A widow in a village in Jharkhand is owed a pension. It exists; it is the law; her name belongs on a list. Claiming it means a form she cannot read, in a language she does not use, at an office open only while she is at work, guarded by a clerk who expects a cut. She is not poor because India lacks the money for her pension. She is poor because the institution standing between her and it does not function.
Multiply her by a few hundred million and you have the actual story of this country. Not a shortage of resources, and not a shortage of plans — India has never lacked five-year plans. A shortage of institutions that reliably do the thing they were built to do.
There is a version of this essay you have already read. AI doctors, AI tutors, vernacular chatbots, credit for the credit-invisible. Every pitch deck has that list, and it isn't wrong. But it is the shallow cut, and it rests on a single idea — that AI makes expertise cheap.
True. Also not the deepest thing happening here.
The deepest thing is that India's binding constraint was never a scarcity of expertise. It was a scarcity of institutions that work — the land registry, the regulator, the court, the primary health centre, the inspector, the agricultural extension officer. The economist Lant Pritchett has a name for a state like ours: flailing. The head — the policy, the Supreme Court, the elite ministries — is world-class. The signal it sends never reaches the limbs. The form of the institution is everywhere. The function almost never arrives.
The Flailing State
A state that has the form of modern institutions — the office, the ledger, the sanctioned post, the law on the books — without the function: the reliable delivery of what those institutions were built to provide. India's poor don't suffer from an absence of the state. They suffer from a state that is present in form and absent in function, and they pay a tax on that gap every single day.
Building the function is supposed to be the slow work of generations. Weber, and every development economist since, will tell you that state capacity — a bureaucracy that actually does its job — takes a century to grow and cannot be rushed. India has been trying for seventy-five years. You do not hire your way to a working land registry across seven hundred districts in a hurry, and we haven't.
That is the assumption that just broke.
Because the real thing AI changes — deeper than making expertise cheap — is this.
The registry as software. The regulator as an agent. The extension officer as a voice on a call. Not the bureaucracy — the service the bureaucracy was supposed to deliver, installed directly, in a language people speak, at the cost of electricity. India has skipped hard things before. This is a chance to skip the hardest and slowest one: a working state.
A word of discipline before the list, because a thesis without a reality check is just a hallucination with references. Indian AI startups raised about $643 million across a hundred deals in 2025. American ones raised $121 billion — a 188x gap that will not close by cheerleading. The money that is moving has turned ruthless: it funds AI performance already showing up in revenue, not AI potential in a deck, and thin wrappers on someone else's model are getting down rounds. The models are also genuinely worse in Indian languages than the demos suggest. So the filter is narrow. Build where the pain is an institutional void, the data you throw off is proprietary, and a real buyer exists — or admit you're building a public good and plan accordingly.
Ten problems, then. I'm leading with the problem in each, because the problem is the point and the solution is the easy half. None of them was a company in 2019. A couple are barely companies now — I'll tell you which, and why.
Problem 01The land you cannot prove you own
Start with the deepest one: in India, you can own land and be unable to prove it.
A farmer can point to the boundary stones his grandfather laid and still hold no document that says the land is his. India runs a presumptive title system — when property changes hands, the state records that a transaction happened, never that the seller actually owned the thing. Title is inferred backwards, through a chain of handwritten deeds across decades and a dozen scripts, and it is always challengeable. The World Bank has put roughly two-thirds of India's civil cases as disputes over land. NITI Aayog says the average one takes twenty years. Nearly four-fifths of an Indian household's wealth sits in real estate it cannot cleanly borrow against — Hernando de Soto's dead capital, and India is its largest museum.
The institution that was meant to prevent all this — a registry that guarantees who owns what — India never built. It has the form of one. A sub-registrar, a ledger, a stamp. Form without function.
The bottleneck was never policy. It was reading. Until recently no machine could parse a faded handwritten register in Devanagari and Tamil and Urdu, then stitch a sixty-year chain of deeds into one verifiable graph; a manual title search is weeks of a lawyer's time. Multimodal models now read the registers, and a language model with a graph behind it reconciles the chain, flags the break, and returns a probabilistic title opinion in minutes.
Don't sell search to consumers. Sell an instant, priced title opinion to the lenders who can't write a mortgage without one. Landeed and AdvaRisk are early and small; there is no category winner and title insurance barely exists here. The moat is the reconciled title graph itself — build it once and every lender in the country has to rent it.
Problem 02The entrepreneur who is, technically, a criminal
Run a small factory in India and you are, on paper, almost certainly a criminal.
An Observer Research Foundation count found 26,134 imprisonment clauses buried in India's business laws — and four-fifths of them were written not by Parliament but by state legislatures, where nobody is watching. Failing to whitewash the factory latrine every four months can carry jail time on par with sedition. Not maintaining a spittoon is, and I am not inventing this, a criminal matter. A single manufacturing MSME navigates something like 1,450 compliance obligations a year, forty-eight registers, fifty-nine kinds of inspector, and roughly nine thousand regulatory changes a year — about forty a day.
No human holds that in their head. So firms do the rational thing and stay small to stay invisible.
The Economic Survey even has a word for the result — "dwarfs," firms more than ten years old that never crossed a hundred workers. They are over half of organized manufacturing and produce about eight percent of its output. An American plant employs roughly eight times as many people at forty years old as at five; the Indian one barely grows at all. That stunting — the "missing middle" — is one of the most expensive facts about the Indian economy, and it is manufactured by the compliance regime itself. The void here isn't an absent institution. It's a hostile one: a state that criminalizes the entrepreneur it needs.
An agent can now read the entire regulatory corpus — all 69,233 compliances, the messy state gazette PDFs no clean API exposes — work out which subset actually binds your firm in your state, watch it change, and file. A rules engine never could, because the rules never sit still. The capability exists but is trapped: the one firm that has mapped the whole obligation graph sells it as enterprise services with no AI and no small-business pricing, and the one AI-native compliance startup that raised real money pointed it at British banks, because that's where the willingness to pay lived. "Which of the 69,233 laws apply to me, and keep me out of jail" is unbuilt for the Indian MSME — hard precisely because two-thirds of the burden is state and municipal, fragmented and undigitized, which is exactly the mess only a model that can read a bad PDF can attempt.
The graveyard is real: the demand is enormous, and a micro-business's willingness to pay to avoid a jail clause it has never been charged under is the open question.
Problem 03The doctor who never went to medical school
For most of rural India, the doctor never went to medical school.
More than seventy percent of rural primary-care visits go to providers with no formal medical training — the RMPs, the "quacks," the man with a folding table and a box of injectables. In stretches of the country, well over half the people calling themselves allopathic doctors hold no qualification for it. The reflex every health-tech deck reaches for is that AI will replace them with a real doctor on a screen.
That is the wrong read, and the data is unusually clear about why.
When Jishnu Das and Abhijit Banerjee ran a randomized trial training these informal providers in West Bengal, correct case management rose and they closed half the quality gap to actual public-sector doctors — and, tellingly, the informal providers were already prescribing fewer unnecessary antibiotics than the qualified government ones. The informal provider is not the disease. He is the distribution network the public system never built: already there, already trusted, already holding the patient. You don't replace him. You put an always-on medical attending in his ear.
A vernacular voice model on a cheap phone can now walk a semi-literate provider through a differential, flag the danger signs, catch the missed TB, and question the reflexive antibiotic — live, mid-consultation, in his language. None of that was possible before models could listen and reason in Bhojpuri. And here is the honest part, which is also the opportunity: this white space is held almost entirely by nonprofits on Google and Microsoft grants, while the venture money builds AI receptionists for urban clinics. It's empty for a reason. The informal provider is cash-poor, and whoever solves who-pays — the state, a pharmacy chain, a diagnostics lab riding along — is the one who turns a decade of pilots into a company.
Problem 04The advice that never comes, above an aquifer running dry
India is draining its groundwater faster than anywhere on earth, one well at a time, and almost nobody is advising the farmers doing it.
The country pumps about a quarter of all the groundwater extracted on the planet — more than the United States and China combined. In the northwest the water table is dropping half a metre a year and most of Punjab's blocks are officially over-exploited. It is a slow, distributed suicide driven by an honest incentive: free electricity and a guaranteed price for water-hungry paddy make it individually rational to grow precisely the wrong crop.
The institution meant to steer the farmer off that path — the agricultural extension service — is functionally dead. It reaches something like seven percent of farmers. The ratio of extension worker to farmer is worse than one to five thousand; India fields around 120,000 extension staff for a quarter-billion farmers, where China has sixty times as many. The advice does not arrive.
You cannot put a human agronomist on 126 million sub-hectare plots. You can put a satellite over every one of them. Cheap, frequent imaging reads crop health and soil moisture per plot; a model turns it into "irrigate Thursday, not today — and take this field out of paddy"; a vernacular voice call carries it to a farmer who will never open an app. The farmer won't pay, so don't ask him to. The novel money is in the layer nobody has built — satellite-verified proof that a farmer grew a thirstier crop's cheaper cousin, sold as a water credit, or as input-linked advisory, or as per-plot risk scoring to lenders and insurers who currently fly blind. SatSure and Cropin and Fasal circle the advisory; the water-aware crop-switch, measured and monetized, is open ground.
Problem 05The quality India cannot guarantee
India keeps being handed the manufacturing prize of the century and keeps fumbling it on quality.
When Apple's contract manufacturer began making iPhone casings at a plant in Hosur, one in two came off the line good enough to ship — fifty percent yield, against a zero-defect standard. That one number explains why "China plus one" keeps resolving to "Vietnam," and why manufacturing has been stuck near seventeen percent of GDP for a decade while every government swears it will reach twenty-five. India captures the press release; someone else captures the order.
The reason is not cost, and it is not labour. It is that a fragmented base of seventy-three million mostly-informal small manufacturers runs almost no process control — quality is a person squinting at parts, not a system catching defects in line — and the zero-defect global supply chains India wants to join do not tolerate squinting. The quality institution China spent forty years accreting — the discipline of the line, the integrator ecosystem, the tacit engineering — India's small-factory base simply never accumulated.
Until recently, vision-based quality control meant a specialist integrator, fixed optics, hand-tuned rules, fifty lakh rupees a line — the preserve of large OEMs. