Flipkart Flippi: conversational discovery before autonomous commerce

External-source research synthesis, not an OfferSignal client result. Based on the retained October 2026 research and source-specific dates.

A shopper can describe a need more easily than a catalogue can interpret it. “Find me a phone for photography” still leaves budget, features, trade-offs and relevant offers unresolved. Flipkart’s Flippi addressed that discovery problem in India with a conversational layer. Public evidence does not show an autonomous buyer, supplier selector, negotiator, checkout or payment agent.

A chat layer for product decisions

Flipkart launched Flippi in October 2023 as a shopping assistant that let customers state what they wanted, receive options and narrow choices through conversation. 1 2 3

The documented task was product discovery and decision support. Flippi could answer questions about promotions and route post-purchase issues toward existing order-history and help interfaces. The product accounts and technical paper do not describe cart mutation, checkout, payment authorisation, binding quotes, supplier selection, negotiation or order placement. The shopper remained the decision-maker. Flipkart’s separate seller tools for generated descriptions and feature summaries are content automation, not evidence of a Flippi merchant agent. 4

How the system grounded conversation

The technical paper describes session history and the current request going to an LLM Standalone Query module, followed by a DistilBERT coarse-intent router. It routed product search, product-specific decision assistance, comparison, offers, customer experience, frequently asked questions, direct response and out-of-scope queries. 2

For decision assistance, the system retrieved catalogue specifications, selected the 20 most relevant and generated an answer. If the evidence could not answer the question, it used a template and retrieved the top three matching frequently asked questions or reviews. Comparison likewise reduced each product to 20 relevant features. This is retrieval-grounded assistance with deterministic routing, not purchasing authority.

The controls were useful but incomplete. Follow-up values were intended to match real assortment facets, yet the paper notes missing or limited facets and an example where the LLM could suggest an irrelevant premium brand for a low-budget request. It did not prove factual correctness, price freshness or neutral ranking.

What was measured—and what was not

Financial Express, with Times Now reporting the same company-reported trend, said more than four million users engaged with Flippi in October 2023, its launch month. That is reported market use, not customer benefit; no exposed-user denominator, engagement definition, sales, satisfaction or retention measure is supplied. 5 6

The authors describe A/B deployment and testing across Search, Browse and Category Landing Pages during an unnamed major sales event, later reducing touchpoints to about 15% of the initial scope. Their production table reports 1.3 million Engaged MAU—a unique user initiating at least one query in a month—about 1.0% last-touch conversion and 68% thumbs-up share. It also reports a 32% session-answerability improvement. Missing dates, populations, controls, definitions and uncertainty make these descriptive KPIs, not demonstrated causal lift or return on investment. 2

Offline diagnostics show why aggregate accuracy needs context. Human evaluation reported 92.34% query-modification accuracy and 95.23% product-disambiguation accuracy; GPT-4 evaluation reported 89.49% query-modification accuracy. Query-modification accuracy fell from 98.65% at turn one to 83.96% at turn five, while support fell from 4,000 to 449 turns. Overall intent accuracy was 97.34%, but macro F1 was 82.04%, product-question F1 58.82% and comparison F1 70.75%. Human summary factuality was 88.47%, and GPT-4 comparison-verdict correctness 88.18%. These are offline component tests, not market error rates or purchase outcomes. 2

A broader Flipkart interview offers a negative lesson rather than a Flippi result: a generative-AI support approach for highly anxious users was redirected to human agents because those users wanted a human component. It also stresses connecting model metrics to product metrics. 7

The practical boundary for commerce teams

For an illustrative retailer or distributor, the transferable pattern is an evidence-linked discovery assistant: canonical product IDs, typed specifications and facets, current price/promotion/availability timestamps, and provenance for each recommendation. Test zero-result rate, wrong-item rate, turn-depth reliability, clarification quality and incremental conversion separately. Keep the authoritative decision with the named commercial owner when an answer affects price, supplier, quote, order or payment.

Flippi’s later status remains unresolved. Economic Times reported SLAP in January 2026 as a phased standalone revamp or replacement, while Flipkart’s statement confirmed only work on initiatives. The official India App Store listing described SLAP as an English iPhone beta with limited early-access spots. That establishes a public successor listing, not definitive retirement of Flippi or autonomous transaction authority. 8 9

Sources and next reading

Flipkart Online Shopping App — Google Play — official Android listing; no current Flippi entry point is shown. Access requires a compatible device, account and regional availability.

SLAP: AI Shopping, Chat & Find — Apple App Store India — official successor listing; India App Store, iPhone and English, with beta early access and limited spots.

Explore OfferSignal · Focused Business & AI Workshop · All research

References