Your card gets declined at a restaurant while you are standing there. Nothing is wrong with your account. A model decided the transaction looked unusual and blocked it. That experience — irritating, common, and almost never explained — is the visible edge of one of the most heavily deployed AI systems in Indian banking.
Why Rules Alone Stopped Working
Early fraud systems were rule-based: block transactions above a threshold from a new location, flag more than five attempts in ten minutes.
Rules have two fatal properties. They are discoverable — fraudsters probe until they find the boundary, then operate just inside it. And they are static — every new fraud pattern requires someone to notice it and write a rule, by which time losses have occurred.
Machine learning models instead learn what normal looks like across millions of transactions and flag deviations, including patterns nobody anticipated.
What the Model Sees
| Signal type | Examples |
|---|---|
| Transaction attributes | Amount, merchant category, channel, currency |
| Behavioural history | Your typical amounts, timing, merchants, frequency |
| Velocity | Transactions per minute, hour, day |
| Device and location | Device fingerprint, IP, geographic plausibility |
| Network patterns | Links to accounts already associated with fraud |
| Session behaviour | Typing rhythm, navigation, time on screen |
The behavioural row does most of the work. The model builds a profile of your normal and scores deviation from it. A ₹40,000 purchase is unremarkable for one customer and a strong signal for another.
Geographic plausibility is a classic: a card used in Vadodara and then in Kolkata eleven minutes later is physically impossible, and no rule needs to be written for that specific pair of cities.
The Speed Constraint
A card authorisation must return in milliseconds. This shapes the architecture more than anything else.
| Layer | Timing | What it does |
|---|---|---|
| Real-time scoring | Milliseconds | Approve, decline, or step up to additional verification |
| Near-real-time | Seconds to minutes | Richer models, may trigger a customer contact |
| Batch analysis | Hours to days | Network analysis, mule account detection, pattern discovery |
The heavy analytical work happens in batch, where a model can examine relationships across accounts. Real-time scoring is deliberately lightweight because it has no time to be otherwise.
The False Positive Problem
This is the central tension, and it explains your declined restaurant payment.
Fraud is rare. If roughly one transaction in a thousand is fraudulent, a model flagging 1% of transactions generates ten false alarms for every genuine catch — even while performing respectably in statistical terms.
| Threshold setting | Fraud caught | Legitimate transactions blocked |
|---|---|---|
| Aggressive | High | High — many customer complaints |
| Balanced | Moderate | Moderate |
| Conservative | Lower | Low — but fraud losses rise |
There is no setting that catches all fraud without blocking legitimate customers. Every bank is choosing a point on this curve, and that choice is commercial as much as technical: the cost of fraud losses versus the cost of frustrated customers who move their business.
This is also why step-up authentication exists. Rather than a binary approve or decline, the system requests an OTP or app confirmation for borderline cases — catching fraud without blocking the genuine customer outright.
The Adversary Adapts
Fraud detection is unusual among ML applications in that the thing being predicted actively fights back.
A credit model’s borrowers do not restructure their behaviour to defeat it. Fraudsters do exactly that — testing thresholds with small transactions, mimicking legitimate patterns, distributing activity across accounts to stay below velocity limits.
Consequences that follow:
- Models require frequent retraining — drift here is adversarial, not just environmental
- Banks disclose little about how detection works, since disclosure aids evasion
- Detection logic is deliberately varied so probing one bank’s boundary does not reveal another’s
See model drift for why this degradation is faster here than in most applications.
Where Fraud Is Heading in India
The RBI’s Financial Stability Report noted that major banks and NBFCs identified AI-enabled cyber threats as their biggest cyber risk over the following twelve months. The tools that improve detection also improve attack.
| Emerging vector | What it involves |
|---|---|
| Voice cloning | Impersonating a family member or bank official in a call |
| Deepfake video | Defeating video KYC and verification |
| Synthetic identities | Fabricated identities built from combined real and false data |
| Scaled social engineering | Personalised phishing generated at volume |
| Mule networks | Layered accounts moving funds to obscure the trail |
See deepfake fraud in finance for what this means for you personally.
Practical Points for Customers
- Tell your bank before travelling. Location change is a strong signal, and forewarning prevents blocks.
- A decline is often protective. Irritating, but it is the system working.
- Never share an OTP. No legitimate bank employee will ask for one. Step-up authentication only protects you if the second factor stays with you.
- Report unauthorised transactions immediately. Liability rules depend materially on how quickly you report.
- Voice is no longer proof of identity. A caller sounding like someone you know is not verification.
Key Takeaways
- Rules are discoverable and static; ML learns normal behaviour and flags deviation
- The model scores against your behavioural profile, not a universal threshold
- Architecture is layered: milliseconds for authorisation, batch for network analysis
- False positives are unavoidable because fraud is rare — every bank picks a point on that curve
- Step-up authentication exists to avoid a binary approve-or-decline
- Fraud detection is adversarial — models degrade faster and retrain more often
- Banks and NBFCs named AI-enabled cyber threats their biggest cyber risk in RBI’s FSR
- Never share an OTP; voice is no longer proof of identity
Frequently Asked Questions (FAQ)
Q: How does AI fraud detection work in banking?
Machine learning models learn normal transaction behaviour from millions of past transactions, then score each new one for deviation. Signals include amount, merchant, device, location, timing and your own historical patterns. High-scoring transactions are declined or escalated to additional verification.
Q: Why did my bank block a legitimate transaction?
Because it deviated enough from your usual pattern to score above the bank’s threshold — an unusual amount, a new location, an unfamiliar merchant. Fraud is rare, so any sensitive model inevitably flags many legitimate transactions. It is the system working, if inconveniently.
Q: What is a false positive in fraud detection?
A legitimate transaction incorrectly flagged as fraudulent. Because genuine fraud is rare, false positives substantially outnumber real catches at any useful sensitivity. Banks tune the threshold to balance fraud losses against customer frustration.
Q: Why do banks ask for an OTP on some transactions but not others?
That is step-up authentication. Rather than approving or declining outright, the system requests a second factor for borderline cases — catching potential fraud without blocking a genuine customer. Lower-risk transactions pass without it.
Q: Can fraudsters defeat AI detection systems?
They actively try. Fraud is adversarial — attackers probe thresholds with small transactions and adapt their patterns, which is why models here degrade faster than in other applications and require frequent retraining. Banks also disclose little about detection logic, since disclosure aids evasion.
Q: Is AI making fraud worse or better?
Both. Detection has improved substantially, but the same technology enables voice cloning, deepfakes, synthetic identities and personalised phishing at scale. RBI’s Financial Stability Report recorded that major banks and NBFCs consider AI-enabled cyber threats their biggest cyber risk.
Q: What should I do if I spot an unauthorised transaction?
Report it to your bank immediately through official channels. Liability for unauthorised electronic transactions depends significantly on how promptly you report, so speed matters. Do not wait to see whether it reverses on its own.
Related Reading: