up
Use additional verification for uncertain cases rather than automatically declining them.
6. Review
Direct human expertise towards genuinely ambiguous activity.
7. Learn
Feed confirmed outcomes back into rules and models.
8. Monitor
Watch for changing fraud and customer behaviour.
Frequently Asked Questions
What is a false positive in fraud detection?
A false positive occurs when legitimate customer activity is incorrectly identified as fraudulent or sufficiently suspicious to trigger unnecessary intervention.
Can AI reduce fraud false positives?
AI and machine learning can help by analysing combinations of behavioural, transactional, device and identity signals, potentially enabling more nuanced risk decisions than simple rules alone.
Should businesses remove fraud rules when adopting AI?
Not necessarily. Rules can remain useful for known patterns and business policies, while machine learning can add more adaptive risk analysis.
How should fraud teams measure performance?
Useful measures include fraud losses, approval rates, false-positive rates, manual reviews, verification challenges and customer abandonment. These should be considered together rather than optimising one metric in isolation.
Product Guide
Fingerprint
Provides device-intelligence technology helping digital businesses identify visitors and detect suspicious activity, including account takeover, payment fraud and other online abuse.
Website: https://fingerprint.com/
Mastercard
Provides payment-network, identity, cybersecurity and fraud-prevention technologies supporting risk assessment and transaction decisioning across the global payments ecosystem.
Website: https://www.mastercard.co.uk/
Mitek
Identity-verification and fraud-prevention specialist providing technologies including identity-document verification and biometric authentication for digital customer journeys.
Website: https://www.miteksystems.com/
SEON
Fraud-prevention platform combining device, behavioural and digital-footprint intelligence with risk scoring and fraud-detection tools for online businesses.
Website: https://seon.io/
From Detecting More Fraud to Making Better Decisions
AI fraud detection creates value when it helps organisations become more accurate, not simply more suspicious.
That means evaluating both sides of every fraud decision:
What fraudulent activity did we stop?
and
What legitimate activity did we allow to proceed?
Combining contextual intelligence, appropriate rules, risk-based verification, machine learning and human expertise can help fraud teams move towards more precise decisions.
The goal is not zero friction or zero fraud at any cost.
It is to apply the right level of friction to the right transaction at the right time.
The Fraud Prevention Summit connects senior fraud and risk professionals with carefully selected suppliers through pre-arranged one-to-one meetings, providing an opportunity to explore AI fraud detection, identity verification, risk intelligence and wider fraud-prevention technologies.
Related Reading
This article follows our earlier AI Fraud Detection Software: What Fraud Teams Should Compare buyer’s guide, covering rules, machine learning, data sources, explainability, integrations, model performance and supplier evaluation.
Together, the two articles move from choosing the right AI fraud-detection capability to using that capability to reduce fraud without unnecessarily blocking genuine customers.
Image credit: https://unsplash.com/photos/a-group-of-people-working-on-computers-in-a-room-3yb7ZsaY0LY
