AI in Commerce: Where the Real Moats Get Built

Both vertical and horizontal marketplaces are built around aggregating demand for a fragmented supply. In reality, these models tend to barely break even on fulfillment. Indicative numbers of a marketplace without any private labels or ads revenue:

The 10% CM1 is structural. A marketplace that barely breaks even on fulfillment with zero marketing has two paths to profitability: private labels and advertising inventory sold to brands. Both degrade the consumer experience. Private labels & ads both push products users did not come looking for. Retention can suffer in such situations.

Acquisition spend rises (LTV/CAC primer) to compensate and hence the unit economics never heal.

Most commerce companies are already inside this trap when they start thinking about AI, and the instinct to use AI to make the broken model more efficient is what produces bad investments.

Where AI Is Being Deployed

Directly or indirectly, AI is being deployed into 3 key tasks:

Research: This is largely commoditized at this point, because plugging a web search module into an app is a weekend project and getting someone to search once is already a solved problem. The product collapses right after that first session, and AI search with no transaction ownership generates no post purchase data, so the model stays generic and the retention curve never improves. More importantly, as a platform you will never be able to benefit from the final purchase data flywheels.

Co-ordination: This layer has genuine near term value, especially in India where last & mid mile logistics still involves a lot of humans on phones. voice AI agents that can handle OTP sharing, pick up / delivery instructions, and exception management are genuinely useful. For now, they represent operational improvements which can be probably used to build structural moats over time, both for the business & also for the end customer.

Execution: The execution layer is moving fast, with browser automation, smarter upsell logic at checkout, and dynamic landing pages by cohort all becoming table stakes within the next couple of years. Almost each of these places have some early Indian incumbents (new cos or large cos adding it as a feature) raising venture $$.

Post-purchase is almost entirely untouched, and that is where I feel the real bet may lie.

From Order Tracking to Outcome Tracking

Most e commerce journeys end identically: order shipped, order delivered, review link, and a 10% discount coupon. The interesting opportunity is moving from order tracking to outcome tracking. Customers want an outcome: better skin, a nicer room, or a good trip.

Workflows can be built that deploy AI as a post purchase concierge natively on WhatsApp or inside the app. The AI knows what the user bought and why. It can answer how to use the item, handle refunds, suggest fixes, and time replenishment based on real usage data rather than fixed (human or automated) cadences. Every photo, chat, and review feeds back into a live user profile. The next session on the app is shaped by this dynamic data. This dynamic shaping is where real Lifetime Value (LTV) expansion comes from.

Having said that, there is still some opportunity to build vertical marketplaces that are AI native & there is some right to play for new entrants.

Which Vertical Marketplaces Are Actually Buildable?

Not every vertical has the same opportunity surface, and a few are worth thinking through in some detail.

Women’s fashion: If you look around yourself or do a cursory look on Helium10 / Junglescout, you will observe that there’s a great sales concentration among top 20 – 30 brands in men but this is not the case for women. This also means that most of the business for women’s fashion comes from the long tail. An AI that can surface the right long tail brand to the right user at the right moment has compounding value that a standard recommendation engine might not be able to produce. The right approach might be building your own inventory with user intelligence rather than just cataloging what already exists.

Beauty: The biggest structural problem in this category is retention. AI skin analyzers exist, but no real data moats are forming because personalization rarely goes beyond the surface. Services could be a viable wedge if someone is willing to go after it seriously, but the retention problem does not have a clean AI solution yet.

Travel: Tickets and hotels are largely commoditized, but activities at destination are highly fragmented with no dominant player, which makes it the most interesting surface to build on. Making users log in via Instagram to build a personality graph from their reels and posts and using that to surface better experiences is the kind of approach that might work here. What still needs to be figured out is another wedge that a platform could build on top of AI recommendations to really push a user to closing the transaction from their platform. Side note: do check out our Travel thesis.

Home & Décor: No vertical commerce platform has reached meaningful scale in India, even though the category is large. Buying intent is high, but friction comes from making decisions. Platforms have built an AR placement layer but combining it with AI design & décor suggestions might close the gap between wanting to redecorate and knowing what to buy. This feels like an AI native problem with no dominant player, and something genuinely large can be built here.

While new marketplaces are cute, there is something that we need to talk about.

The Cold Start Problem

Every AI-native commerce product hits the same wall early, because the model needs data to produce good outputs, good outputs drive engagement, and engagement generates data, but the cycle has to start somewhere and that starting point is where most products quietly die before the flywheel ever gets going.

Image: Gemini

There are 3 broad points of failure here

1. Recommendation lag: Most models retrain on daily or weekly cycles, while a shopper’s intent shifts within hours. For instance, in beauty and skincare, someone casually exploring anti ageing content on a Monday morning is a very different use case as opposed to the same person urgently looking for acne solutions before a Friday night event. If your personalization model only refreshed overnight, it is already out of sync with what they actually need right now.
   
2. New SKU burial: Marketplaces constantly add thousands of new products, but the model has no interaction data on fresh listings. The usual workaround is to manually boost anything tagged as ​“new,” which kills true personalization and turns the experience into simple editorial curation pretending to be AI led.
   
3. The new user dead zone: For a first time user, the AI native experience can feel identical to a non AI product for the first few sessions. This is exactly when the hook needs to kick in, but most teams wait for the model to ​“warm up,” which is not an option. Think of a beauty app that promises personalized routines: if the first 3/4 visits show generic bestsellers and broad categories, the user never experiences the intelligence you are selling, and they churn before the model ever really learns who they are.

The answer could be using non behavioral signals earlier: social graph data to infer taste before the first browse session, onboarding flows designed to extract meaningful preference signal upfront, and catalog metadata rich enough to give the model something substantive to work with on day zero. The products that figure this out will have structurally better retention curves than those waiting for the flywheel to self start.

Where One Could Place Their Bets

The sequence matters more than the specific technology choices, because research layer AI is already table stakes and post purchase AI is where the real moat might get built.

Own the outcome, not just the order: The brands that build durable businesses will be the ones tracking whether customers actually got what they came for. That means building an ongoing, AI assisted relationship that starts at landing page (and where they came from), goes to delivery, and eventually compounds over time. To do this well, you need to own the post-purchase surface directly instead of relying on a marketplace to provide it.

Pick verticals where long tail supply meets high intent demand: Women’s fashion, home and décor, and destination experiences are all categories where supply is so fragmented that discovery is genuinely broken. In these spaces, AI can do real, compounding work by surfacing the right thing to the right person. In contrast, categories where supply has already concentrated don’t offer the same surface area for this kind of differentiation. If I had to visualize the point I’m trying to make here, it would look like a 2×2 matrix like this:

Founders who get that sequencing right early will build data assets that take years to replicate, and the ones who spend 2025 building better search will find themselves in a feature war against platforms with ten times the data and ten times the engineering budget.

Commerce is being restructured, and and while I don’t have any answers, it would be interesting to see how all these hypotheses play out.