Theses

The bar at seed doesn't hold constant across categories. What has to be true for an AI company looks nothing like what has to be true for a fintech company, and treating every pitch against the same checklist is how good sectoral judgment turns into generic advice. Different categories fail for different reasons, so the one thing that has to be true before anything else, differs too.

This isn't a checklist. It's the one question I actually ask myself for each category, before anything else about the pitch matters.

AI. Is this AI actually doing the job end to end, live, with a real data or distribution moat forming from doing it, not a wrapper on someone else's model? Most AI pitches sound like this is already true. Very few of them actually are, once you ask what breaks the moment the underlying model changes.

Fintech. Is there a real, live financial relationship already proving itself, a loan repaid, a policy attached, a transaction cleared, or is this still a plan waiting on distribution? Distribution is the whole game in a regulated market, and a deck can describe a partnership that hasn't actually been signed just as easily as one that has.

Consumer. Would people pay for this and come back on their own, without a subsidy or a discount holding them there? Retention that survives without one is rare enough that its absence is usually the real story, however good the growth chart looks on its own.

Commerce. Does this make money on a real transaction without subsidizing it, and would it survive losing its first anchor customer or platform? Commerce businesses tend to look durable long before they actually are, and the one relationship that made the first year work is exactly what a bad quarter takes away first.

Different categories, different questions, and none of them are answered by a good story about the market. The pitch that can't answer its own question honestly is the one I pass on, whatever else is in the deck.

The old assumption behind a seed round was simple. It bought eighteen months of runway, and runway meant payroll, enough people to reach product-market fit before the money ran out. Burn was one number, and it stayed roughly fixed until the company hired again.

I keep seeing smaller teams building bigger things than that math would have predicted, two or three people now shipping and iterating on what used to need a team of ten. That alone would just mean smaller rounds buy the same runway. It doesn't stop there.

Payroll shrinking is only half of what changed. The half nobody's pricing into the round yet is inference. Every one of these smaller teams is running on a model, and that spend doesn't behave like payroll. Payroll is fixed until the company hires again. Token spend scales with usage, so it grows exactly when the product starts working, which is the one moment nobody wants to discover their burn model was wrong.

A round modeled on one flat number for eighteen months misprices exactly the moment that matters. It needs two lines instead, one that's shrinking, headcount, and one that grows with traction, inference, and collapsing them into a single burn figure hides the one that's about to move.

The check should get smaller for the same milestone, that part is real. But the day a team finally hits product-market fit is the day the second number starts outweighing the first, and a fund still underwriting last decade's cost structure against this decade's product is going to misprice the round in exactly the direction that hurts.

Every AI coding tool this year competes on the same axis: how fast it can generate code. That axis is commoditizing in real time. When several different model providers can generate a working function from the same prompt, generation speed stops being a moat. The interesting question isn't who writes the code fastest anymore, it's who you can trust to ship what got written.

The bottleneck moved from writing to verifying. Engineering teams are already reporting that code review, not code generation, is now their biggest time sink. Nobody sized their review process, their test coverage, or their deployment gates for a world where a single engineer can produce ten times the code volume in a day. That gap is where the real product sits, not in the generation layer.

AI-native engineering platforms worth backing aren't building a faster autocomplete. They're building the trust infrastructure underneath it: automated test generation that actually exercises edge cases, agentic code review that catches what a tired human reviewer misses, deployment gates that roll back before a bad change reaches a real user. The product is judgment, encoded as tooling, not raw generation.

This isn't a new problem dressed up in AI language. Every step change in developer productivity, from compilers to CI/CD to infrastructure-as-code, moved the bottleneck downstream rather than removing it. AI just moved it further and faster than the last few shifts did, and most tooling hasn't caught up.

I'm looking for founders who treat verification as the product, not an afterthought bolted onto a generation tool, and who have real opinions about where automation should stop and a human should sign off. Early stage, conviction on judgment about the failure modes of AI-written code, not on whichever model API happens to be fastest this quarter.

Every thesis on this list has had a filter running quietly underneath it that I've never named directly. It's time to. This one is about who, not what.

The pattern: operators who spent real years inside a company at genuine consumer or product scale, Flipkart, Swiggy, Razorpay, Goibibo, the names that keep showing up across this whole list, before starting something of their own. Not first-time builders straight out of college, and not serial founders chasing whatever category is loudest this quarter.

This isn't a credentialism filter. A polished resume alone tells me nothing. What it actually predicts is scar tissue: this person has already made expensive mistakes on someone else's balance sheet, has watched what actually breaks at scale, retention, trust, operations, not just a launch that looked good in a deck, and has the pattern recognition that only comes from being wrong at a size that mattered.

