Explore

Kaptur - Flow post claims

Summary

A fresh sizing based on 3.3k SF claims was done where all images in the RVP leg were present (Pickup, RPC, ICSD, Evidence)

This was split between our Product team (1.5k) and Agent audits (1.8k)

Both audits converged to a similar result, thereby building conviction in the sizing

Of 16% CR-CAS, sellers abuse is the leading cause with 7.5pp (47% of claims) contribution

Specifically, 2 use-cases of Sellers abuse:

Swap Product Claims - 6.5 pp (~40.5% of claims) | Cases where the Seller received the correct product, but swapped It with a completely different product to raise a claim (Evidence <> ICSD)

Same Product Claims - 1 pp (~6.5% of claims) Cases where Sellers received a correct product and raised a claim on that same product (Evidence =Catalog)

Remaining 8.5pp of CR-CAS is composed of:

WFR - 4.4p (27% of claims)

3P swaps - 1.8pp (11% of claims)

Incorrect pickups - 1.4pp (9% of claims)

Subjective cases - 1pp (6% of claims)

Subjective cases contribute 1pp to CR CAS. These are cases where the product that was picked was delivered and shown in the claim evidence. While, the product is very similar to catalog, it is not an exact match. These were classified as Seller Abuse in previous sizing exercises.

This is line with the

previous sizing exercise⁠

done during R2R where Seller-side issues accounted for 43% of total claims.

Following are 2 key call-outs:

For Seller-abuse (7.5 claims pp), focus will be to double-down & scale identified WIP levers: One key difference we see from previous sizing is that Seller side fraud MO has shifted from Same Product Claims to Swap Product Claims. This shift was expected due to graded penalty running on EvsC flagged cases. Hence, we remain confident on our seller side claim reduction levers that were presented at start of R2R to tackle both these problems (ICSD for Swap Product Claims detection; EvC for Same Product Claims detection; Graded Penalty as the uber-penalty lever)

Attribution

Fraud MO

CR CAS (pp)

Claim OC%

Initiatives

Status

Seller

Unwarranted Claim/Seller Abuse

0.97

Evidence vs Catalog (EvsC)

Graded Penalty

Live

For graded penalty, apart from EvsC flagged tickets, we also re-audit 10% of all rejected tickets

We are also working on adding new signals to the DS model to identify the riskiest AWBs for flagging images

Seller

Seller Swap

6.54

40.8%

Image at Seller Doorstep (ICSD)

Graded Penalty

Live on 9.8% RVP OC (26.17% Claims OC)

⁠

2. For WFR (4.4 pp) reduction, focus is on better instrumentation/deterrent experiments for right attribution & actioning:

We have started a sizing exercise from a WFR claims perspective to understand where the problem lies: Seller, FE, User. Post sizing, we will work on the following three pillars

Solutioning Lever

Stage in Supply Chain

Initiatives

Status

Improve Seller Accountability

at Pickup

We will be exploring the following experiments

Mandating Scan and Pack for P4-P6 sellers and high WFR claiming sellers. We have identified a pareto set of 500 sellers having high WFR claim rate.

Mandating BP-TP via claim blocking for high WFR sellers.

Image Capture at Pickup for High WFR P1-P3 Sellers

Scan and Pack, Claim Blocking To go live by 9th Aug

Image Capture at Pickup is under exploration.

Improve Seller Accountability

at Claim Raise

Extend Evidence vs Catalog to compare Evidence Image with Scan and Pack Image to prevent Same Product Claims.

Efficacy Sizing Exercise: See clarity of scan and pack images and size where catalog image NOT = scan and pack images. ETA 15 Aug

Integration into existing EvsC model: High level 2-3 weeks tech effort

Improve FE Accountability

at User Delivery

For Risky WFR orders, we will do

Image at customer doorstep based on supplier cohort and order value

We will compare Image at Customer doorstep (ICCD) image with Scan and Pack Image to identify fraudulent FEs

Pilot ongoing for Image at Customer Doorstep

Audit are being done to see efficacy of comparing ICCD with Scan and Pack Image.

Pilot and audit to be concluded by Aug.

Improve User Accountability

at Return Initiation

Increasing coverage of user uploaded image in return raising process basis user grading model and order value.

At time of return initiation compare user uploaded image with ICCD to decide if return has to be initiated or not

Audit are being done to see efficacy of comparing ICCD with User uploaded image.

Audit to be closed by 13th Aug.

Improve User Accountability

at Return Pickup

Compare ICCD with Pickup Images via FE at time of pickup OR post pickup via agent audits. Fraudulent users will be barred from platform

Pilot to be started with Valmo on selected Hubs in Aug

⁠

⁠

Full list of Initiatives for each problem type are detailed

here⁠

⁠

Details

Methodology

We conducted two separate audits to understand where abuse happens in RVP (customer) returns on the platform. These audits were conducted on claims raised under “I have received wrong return”. Audits were conducted by Product team (1.5k cases) and Agents (1.8k cases) separately.

During these audits we separated out subjective cases—where the product received by sellers closely matches their catalog—from other types of returns. This separation is intended to ensure that subjective cases are not misclassified as seller abuse, as occurred in previous sizing exercises.

