1. Why False Positives Are AML's Flagship Problem

Typical Rule-Based System:
Rule: Cash Deposit > Rs 10 Lakh in Savings account in a day = STR
Result: A jeweller, property dealer, hospital - all flagged daily. Compliance team drowns.

Legal Problem:
PMLA mandates banks must file STR within 7 days of suspicion to FIU-IND. But if you file 10,000 false STRs, FIU-IND considers it defensive reporting and will audit you. If you miss 1 true laundering case, ED will prosecute under Vijay Madanlal judgment.

RBI allows from 1 April 2026: Additional risk-based checks beyond standard 2FA, including AI/ML for AML triage.

2. Legal Framework - What Law Requires

A. PMLA, 2002 Sec 12 - Obligations of Reporting Entity:

• Maintain records of transactions
• Furnish information to FIU-IND (CTR, STR, CCR)
• Verify identity of clients - KYC

B. Supreme Court - Vijay Madanlal Choudhary v. Union of India (2022) 3-Judge Bench - 2022 SCC Online SC 929
This is the constitution of AML in India.
Held:

• Upheld core of PMLA - ED powers of arrest, search, attachment, stringent twin conditions for bail under Sec 45 are constitutional
• For provisional attachment under Sec 5, registration of predicate offence is NOT mandatory
• ED officers under PMLA are not police officers - their statements admissible
• If person is finally discharged/acquitted of scheduled offence, no PMLA offence remains

Implication for False Positive: Supreme Court emphasized KYC procedures as preventive measures that protect integrity of financial system while balancing individual privacy. You cannot do blanket blocking. You need risk-based approach.

C. RBI Master Directions on KYC 2016 (Updated 2024) + FIU-IND Guidelines:

• Risk categorization: Low, Medium, High - based on customer type, product, geography
• Periodic KYC update
• Transaction Monitoring - Must use Risk-Based Approach, not rule-only

D. Recent Judgments on Account Freezing:

• State Bank of India v. Rajesh Agarwal (2023) and SBI v. Amit Iron (7 April 2026) - We discussed - Before tagging account as fraud / freezing, give notice + share forensic audit report. Can't freeze on mere false match.
• Kerala High Court - Judicially Crafted SOP on Bank Powers to Freeze under PMLA - Monitoring provisions mandate filing STR, but do NOT confer express power to freeze. Freezing more prejudicial than closure.

3. Proven False Positive Reduction Techniques - Accepted by Regulators

Regulators globally accept this: ML fixes false positives mainly by triage and segmentation on top of rules, never by silently deleting regulated scenarios.

Technique 1: Customer Segmentation + Dynamic Thresholds
Don't use Rs 10 Lakh for all.

• Salaried person: Rs 2 Lakh cash deposit = alert
• Jeweller: Rs 50 Lakh = normal - set threshold at Rs 1 Cr based on past 12-month average
• Use peer group: Compare jeweller with jewellers in same pin code turnover
  Result: Infosys BPM case study for SCB - Reduced false positives by 40% through segmentation.

Technique 2: Multi-Factor Risk Scoring Instead of Single Rule
Combine:
Behaviour (velocity, deviation from past) + Customer Risk (High risk geography, PEP) + Product Risk (trade, crypto) + Network Risk (linked to mule account)
Score: 0-100. Only score >75 goes to analyst.
Example: Cash Rs 9.9 Lakh alone = 30 score. Same + account opened 7 days ago + address is high-risk district + 5 accounts at same address = 90 score = True alert.

Technique 3: AI Triage + Feedback Loop (RBI approved from April 2026)
Tookitaki deployed at United Overseas Bank (UOB) Asia:

• Two objectives: Reduction in False Positives + Better detection
• Use supervised ML trained on analyst decisions (True / False) for 2 years
• Auto-close: If model confidence False Positive > 95% and pattern matches historical false - auto-close with audit trail
• Human review: Only 10% highest risk

This minimizes superfluous alarms, allowing AML teams to concentrate on high-risk.

Technique 4: Scenario Logic Validation and Calibration
Common pitfalls:

• Same PAN used for threshold aggregation but joint accounts counted twice
• Round amount detection - Rs 10,00,000 flagged but Rs 9,99,999 not - Fraudsters exploit. Use bands, not exact.
• Time window: 24h rolling vs calendar day

Quarterly validation: Run scenario on past 6 months - how many alerts became STR? If conversion <2%, recalibrate.

Technique 5: Whitelisting and Good Guy List with Maker-Checker
Delhi HC in Maker-Checker case 2026 W.P.(C) 1571/2026 - Sakshi Sharma v UoI: Maker-checker is only risk-control, not immunity. But for AML, after investigation, if customer is hospital receiving insurance payments daily, whitelist with:

• Reason
• Expiry (6 months)
• Approver - MLRO + Principal Officer
• Audit

Prevents same false alert daily.

4. How Courts View False Positive Reduction

If you reduce too aggressively and miss laundering:

• Vijay Madanlal will apply - ED will treat bank officials as facilitators if wilful blindness proved. Case 10: Supreme Court invalidated bail conditions but upheld strict liability.

If you don't reduce and harass customers:

• Banking Ombudsman + Consumer Court will hold deficiency. Example: Customer Did Not Share OTP but SBI's 2FA Failed - Delhi HC held glaring deficiency.
• SBI vs Rajesh Agarwal - No notice before fraud tag = violation of natural justice.

Balance Test given by Supreme Court 29 Oct 2026 in Digital Arrest Suo Motu SMW(Crl) 3/2025:
Banks must compensate if due diligence fails. Court proposed directing banks to use AI and transaction history to identify and temporarily hold suspicious transactions. So reduction is allowed, but must be intelligent, not blanket suppression.

5. Recommended Implementation Framework - RBI Compliant

Step 1: Data Quality First
Poor quality data = false positives. As per KYC/AML Audit checklist: Gap in KYC - poor quality data creates false positives & less time to focus on real risks. Clean PAN, address, occupation.

Step 2: 3-Layer Architecture
Layer 1 - Rules: Regulatory mandatory (CTR >10L, cross-border >5L)
Layer 2 - Risk Scoring: Behaviour + Peer
Layer 3 - AI Triage: Auto-close low, escalate high

Step 3: Documentation
For every auto-closed alert, keep:

• Reason code
• Data points used
• Model version
• As per OFAC Name Matching guidance: Compliance documents a match-analysis memo (true match / false positive) with sources.

This protects you in ED audit.