AI Deep Dive: India's E-Commerce Marketplaces

Which parts of a marketplace AI will transform, which it won't, and where the next generation of startups will be built - through the lens of India's largest marketplaces

Over the past weeks, we sat down with senior technology leaders at some of India's largest consumer marketplaces across new-age retail tech, hyperlocal commerce, e-commerce, traveltech, and social commerce. The question we asked each of them: what should a new-age marketplace know about AI adoption, based on what you’ve already lived through? Everything below is shared on the condition of anonymity: no companies, no individuals, learnings only.

Two things are happening inside India’s consumer platforms at once, and they pull in opposite directions.

The first is compression. Support, cataloging, onboarding, QA, analytics, coding, creative production, invoice review, hiring screens, warranty triage, and internal operations are getting cheaper. Work that used to require a queue, a dashboard, a team, or a vendor can now be delegated to a model, an agent, or a small internal team. Every workflow that is repetitive, measurable, and trapped inside a bad interface is being rethought.

The second is expansion. AI is not just making marketplaces cheaper to run. It is pulling new users, sellers, products, categories, and transactions into the market. Voice can reach consumers who never became fluent in e-commerce search. Catalog agents can onboard sellers who never had clean product data. Recommendation systems can create demand in categories where the user does not know what to ask for. AI-generated creative can make millions of micro-campaigns economically possible.

The old marketplace question was: can you aggregate enough demand and supply to make the flywheel spin? The new marketplace question is: can you convert messy intent, messy supply, and messy offline reality into a governed transaction better than anyone else?

But the most important shift is subtler than both. The consumer interface is not converging on a single universal chat box. It is fragmenting by category behaviour. For transactional needs, the user wants the answer. For expressive categories, the user wants to shop. That distinction sounds small. It decides whether an AI interface becomes the marketplace or another failed experiment sitting alongside search, feed, and filters.

The tempting thesis is that AI agents become the new consumer front door and marketplaces get reduced to commodity fulfillment layers. That will be true in some categories. It will be false in others.

The AI Marketplace Matrix

The easiest mistake is to treat marketplaces as one category. A food-delivery marketplace, a fashion marketplace, a quick-commerce app, an eyewear retailer, a travel aggregator, a services marketplace, and a grocery platform may all look like consumer internet businesses but structurally they are different machines.

The useful map has two axes:

• Does the user want a definitive answer or an exploratory journey?

• Is the transaction mostly informational or physically constrained?

Each quadrant has a different AI endgame. In answer-seeking categories, AI can collapse the funnel. The agent asks the right questions, reasons through the options, and presents a small set of choices. The old browse/search interface loses power because the user did not want to browse in the first place.

In exploratory categories, AI cannot simply compress the journey into three recommendations. The act of shopping is part of the product. The user wants to compare, wander, refine taste, react visually, and feel agency. In these categories, AI has to enrich discovery without killing the pleasure of exploration.

In physically constrained categories, the interface matters, but the operating system matters more. Availability, density, ETA accuracy, substitutions, returns, trust, price, freshness, support, and fulfilment determine whether the transaction actually works. AI can change the front door. It cannot wish the delivery network, supply quality, or local operating memory into existence.

What Marketplaces Actually Are

A marketplace is four businesses hidden under one brand.

First is the demand engine: acquisition, activation, search, discovery, personalisation, conversion, cross-sell, support, and habit formation. This is the visible layer. It is also where consumer AI gets most of the attention.

Second is the supply engine: seller onboarding, merchant tooling, catalog ingestion, content generation, pricing, inventory quality, ad spend, availability, reliability, and retention. This layer is less visible, but it is often where liquidity is actually created.

Third is the market-making core: ranking, matching, recommendations, trust, fraud, logistics, pricing, demand shaping, movement planning, forecasting, and fulfillment orchestration. This is the compounding layer. It decides whether the marketplace is merely a directory or a machine that can create transactions.