The specific question I'm asking isn't "is this person smart," it's "has this person owned something real under real constraint." A P&L, a compliance mess, a retention number that had to move or the business didn't survive. That's a different kind of judgment than shipping a good feature inside a well-run machine someone else built.

This is why "early stage, conviction on team and problem over traction" keeps recurring across this whole list. The team is the underwriting. I'm looking for the specific kind of operator scar tissue that predicts someone can do this again, on their own, from scratch, under real constraints, not a good story about why now.

The last thesis was about practicing before the real thing. This one is about the real thing. An agent that books the flight, moves the money, files the return, is a fundamentally different bet than one that role-plays the conversation first. Most of what gets called an "agent" right now is a chatbot with function-calling bolted on. That's not the same claim.

The tell is reversibility. A simulation's whole value is that the mistake is free. An acting agent's mistake is not. That single fact changes everything about what has to be built underneath the model: not better prompts, but a trust and liability structure for what happens when it's wrong, because at scale, it will be.

This is the professional-services liability argument again, generalized past law and audit. Someone has to be accountable when an autonomous action goes wrong, whether that action is a refund, a trade, a contract signature, or a support ticket resolved the wrong way. Founders who treat that accountability as a legal afterthought instead of the actual product are building on top of a hole they haven't noticed yet.

The domains worth building in aren't the highest-stakes ones or the lowest, they're the ones with bounded, recoverable consequences: real enough that a human doesn't want to do it by hand, contained enough that a mistake doesn't end a company or a life. Moving a person's life savings unsupervised is the wrong place to start. Rebooking a flight when one leg gets cancelled is closer to right.

What actually makes this buildable is staged autonomy, not full autonomy from day one. The agent proposes, a human approves, and only after a track record earns it does the loop tighten. Audit trails and insurance get designed in from the first version, not retrofitted after the first expensive mistake. I'm looking for founders building that trust ladder as the core product, in a domain they picked because the stakes are real but bounded, not founders who wrapped function-calling around a chat model and are calling it agentic.

Simulation has become a label people stick on any product with a roleplay prompt. A chatbot pretending to be an angry customer is not a simulation, any more than a script read by an actor is a flight simulator. The distinction isn't semantic, it's the whole thesis.

A real simulation models a distribution, not a single scripted path. It has to produce the edge cases, the adversarial responses, the ways real systems actually fail, not just the polite average case a demo is built to show off. The fidelity of the modeled environment is the product. Everything else, the chat interface, the dashboard, is a wrapper around that core bet.

The use case changes, the underlying bet doesn't. Sales and negotiation practice, support teams rehearsing a hostile customer before a product ships, clinical training, enterprise software absorbing simulated load and edge-case users pre-launch, autonomous systems logging millions of miles in a simulated world before one mile on a real road. Different domains, same question: does practicing inside this system actually transfer to the real one, or does it just feel like practice?

Willingness to pay tracks the cost of the mistake being avoided, not how novel the AI feels. A bad sales call costs a deal. A bad clinical call costs a life. A bad autonomous driving decision costs a person. The categories worth building in are the ones where getting real-world reps is slow, expensive, or dangerous, because that's exactly where a high-fidelity substitute is worth real money, not a nice-to-have.

I'm looking for founders treating the simulated environment itself as the defensible core, the modeling work, the edge-case coverage, the domain expertise baked into what the simulation actually reproduces, not founders wrapping a roleplay prompt around a general model and calling it a simulator. The wrapper is easy. The fidelity is the moat.

Quick commerce isn't one bet, it's a 2x2. Consumer or B2B, generic or vertical, and most of the conversation only ever happens in one quadrant. That's a problem, because the quadrant everyone talks about is the one I'd least want to write a check into right now.

Generic consumer quick commerce is Zepto, Blinkit, and Swiggy Instamart fighting a capital war over dark-store density, not a product contest. The demand itself was never the risk, Hari Menon said as much stepping down as BigBasket's CEO. "If you'd asked 100 people back then whether they wanted groceries in 10 mins, most of them would've said no. But the moment we gave it to our customers, they lapped it up." The playbook is settled, whoever can absorb the most burn on warehouses and delivery fleet density wins the metro, and the metros are already spoken for. This is a logistics and balance-sheet game now, not an early-stage one. Dunzo's arc is the cautionary tale, not the opportunity.

