Fundle.ai: Retail Loyalty Automation with AI Agents Combats Fraud

Types of Loyalty Fraud in Indian Retail

Before deploying any detection system, a CRM Head needs a precise taxonomy of what they are defending against. Loyalty fraud in Indian retail clusters into six distinct categories, each with different signatures and different damage profiles.

The first and most prevalent is phantom transaction fraud — the submission of fabricated or inflated receipts to earn points that were never legitimately generated. In formats like Apollo Pharmacy or Reliance Trends where high-frequency, small-ticket purchases are the norm, fraudsters submit doctored bills or reuse genuine receipts from multiple accounts. The average phantom bill in pharma loyalty runs ₹800–1,200, small enough to stay below manual review thresholds but large enough to accumulate meaningful point balances across hundreds of accounts over weeks.

The second category is account takeover (ATO). As Indian consumers consolidate high-value loyalty balances — a Tanishq member with ₹15,000 in accumulated points, or a mall loyalty member with a ₹8,000 reward balance — these accounts become worth targeting. Credential stuffing attacks using leaked databases from unrelated breaches are the most common vector. Platforms without behavioral biometrics or device-fingerprinting have no way to distinguish a legitimate login from an ATO attempt.

Third is referral and enrollment fraud. Programs offering sign-up bonuses or referral incentives — common at Lifestyle, Pantaloons, and FabIndia — are particularly vulnerable to bot-driven mass enrollment using disposable phone numbers. India's low cost of prepaid SIM acquisition (₹49–99 per connection) makes this trivially cheap to execute at scale.

Fourth is employee collusion, which the data suggests accounts for nearly 62% of all confirmed loyalty fraud cases in organized retail. A store associate at a Cafe Coffee Day or a Manyavar outlet can award points to their own account or that of a confederate by scanning a loyalty card against real but unrelated customer transactions. POS-level controls that don't cross-reference loyalty ID to the authenticated billing customer are blind to this.

Fifth is return fraud combined with point laundering. A fraudster purchases merchandise, earns points, then returns the merchandise for a cash or wallet refund — retaining the points. Without a points-reversal trigger that is tightly coupled to the returns management system, this cycle can be repeated indefinitely.

Sixth, and increasingly sophisticated, is synthetic identity fraud — the construction of entirely fabricated customer profiles using combinations of real and fictitious PAN, phone, and address data. These profiles pass basic KYC checks but exhibit behavioral patterns — rapid point accumulation, no genuine purchase diversity, immediate high-value redemptions — that only a machine learning model will catch.

Using AI Agents to Detect and Prevent Fraudulent Activities

The fundamental architectural shift that AI agents bring to loyalty fraud prevention is the move from batch detection to continuous, real-time inference. A rules engine asks: 'Did this transaction exceed a threshold?' An AI agent asks: 'Does this transaction fit the behavioral fingerprint of this specific customer, at this specific time, from this device, in this location, relative to every other transaction this account has ever generated?' These are categorically different questions, and only the second one catches sophisticated fraud.

AI agents in a loyalty context operate across three distinct defense layers. The first is enrollment-time identity verification. Before a point balance can even be created, an AI agent evaluates the incoming enrollment signal — phone number velocity (has this number been used for three other enrollments this week?), device fingerprint (is this the same handset that created 14 accounts last month?), and network graph analysis (does this referral chain look like a legitimate social graph or a synthetic ring?). Programs running on legacy CRM stacks like basic configurations of EasyRewardz or older Capillary setups typically lack this layer entirely.

The second defense layer is transaction-time anomaly scoring. Every point-earning event is scored against a multi-dimensional model that incorporates: time-of-day against customer's historical pattern, basket composition against category affinity model, location against home-store cluster, transaction value against cohort median, and inter-transaction velocity. A Reliance Trends customer who typically shops twice a quarter suddenly submitting three receipts in 90 minutes from three different city locations is an automatic hold — not a manual review queue, but an actual hold that prevents point posting until a secondary validation clears.

The third layer is redemption-time authorization. This is where the financial loss actually occurs, making it the highest-value intervention point. AI agents apply a risk score to every redemption request that factors in: time elapsed since earning (very fast earn-and-burn is a classic fraud signal), redemption value as a percentage of total balance, channel of redemption request, device consistency with enrollment device, and peer-group behavioral norms. If a member's redemption request scores above a defined risk threshold, the agent can automatically route to step-up authentication — OTP, manager approval, or delayed fulfillment — without any human intervention in the workflow.

What makes this approach materially different from competing platforms like MoEngage or WebEngage (which are excellent engagement orchestration tools but were not architected for fraud prevention) or even Xeno and Almonds.ai (which focus on campaign automation) is that the AI agent layer is not bolted on — it is native to the transaction processing pipeline. Detection latency matters enormously: a fraud event detected in 200 milliseconds can be stopped; one detected in 24 hours cannot.

Fundle's Fraud Detection Mechanisms

Fundle's AI-driven loyalty automation reduces fraud risk across 270+ partner brands — not by adding a fraud module on top of a loyalty platform, but by treating fraud prevention as a first-class concern in the platform's core architecture. This distinction is operationally significant for a CRM Head evaluating vendors.

The Fundle AI Platform embeds what the team calls 'trust scoring' into every loyalty event object. When a transaction is ingested — whether from a Petpooja POS, a POSist terminal, a GoFrugal retail management system, or a Wondersoft fashion retail POS — the platform's Fundle AI Agents immediately enrich that event with identity signals, behavioral signals, and network signals before the event is written to the ledger. This means the point balance a customer sees in their app reflects only verified, clean transactions.