Audits were done by utilizing images available from various stages in the supply chain to understand where abuse occurs. The images used for this audit included:

Customer Doorstep Pickup Image

RPC Image (taken by the hub manager at the seller's RTS hub)

Image at Seller Doorstep (ICSD) (taken by the field executive at the seller's doorstep)

Evidence Image/Video (provided by the seller during the claim raising process)

In addition to categorizing claims by customer return type (Param and Smart QC), we also differentiated based on the type of product received by the seller (as compared to the Catalog Image). For sizing purposes, we split them into

Completely different category

Major Design/Color change

Subjective cases (Slight design/colour change) - called out separately (previously part of Seller Abuse)

Same Product (attributed to Seller Abuse)

Results

Overall Attribution

Overall Attribution

⁠

QC/Non-QC Split

QC/Non-QC Split

⁠

Next Steps

Attribution

Fraud MO

Claim OC%

Initiatives

Status

Seller

Unwarranted Claim/Seller Abuse

Evidence vs Catalog (EvsC)

Graded Penalty

Live

For graded penalty, apart from EvsC flagged tickets, we also re-audit 10% of all rejected tickets

We are also working on adding new signals to the DS model to identify the riskiest AWBs for flagging images

Seller

Seller Swap

40.8%

Image at Seller Doorstep (ICSD)

Graded Penalty

Live on 9.8% RVP OC (26.17% Claims OC)

Incorrect Pickup

8.6%

Pickup vs Catalog (PvsC)

Subjective QC

PvsC - Live with agent auditing on 15% RVP OC.

Subjective QC - We are exploring if we can add extra questions in the QC pickup process subject to the SSCAT being picked up that will help FEs make a better call and subsequently reduce incorrect pickups. (Eg: for collared tshirts sscat add a question saying "Does Tshirt have a collar?").

We are in discussion with Category and Catalog team for the same

LM FE Swap

3.8%

Secondary QC

RPC vs Pickup Image

Secondary QC - Deprioritised this R2R in favour of ICSD.

RPC vs Pickup Image - Live

FM FE Swap

7.2%

RPC vs ICSD Image

Seller Level Bagging (Exp)

RPC vs ICSD Image - Live

Seller Level Bagging - Experiment went live last week with SF

WFR

Subjective cases

6.2%

Product Level Tagging (Exp)

Experiment To go live in August

WFR

Forward Leg Issues (Seller/FE/User)

27.5%

Sizing Exercise to attribute WFR within Seller/FE/User in Forward Leg

Based on sizing, initiatives to be picked up based on these

solutioning levers

⁠

Sizing Exercise - To be concluded by 21st Aug

⁠

WFR solutioning levers

Solutioning Lever

Stage in Supply Chain

Initiatives

Status

Improve Seller Accountability

at Pickup

We will be exploring the following experiments

Mandating Scan and Pack for P4-P6 sellers and high WFR claiming sellers. We have identified a pareto set of 500 sellers having high WFR claim rate.

Mandating BP-TP via claim blocking for high WFR sellers.

Image Capture at Pickup for High WFR P1-P3 Sellers

Scan and Pack, Claim Blocking To go live by 9th Aug

Image Capture at Pickup is under exploration.

Improve Seller Accountability

at Claim Raise

Extend Evidence vs Catalog to compare Evidence Image with Scan and Pack Image to prevent Same Product Claims.

Efficacy Sizing Exercise: See clarity of scan and pack images and size where catalog image NOT = scan and pack images. ETA 15 Aug

Integration into existing EvsC model: High level 2-3 weeks tech effort

Improve FE Accountability

at User Delivery

For Risky WFR orders, we will do

Image at customer doorstep based on supplier cohort and order value

We will compare Image at Customer doorstep (ICCD) image with Scan and Pack Image to identify fraudulent FEs

Pilot ongoing for Image at Customer Doorstep

Audit are being done to see efficacy of comparing ICCD with Scan and Pack Image.

Pilot and audit to be concluded by Aug.

Improve User Accountability

at Return Initiation

Increasing coverage of user uploaded image in return raising process basis user grading model and order value.

At time of return initiation compare user uploaded image with ICCD to decide if return has to be initiated or not

Audit are being done to see efficacy of comparing ICCD with User uploaded image.

Audit to be closed by 13th Aug.

Improve User Accountability

at Return Pickup

Compare ICCD with Pickup Images via FE at time of pickup OR post pickup via agent audits. Fraudulent users will be barred from platform

Pilot to be started with Valmo on selected Hubs in Aug

⁠

An overall image-triangulation data layer that will run on top of these accountability levers to compare- and assign attribution to right stakeholder.

⁠

Samples (Optional)

Cases of seller swap:

Seller Swap was verified as Pickup Image = RPC image = ICSD Image BUT ICSD Image is != Claim/Evidence Image. Further be verified the Packet ID and Timestamp to make sure the FE had delivered the correct shipment

https://docs.google.com/spreadsheets/d/1XTecnOG1L1DDJ9eeFxwh-jx1F7aBG9uL-X1-m0_0l9w/edit?gid=1079512543#gid=1079512543⁠

⁠