There's a second reason I'd stay out of that quadrant, and it isn't capital intensity. The ten-minute promise ran on riders weaving through traffic under a hard clock, and in January 2026 the government made every major platform drop the ten-minute claim entirely after road-safety concerns and gig-worker protests over Christmas and New Year's. That's not a one-off headline, it's what happens when a growth model's core promise depends on people breaking traffic rules to hit a number. A regulator has already stepped in once. I don't expect Indian metros to absorb this pace of delivery density indefinitely without a real cost to road safety and quality of life, and generic consumer quick commerce is the quadrant carrying that risk, vertical B2B mostly doesn't run on a public speed promise at all.

Vertical consumer quick commerce only clears the bar where the urgency is real, not manufactured. Medicine is the standout case, a genuine, non-discretionary need where ten minutes matters and the customer will pay for it every time. Home services is proving out as the second. Urban Company already holds 65 percent of that market on ₹1,556 crore in revenue. Snabbit dispatches over 40,000 jobs a day, Pronto's bookings are up eighteen-fold in a year on a ten-to-fifteen-minute promise, and both are pulling in real capital because the urgency is time scarcity, not marketing. Instant flowers, instant gadgets, instant anything-a-normal-person-can-wait-a-day-for is marketing dressed up as need, and it burns the same cash as the generic players without their density advantage.

The quadrant worth actually spending time in is B2B. The same dark-store and hyperlocal delivery rails the generic giants already paid to build now exist as infrastructure someone else can build on top of, serving a shopkeeper, restaurant, or pharmacy whose reorder behavior is operational necessity, not an impulse a push notification triggered. That's a fundamentally better retention story, and it doesn't require the same marketing spend to manufacture urgency that isn't there.

Inside B2B, vertical beats generic again. A founder who owns fast replenishment for one category, restaurant ingredients, pharmacy stock, hardware for contractors, understands that buyer's order cycle and margin structure in a way a horizontal "quick commerce for every business" platform never will. I'm looking for founders building narrow, B2B-first, riding infrastructure that already exists instead of re-fighting the dark-store capital war the generic consumer players have already decided. Same capital-efficiency lens as the rest of this list, applied to logistics instead of software.

Every other lane on this list has a global counterpart, so this one should too. It doesn't get one as easily, and I want to be honest about why. Consumer applications built in India, for the world is the hardest pairing here, not the same pattern with the labels swapped.

AI for the world and professional services for the world both had proof points to point to. Zoho, Freshworks, Postman, Chargebee won because B2B buying is spec-driven: a feature comparison and a price don't care where the founder grew up. Consumer doesn't work that way. A user doesn't evaluate a product, they feel something about it in the first five seconds, and that feeling is built from a cultural fluency no engineering discipline substitutes for.

Cost and execution speed still transfer, same as the other lanes. Taste and distribution instinct do not. A founder who has never lived inside American or European consumer culture is building blind on the parts that actually decide whether a consumer product spreads, no matter how sharp the team is on everything else. This is where "built in India does not mean designed for India" runs out of runway, because the honest version now is: designed for India doesn't transfer to designed for anywhere else either.

The categories where this has any real shot are the ones where the underlying behavior is closer to universal than culturally coded. Utility, mechanics-driven games, infrastructure that happens to sit in a consumer-facing wrapper, not social, dating, fashion, or food, where the product basically is the culture. Truecaller is the closest thing to a proof point, and even that leaned on identity infrastructure more than on taste. That thinness of examples is itself the signal.

What this actually needs is a co-founder or early exec who has lived and consumed inside the target market, not a remote growth hire brought on after launch. Cost advantage can fund that person. It cannot replace them. I'm looking for founders who pick a narrow, behavior-driven wedge on purpose and pair it with someone who has the cultural fluency the founder doesn't, rather than founders assuming the Zoho playbook just carries over because it worked once, in a completely different kind of business.

This one predates AI and will outlast it. Consumer applications for Indians is its own lane, separate from AI for Indians, because most of what breaks here has nothing to do with model quality. It's the same mistake global playbooks keep making, treating India as a market to enter rather than a product to design from scratch.

Monetization is the first place the playbook fails. Subscription pricing tuned for a US willingness to pay doesn't survive contact with a price-sensitive, cash-flow-aware user. The apps that actually won here monetized around the edges of engagement instead, commerce take-rates, lending, ads, a small cut on a transaction the user was already going to make. Dream11, CRED, Meesho didn't out-subscribe anyone, they built revenue into behavior that was already happening. AstroTalk and PocketFM prove the same point from a different angle, pay-per-minute consultations and pay-to-unlock episodes, not a flat monthly fee. AstroTalk went from 400,000 users in 2020 to 8 million by 2024 and turned a real profit doing it, PocketFM's revenue jumped 68 percent to over 1,700 crore in a single year. Neither is a tech-elite product, astrology consultations and Hindi audio drama reach further into India than any English-language subscription app ever has.