Fundle Mall Loyalty specifically addresses the cross-tenant fraud challenge that pure brand-level programs cannot see. In a mall like Select CITYWALK or a Phoenix Marketcity property, a fraudster can exploit the earn structure across 80–100 brand tenants simultaneously — earning at a fashion anchor, a food court operator, a multiplex, and a pharmacy in the same session using different devices or accounts. Fundle's cross-brand graph model treats the entire mall as a single behavioral universe. Suspicious activity in one tenant immediately elevates the risk score for that device and identity cluster across all other tenants in the same property.

Fundle Brand Loyalty adds a category-specific fraud model layer. A jewelry brand like Tanishq has completely different fraud vectors from a QSR like Cafe Coffee Day. High-value, low-frequency purchases at jewelry require models tuned for outlier detection on transaction value. High-frequency, small-ticket formats require velocity and receipt authenticity models. Fundle's model library maintains category-calibrated parameters rather than applying a one-size-fits-all threshold.

Fundle Agentic AI goes a step further than detection — it acts. When a fraud signal crosses a defined confidence threshold, Fundle AI Workflow triggers an automated response sequence: point posting is suspended, the account is flagged for step-up authentication, the brand's CRM team receives an alert with evidence package, and — for confirmed fraud — the account is quarantined and the fraud event is logged to the compliance ledger. No analyst needs to wake up at 2 AM to stop a redemption campaign from being drained.

Real-Time Analytics and Alerts for Loyalty Program Security

Operationalizing fraud prevention requires more than a detection model — it requires a monitoring and alerting infrastructure that puts actionable intelligence in front of the right people at the right time. This is where many technically capable platforms fall short: they generate signals but bury them in dashboards that no one checks until month-end.

A mature real-time analytics layer for loyalty fraud should surface three categories of intelligence. The first is transaction-level alerts — immediate notifications when a specific transaction or account crosses a fraud threshold. These are the alerts that need to reach a mall operations manager or brand CRM manager within minutes, not hours. They should include the specific signal that triggered the alert (velocity anomaly, device mismatch, cross-tenant ring pattern), the financial exposure at risk, and a one-click action to suspend, escalate, or clear.

The second category is pattern-level intelligence — daily or weekly aggregated views of emerging fraud vectors. If receipt manipulation at a specific brand-tenant starts spiking on weekends, this pattern may not trigger any individual transaction alert but represents a coordinated campaign that needs a structural response (tighter receipt validation, manager approval for weekend point posts above a threshold). Pattern intelligence is what transforms a reactive fraud team into a proactive one.

The third category is program health metrics — the KPIs that tell you whether your fraud controls are calibrated correctly. The five metrics every CRM Head should track are: fraud reversal rate (percentage of posted points subsequently reversed as fraudulent, target below 0.5%); false-positive rate (percentage of legitimate transactions flagged, target below 2%); mean time to detection (MTTD, target under 5 minutes for high-confidence fraud); net loyalty liability exposure (total unredeemed points adjusted for estimated fraudulent balances); and fraud-as-percentage-of-gross-merchandise-value redeemed (your single most comparable benchmark against industry peers).

The Fundle AI Platform surfaces all three intelligence categories in a unified operations console. Mall Marketing Directors get a property-level view; brand CRM heads get a tenant-level view; and program managers get a transaction-level drill-down — all from the same data pipeline with no reconciliation lag.

How Fundle solves this

Vineet Narang founded Fundle on the conviction that India's loyalty market needed a platform built for the complexity of Indian retail — multi-format, multi-brand, multi-POS, and increasingly under siege from sophisticated fraud that legacy point platforms were never designed to handle. That conviction is visible in every layer of the Fundle AI Platform's architecture.

Fundle Loyalty's core data model treats every loyalty event as a rich, multi-signal object rather than a simple debit-credit entry. When a transaction arrives from a GoFrugal terminal at a Reliance Trends store or a Wondersoft system at a premium fashion retailer, Fundle AI Agents enrich it with identity confidence scores, behavioral anomaly scores, and network risk scores — all before the point balance updates. This means fraud prevention is not a post-processing step; it is embedded in the transaction lifecycle itself.

For mall operators, Fundle Mall Loyalty provides the cross-tenant visibility that no single-brand CRM platform can offer. The unified event stream from all brand tenants feeds a property-level graph model that identifies coordinated fraud rings operating across multiple stores in a single visit — a pattern that is completely invisible to any tenant-level system. A Select CITYWALK or Phoenix Marketcity property manager can see, in real time, if a cluster of accounts is exhibiting coordinated earn-and-burn behavior spanning a fashion anchor, a food court, and a multiplex — and halt the ring before full redemption occurs.

Fundle Brand Loyalty extends the same intelligence to standalone brand programs — whether that is an Apollo Pharmacy chain running tens of thousands of daily transactions or a premium brand like FabIndia managing a high-value, relationship-driven member base. The category-calibrated fraud models mean that the detection logic is appropriate for the business model, not a generic threshold applied uniformly.

Fundle Agentic AI and Fundle AI Workflow close the loop between detection and response. When a fraud signal fires, the workflow engine executes the appropriate response sequence automatically — step-up authentication, account hold, brand alert, compliance logging — without requiring a human to initiate each step. Fundle AI Agents continuously monitor redemption queues, enrollment pipelines, and earning events, escalating only the cases that genuinely require human judgment. The result is a fraud operations function that scales with transaction volume without scaling headcount — exactly the operational leverage that a CRM Head managing a 200-brand program needs.