Fourth is the operating system: customer support, finance, audit, warranty, hiring, vendor reconciliation, ERP/WMS integrations, warehouse workflows, store workflows, analytics, and developer productivity. This is where AI releases margin, but also where the ugly truth of enterprise systems shows up.

AI touches all four. But it does not touch them evenly. On the demand side, AI changes the interface from browse to intent. On the supply side, AI changes the bottleneck from seller acquisition to seller readiness. In the market-making core, AI changes matching from static rules to adaptive models. In the operating system, AI changes marketplaces from human-process businesses into agent-supervised execution systems.

The best companies are asking which layer controls the transaction, and where can AI move that control point?

Demand: Chat Is Not The Universal Commerce UI

The first wave of commerce chatbots taught the wrong lesson. Many did not fail because the model was too weak. They failed because chat was the wrong interface for the category.

Chat works when the user wants a definitive answer. Find me a flight under this budget. Reorder this medicine. Which policy covers this claim? Which product should I buy for this use case? The user has a job and wants resolution. Chat is weaker when the user enjoys the act of browsing. If someone is buying a saree, a t-shirt, a frame, a lamp, a lipstick, or a piece of decor, three perfect recommendations may still be unsatisfying. The point is not only to complete the transaction. The point is to explore taste. That means the future of AI commerce is not simply conversational shopping. It is interface-market fit.

AI discovery matters most where the user is still forming intent. Fashion, beauty, eyewear, home decor, travel, and other low-repeat or high-variety categories can benefit meaningfully from AI-assisted discovery. The consumer does not merely want a faster search box. They need help forming the choice without losing the pleasure of choosing. A food-delivery marketplace is a useful counterexample. In high-frequency, logistics-heavy categories, the moat remains availability, reliability, density, pricing, substitutions, support, and fulfillment. The interface can change; the operating system still has to work.

Voice is part of the same story. In markets where the interface itself has been the bottleneck, voice can unlock users who were never fully served by text-first commerce. But voice is not one product. In production, it often becomes a system of choices: ASR, reasoning, TTS, latency, language coverage, tool use, escalation, and cost. Cascaded voice systems win where control and composability matter. Speech-to-speech models will matter where emotion, latency, and natural turn-taking are the product.

So the AI threat is not one generic shopping agent. It is category-specific. In some markets, the AI front door becomes the marketplace. In others, it becomes a thin layer on top of the real moat.

Supply: The Catalog Becomes The API

Most marketplace conversations over-index on demand because consumers are visible. But many of the hardest marketplace bottlenecks are supply-side. Can the seller onboard without a human ops team? Can messy catalog files become structured SKUs Can bad images be repaired or rejected? Can the platform infer attributes, categories, styles, variants, compliance requirements, and substitutions from incomplete seller data? Can the seller understand which products to push, which ads to run, which inventory to refresh, and which demand signals matter? Can a local merchant participate in a digital marketplace without becoming a systems integrator?

This is where AI becomes less glamorous and more important. A consumer marketplace is only as good as the structured representation of its supply. If the catalog is wrong, every downstream model is wrong: search, recommendations, pricing, availability, ad targeting, personalisation, and support. The bottleneck is not just extraction. It is normalisation, enrichment, deduplication, image quality, SKU mapping, category inference, attribute completion, compliance, and continuous refresh.

The catalog is becoming the API for supply. This also changes seller retention. A seller agent that improves catalog quality, recommends content, manages ad budgets, explains demand, and reduces operational work becomes more than a tool. It becomes the seller’s interface to the marketplace. If the consumer front door is where AI gets attention, the seller cockpit may be where value quietly compounds.

Core: The Recommender Stack Is Being Rewritten

The strongest marketplaces have never been mere aggregators. They are market-making systems. They decide what to show, when to show it, which supply to allocate, how to price, how to forecast demand, how to place inventory, how to route jobs, how to detect fraud, when to subsidize, when to intervene, and how to trade off short-term conversion against long-term liquidity. AI makes that layer deeper. The architectural shift is from cascaded recommenders to generative recommenders.