Even the categories that look like classic subscriptions are shifting, and the shift itself is the signal. Indians weren't paying for cloud storage a few years ago, now iCloud counts India as its fastest-growing major market, and Google routes its own storage plans through Airtel's billing instead of asking for a card number directly. India's digital subscription revenue grew 60 percent in a single year, per FICCI-EY, crossing sixteen thousand crore rupees. None of it happened by importing a US subscription card flow, it happened by riding a bill Indians already trust, or a payment as small and specific as one consultation.

Distribution doesn't run through the app store either. Discovery in India is social, not algorithmic search. A reseller in a WhatsApp group, a family member's recommendation, a regional-language creator on Moj or Josh carries more weight than a category ranking. Vernacular content apps like Kuku FM and Pratilipi grew by going where the audience already was, not by winning an English-language app store category no one in their user base was browsing.

Retention follows the same logic. Habit loops built for a single user on a personal device assume a kind of privacy and consistency that doesn't hold for a shared family phone. The products that stuck were the ones designed for interruption, for someone else picking up the phone mid-session, for trust that's earned through a community or a reseller network rather than a streak counter.

I'm looking for founders who start from how Indians actually discover, pay for, and stay with a product, not from a global consumer template with the currency symbol swapped. The constraint is the product brief, not an afterthought once growth stalls. Early stage, same conviction on team and problem as the rest of this list.

The localised version of the last thesis is a different business, not a smaller one. India's professional services market for small businesses is enormous, fragmented, and runs almost entirely through a network of local CAs, company secretaries, and compliance agents most people have never heard of outside their own city.

Global mid-market buyers self-serve. Most Indian MSMEs don't, and won't, no matter how good the interface is. They pay for a relationship they trust, usually a CA who's filed their returns for a decade, not a dashboard. Any AI-native firm here has to keep that trust layer, even while replacing the manual work behind it.

The actual opportunity is the gap underneath the trusted relationship. GST filings, MCA compliance, labour law paperwork are rule-based, repetitive, and exactly the kind of work AI does well. AI doing the compliance work, a local human doing the sign-off and the relationship, is the right shape here, not a pure self-serve product.

I've watched this exact split work, not hypothetically. My brother runs a CA firm, and he's automated the auxiliary work himself, bookkeeping and data entry, GST and tax filing prep, reconciliation, client reporting, invoicing and payment collection. None of it touched the relationship. He's still the person a client calls, still the one who signs off. The work that moved is precisely the rule-based layer this thesis is about, not the trust layer sitting on top of it.

Most small businesses in India don't avoid compliance because they don't understand it. They avoid it because a CA relationship is expensive and slow relative to what a tiny business can pay. A firm that cuts that cost by five to ten times doesn't just take share from existing CAs, it brings in businesses that were informal by default.

I'm looking for founders who've actually sat inside this ecosystem, who understand why a local agent is trusted and a SaaS login isn't, and are rebuilding the economics around AI without assuming the buyer behaves like a global self-serve customer. Same lane as AI for Indians, applied to compliance instead of consumer products.

Professional services is the next place this plays out. Law, audit, tax, and consulting firms sell judgment by the hour, and every AI copilot pointed at them so far just makes the existing associate cheaper to run, not the firm structurally different.

Incumbents can't rebuild themselves around AI, because the billable hour is the business. A partner who halves the hours a matter takes has just cut their own revenue. That's an incentive gap, not a technology gap, and no copilot sold into these firms is going to close it.

There's no need to invent a new delivery model to get around that trap either. India already built one, decades of BPO, and more recently GCCs, executing professional-services work for global clients at real scale and real quality. That story has already been won, law firms, audit firms, and consultancies already route research, drafting, and first-pass review through Indian teams today. The infrastructure and the trust are already there.

The wedge is re-pricing that existing delivery layer, not replacing it. A BPO or GCC operator doesn't lose personal income by cutting hours the way a partner does, cutting hours is how they compete on margin, not how they lose it. A team that equips these providers with AI systems for the actual work, drafting, research, review, and helps them price the output instead of the hour, wins on both sides, the provider captures more margin per engagement, the global client pays for a result instead of a timesheet.

The constraint that never goes away is liability, and it doesn't disappear just because the delivery model already exists. Professional services carry malpractice risk, regulatory exposure, and a client who needs someone accountable when it's wrong. The winning version of this tooling is built around that from day one, human sign-off, credentialed review, insurance, layered into the provider's existing workflow, not bolted on after the fact.

