Swiggy
Swiggy Instamart uses vision AI to settle refund claims and verify returns
Swiggy Instamart uses vision-language AI to check customer claim photos and to verify items at return pickup, replacing manual agent review for clear cases. Swiggy reports that the return check's false Not Match rate fell from 60% at launch to 17%, and that more than 5000 return tickets a day are resolved without a human.
Key figures
Copied exactly from the sources. Copy any figure with its citation.
60%
to 23% to 17%
False Not Match rate on returns, from launch to after automated prompt optimisation1
98.5%
Accuracy of the Reverse Quality Check (RQC) on returns1
>5000
RQC tickets solved without human intervention each day1
up by 6%
Change in customer satisfaction score1
What it took
Cost, team size and timeline were not disclosed in the sources.
The problem
Swiggy Instamart receives more than 40,000 customer claims a day about wrong, expired, damaged, poor-quality or badly packed items. Whenever rules and fraud models could not settle a claim from a photo, a human agent compared it with the catalog and decided. Costs grew in step with volume, decisions were inconsistent and customers waited. For returns, delivery staff were meant to check the item at pickup but were paid for speed, so most returns were accepted unchecked.
What they built
Swiggy built two systems that use vision-language models, which are AI models that read both images and text. The Image Validation Model (IVM) examines a customer's claim photo against the catalog image and returns valid, invalid or ambiguous. The Reverse Quality Check (RQC) compares the delivery partner's photos at pickup with the customer's claim photos and returns match, not match or ambiguous.
How they did it
Each claim type has its own handler and prompt, and the customer photo is treated as the main evidence. Code, not the model, handles routing, combining several photos and the final approve, reject or escalate decision. Expired-item claims use two ways of reading label dates in parallel. If RQC cannot confirm a match, the delivery partner gets a specific message and can retry, up to 2 attempts, before manual review. If the system is down, the claim goes to manual review.
Results
The RQC false Not Match rate, meaning legitimate returns wrongly flagged as a mismatch, started at 60%, fell to 23% with prompt work and to 17% with automated prompt optimisation. Swiggy also lists more than 5000 tickets a day solved without human intervention, 98.5% RQC accuracy, a 6% rise in customer satisfaction score and 20,000 fewer weekly Image Validation tickets.
What to copy
Compare candidate models on your own labelled cases, including cost and speed. Keep a human route for unclear cases. Give the people on the ground precise feedback instead of a generic error.
All reported figures
| Measure | Figure | Citation |
|---|---|---|
| False Not Match rate on returns, from launch to after automated prompt optimisation | 60% to 23% to 17%1 | |
| Accuracy of the Reverse Quality Check (RQC) on returns | 98.5%1 | |
| RQC tickets solved without human intervention each day | >50001 | |
| Change in customer satisfaction score | up by 6%1 | |
| Decrease in weekly Image Validation ticket volume | 20,0001 | |
| Claims processed each day | more than 40,0001 |
Lessons for your project
Give the model the visual evidence and keep decisions in code. Swiggy lets the model judge photos, while routing, rules and money decisions stay deterministic.
Test more than one model on your own production data. Swiggy uses a different model for expiry-date reading than for other claim types.
Tell field staff what to fix. Swiggy says about 60% of first failed matches came from bad photos, so specific retry prompts help.
Expect a poor start and plan to tune. The return check began with a 60% false Not Match rate and was improved by prompt work.
Caveats
- All figures are Swiggy's own, from one engineering blog post. No outside party confirms them.
- The post lists its results without definitions. It does not say how the 98.5% RQC accuracy or the 6% satisfaction change were measured, or the baseline period.
- The post gives no launch date. It says the systems have run in production for over 6 months, so the deployed date shown is the post date and the actual launch was earlier.
- The post names the models used: Gemini 2.5 Flash for most claim types and return checks, and GPT-4o for expired-item claims. It also benchmarked Claude Sonnet 4, which was not chosen.
- The business impact on fraud losses and inventory leakage is described in words only, without a figure. About 10 to 15% of returned orders being incorrect or short is Swiggy's pre-launch estimate.
- No cost, team size or build time is given.
Sources
- Building Production-Scale Visual AI for Claims and Returns at Swiggy Instamart
Swiggy Bytes (Medium) · 2026-06-16 · Company blog · medium.com
Summarized in our own words from the sources above. Spotted an error? See corrections.