The old stack split the work into retrieval, ranking, re-ranking, business rules, and objective-specific models. The emerging stack tries to collapse more of that into larger transformer-based systems, often using semantic representations of products, users, and actions as the vocabulary of recommendation. Instead of many small models passing candidates down a pipeline, the direction of travel is toward fewer, larger systems that can reason across products, users, objectives, and context. This does not mean every marketplace model becomes a generic API call. The model is not the moat. The production loop is.

The moat is the combination of proprietary marketplace data, feedback loops, local operating constraints, transaction outcomes, and the ability to measure whether a model actually improved conversion, availability, ROAS, CSAT, attach rate, fraud loss, returns, inventory turns, or fulfillment quality. The more physical the marketplace, the more important this becomes.

In a pure digital marketplace, AI can crawl, compare, recommend, and transact with fewer real-world dependencies. In physical marketplaces, every recommendation has an operational shadow. Inventory may be wrong. Delivery may fail. A product may be unavailable. A return may be expensive. A vendor may be unreliable. A store may be crowded. A customer may churn after one bad experience.

The agent can shape demand. The marketplace still has to fulfill it.

Operations: Marketplaces Become Agent-Supervised Companies

Marketplaces contain a shocking amount of internal bureaucracy. Support teams jump across old dashboards. Finance teams reconcile vendor systems, WMS, ERP, and revenue recognition. Audit teams review invoices and exception patterns. Warranty teams decide claims. Hiring teams screen applicants. Analysts generate reports. Developers write and maintain internal tools. Ops teams move information between systems that were never designed to talk to each other.

This is the labor-heavy underside of consumer internet. AI attacks it directly. Support is the obvious first wave because the workflow is measurable: cost per interaction, resolution rate, CSAT, response time, escalation rate. When token costs fall and quality improves, the ROI becomes obvious.

But support is only the beginning. Warranty decisions, fraud checks, invoice review, candidate screening, ledger reconciliation, store operations, catalog refresh, ad creative generation, campaign analysis, and developer productivity all sit inside the same pattern: high-volume knowledge work trapped between legacy systems and human queues.

The more interesting internal use case is not coding productivity by itself. It is turning non-engineering operators into builders. When business owners, category managers, analysts, ops leaders, and P&L owners can build their own tools, the organization changes shape. The gain is not only individual productivity. It is lower coordination cost. A single person can understand the KPI, inspect the data, build the workflow, test the fix, and own the outcome. That collapses the handoff between business, product, data, and engineering.

In marketplace operations, this matters enormously. The person closest to the exception often knows what needs to change but historically lacked the technical leverage to change it. AI gives that operator a builder persona. The old operating model was a human process plus software. The new one is a software process plus human escalation.

The Disruption Vectors

Five mechanics are doing the actual work.

Vector 1: The interface splits by category behavior

The app is no longer the only consumer interface. A user can speak, describe, upload an image, show a preference, ask for a bundle, or delegate a task. The marketplace that understands intent best gets the first shot at the transaction.

But the interface does not always become chat. In answer-seeking categories, the agent can collapse the journey. In taste-led categories, the agent has to enrich discovery without killing the joy of exploration. In logistics-led categories, the best AI may never be visible to the consumer at all.

The risk to incumbents is not that a model knows more about commerce. It is that the model becomes the user’s default interpreter of intent in the categories where intent is the bottleneck.

Vector 2: Supply becomes machine-readable

AI does not work on “supply.” It works on structured representations of supply. A marketplace with clean catalogs, rich attributes, good images, reliable availability, historical interaction data, and transaction feedback has a compounding advantage. A marketplace with messy PDFs, poor images, inconsistent attributes, duplicate SKUs, and vendor-specific formats has an AI tax.

This is why catalog automation is strategically underrated. The company that makes supply machine-readable can sit upstream of search, recommendations, ads, personalization, seller success, and marketplace liquidity.

Vector 3: Model strategy splits by value, volume, and maturity

The winning model stack is not simply frontier or open source. It is a router. Frontier models dominate exploration, long-tail use cases, multimodal reasoning, complex review, and high-value workflows where quality matters more than unit cost. Open-weight or self-hosted models absorb high-volume, mature, accuracy-critical workloads once the task is well understood and the economics matter.

This is why the answer to “open or closed?” depends on what you are measuring.

The count of use cases may be frontier-heavy because experimentation starts on the most capable models. Request volume may be open-heavy because mature, high-scale workflows eventually move toward cheaper or fine-tuned systems. Economic spend can sit between the two, depending on where quality still commands a premium.

That means AI infrastructure decisions will increasingly be made at the workload level, not the company level. The model is not the strategy. Allocation is the strategy.

Vector 4: ROI governance replaces AI theater

The first wave of AI adoption rewarded experimentation. The second wave will reward accounting Marketplaces are starting to ask a harder question: which AI use cases create value, which merely create activity, and which looked good only because nobody was measuring the full cost?

The cleanest ROI cases are still traditional business cases with AI inside them. ROAS improves. Cost per support interaction falls. Activation rises. Conversion improves. Fraud loss decreases. Hiring throughput increases. Warranty turnaround falls. Developer velocity improves. Team size ratios change.

The hardest cases are internal productivity tools where benefit is real but attribution is fuzzy. Coding agents, analyst copilots, and general employee automations can feel obviously useful, yet still require better measurement: shipping velocity, maintenance burden, discoverability, rework, quality, and total cost.

This is where many companies will discover that AI spend is not the same as token spend. The cost base includes GPU infrastructure, ranking models, video processing, voice, multimodal pipelines, vendor contracts, cloud commitments, self-hosting, observability, evals, safety, and people needed to maintain the system.

The CFO will not care that the demo was impressive. They will care whether the workflow creates revenue, reduces cost, reduces risk, or changes the operating leverage of the company.

Vector 5: Physical marketplaces are defended by operational memory

The most important distinction in marketplace AI is physical coupling. This is why a food-delivery marketplace behaves so differently from a low-repeat fashion or travel marketplace. In one category, AI may reshape discovery. In the other, discovery is only one small part of the transaction.

AI can reason across information markets quickly. It can compare digital goods, recommend products, summarize options, generate content, and automate coordination. But in physical-world marketplaces, the transaction is only partly informational.

The rest is local density, trust, vendor reliability, logistics, returns, inventory, payments, compliance, support, and operating history. That does not make incumbents invulnerable. It makes the battle more specific.

The AI-native challenger has to decide whether it is attacking the front door, the supply cockpit, the market-making core, or the operating system. Each requires a different product, GTM motion, data asset, and moat.

The incumbent has to decide whether it is using AI as a feature, a cost program, or a new control point. Both extremes are wrong. AI will not kill marketplaces wholesale. But marketplaces that are only traffic aggregators will be exposed.

Efficiency Tokens And Value Tokens

The most useful way to think about AI spend in marketplaces is to split it into two pools. Efficiency tokens reduce cost. Value tokens create revenue.

Efficiency tokens show up in support, code, QA, analytics, catalog processing, invoice review, hiring screens, warranty triage, and internal workflows. These use cases matter because they release margin. But over time, they become cost-sensitive. Once a workflow is stable, high-volume, and measurable, the pressure to move it to cheaper models, open-weight systems, caching, fine-tunes, or self-hosting increases.

Value tokens show up in activation, conversion, personalization, recommendations, ad performance, seller ad budget growth, demand shaping, cross-sell, category expansion, and new user unlocks. These use cases justify frontier spend longer because the upside is tied to GMV, take rate, frequency, or ad revenue.

That split explains why AI budgets can look irrational from the outside. A company may aggressively optimize token cost in support while spending heavily on multimodal discovery. It may self-host one workload and pay frontier prices for another. It may use open models for volume and closed models for judgment. It may appear model-agnostic because the real strategy is not allegiance to a model vendor. It is margin-aware routing across workflows.

This is also where the marketplace AI stack becomes investable. Every large marketplace will need cost attribution, evals, routing, quality measurement, prompt/version control, data governance, model monitoring, and business-outcome linkage across dozens or hundreds of AI workflows.

Where Potential New Company Formation Opens Up

If AI compresses parts of the old marketplace model, where do new companies get built?

We think the opportunity is in the infrastructure and workflow layers where models meet marketplace reality: messy supply, multilingual demand, physical fulfillment, ROI measurement, and high-volume operations.

The common thread is not “AI for commerce”. It is workflow ownership. If you are demand-side, prove activation, conversion, frequency, or GMV lift. If you are supply-side, prove onboarding speed, catalog quality, seller revenue, ad efficiency, or supply liquidity.

If you are core marketplace infrastructure, prove matching quality, attach rate, fraud reduction, forecasting accuracy, inventory turns, or contribution margin. If you are on the operations side, prove cost per workflow, turnaround time, accuracy, escalation rate, auditability, and labor leverage.

The startup opportunity is not evenly distributed. The most glamorous layer may not be the most defensible. The most defensible layer may not be the easiest wedge. And the biggest spend may not produce the best venture outcome if it is only cost compression with no control point.

India AI News Roundup

The most important AI developments shaping India’s market structure this week.

Google picks 20 AI startups for India Accelerator 2026 cohort

HCLTech lands a $1.1B AI-led enterprise deal, reportedly with Mercedes-Benz

Dell pushes sovereign AI storage in India with PowerStore Elite

ESDS launches Swaraj Cloud, an India-operated sovereign AI cloud platform

Nurix AI acquires Verloop.io to strengthen enterprise conversational AI platform

Startup Signals

Early AI companies from India we are watching because they point to where the market may be forming.

Observal: The system of record for AI coding agents

Observal is an open-source registry and analytics platform for AI coding agents. It tracks skills, MCP servers, hooks, prompts, sandboxes, components, agents, session traces, and usage across tools like Claude Code, Cursor, Gemini CLI, VS Code, and GitHub Copilot.

The insight is simple: AI coding agents are becoming part of the software supply chain, but most teams do not yet have a clean inventory of what agents are using, what components they depend on, or which agent workflows are actually helping. If engineering work increasingly happens through agents, then agent components need versioning, observability, and governance.

Why it matters: the next developer platform may not just manage code. It may manage the agents, tools, prompts, and reusable skills that produce the code.

Shadow Labs: Agents that execute while the call is still happening

Shadow Labs is building an AI coworker for sales and customer-facing teams, moving beyond the now-crowded “meeting summary” layer into real-time execution. The sharper wedge: the call itself becomes the prompt, and background agents can pull context, update systems, draft follow-ups, and move work forward before the conversation ends.

Why it matters: enterprise AI is shifting from passive copilots to action-taking operators. The meeting stack is a natural insertion point because it sits closest to customer intent, but the durable company here is not the best transcriber. It is the one that owns the workflow after intent is captured.

Dreabee: Creator intelligence for influencer marketing

Dreabee is building a creator intelligence platform for brands and agencies, positioning itself as a more data-driven decision layer for influencer marketing. The wedge is simple: as creator spend grows, teams need better ways to discover creators, underwrite audience quality, benchmark performance, and allocate campaign budgets.

Why it matters: influencer marketing is still too taste-led and relationship-led for the amount of money now flowing through it. The opportunity is to turn creator selection into something closer to a research, underwriting, and measurement workflow. If AI makes supply discovery cheaper, the scarce layer becomes trust, signal quality, and decisioning.