I'm looking for founders who equip India's already-proven offshore and GCC delivery layer, not founders trying to build a brand-new client-facing firm to compete with Big Law or the Big Four head-on. The wedge is usually narrow, one workflow, re-priced from hours to output, before the provider expands into the next one.

The second lane from the earlier thesis deserves the same treatment. AI for the world, built in India is not a talent-arbitrage story dressed up in a new label. It's a different claim entirely.

The old model rented out engineers to someone else's roadmap. This one is founders who own the product, the brand, and the customer relationship, and simply chose to build the company from India. That distinction is the whole thesis. A services mindset optimizes for utilization. A product company optimizes for the customer's next best experience.

The advantage people reach for first is cost, and it's real, but it's not the interesting part. Cost buys runway. What you do with that extra runway is what separates a discount competitor from a category leader: more iterations per dollar, a team that can afford to rebuild something twice before a well-funded rival ships once.

The teams worth backing already have a decade of building for consumer scale inside companies like Flipkart, Swiggy, and Razorpay, not just a computer science degree. That maturity has to travel outward, though. Built in India does not mean designed for India. Once the customer is global, the UX, support hours, pricing psychology, and payment rails all have to fit that customer, not the founder's own defaults.

Zoho, Freshworks, Postman, Chargebee already proved the pre-AI version of this works. I'm looking for the AI-era version of the same founders: technically deep, globally fluent from day one, and treating the India base as an engine for speed and craft, not an excuse to ship less of it.

The first lane from the last thesis needs its own space. AI for Indians is not one product decision, it's four or five, and most founders only make one of them.

Everyone gets language first. Support Hindi, add a few regional languages, call it done. That's not what real usage looks like. People don't speak in clean Hindi or clean Tamil, they code-switch mid-sentence, drop in English nouns, and shift dialect depending on who they're talking to. A model trained on textbook translations breaks the moment it meets a real voice note.

Language is the visible constraint. Bandwidth and device class are the invisible ones. Most of the next wave of users are on a shared, low-RAM Android phone with inconsistent 4G, not a laptop on fiber. Every design decision, from model size to how much you round-trip to a server, has to assume that from day one, not retrofit it after a demo works on a founder's iPhone.

Trust here isn't a UI pattern, it's a track record. People pay for what a neighbour or a WhatsApp forward vouches for, not what a landing page claims. And the price point isn't a discount off the US price, it's a different unit of value entirely, pay per use, pay per minute, pay inside a bundle someone already trusts, not a monthly subscription card.

This runs deeper than any single product decision, localization here isn't a checklist, it's built around behavior, and language is only the most visible piece of it. Flipkart proved this first with cash on delivery, a response to how little Indians trusted paying online in the 2010s, and Amazon fell in line rather than compete without it. Ola did the same with ride verification, adding a per-ride OTP in 2017 after India's low-trust environment made pickup fraud a real problem, Uber didn't add an equivalent PIN in India until three years later, in 2020. Global products don't just get localized here, Indian behavior has repeatedly forced them to change on India's terms.

India has already produced the proof, Jio, UPI, Meesho didn't win by being cheaper versions of a global product. They won by being built for constraints global products never had to solve. That's the bar for AI here too. I'm looking for founders who start from language, device, trust, and price as the design brief, not as a localization checklist bolted on after product-market fit in English.

Most conversations about AI in India collapse into one bet. They shouldn't. There are two, and they call for different products, different teams, and different economics.

The first is AI for Indians. Language is the wall here, not intelligence. A model that reasons well in English and stumbles on a voice note in Bhojpuri hasn't solved the problem for the next 500 million users online. Trust, price, and bandwidth matter as much as accuracy. The winning products here won't look like a chat app with a translation layer bolted on. They'll be built assuming low bandwidth, shared devices, and a first internet experience that is voice and vernacular, not text and English.

The second is AI for the world, built in India. This is not the old outsourcing story with a new label. Indian teams have shipped globally-used products before, in payments, developer infra, and SaaS. The same depth of engineering, now applied to model work and applied AI, at a cost structure competitors can't match, is a real edge. The founders worth backing here aren't running an offshore delivery arm for someone else's roadmap. They own the product and the customer, and India is where they chose to build, not where they were forced to.

What ties the two together is capital efficiency. Small teams, tight loops between building and shipping, and a default to doing more with less compute and less headcount. That instinct travels well in both directions: serving a price-sensitive Indian user and out-executing a well-funded competitor come from the same discipline.

This is the lens I'm underwriting through: founders solving a real problem in one of these two lanes, with product decisions that follow from who they're actually building for, not from what's trending. Early stage, conviction on team and problem over traction.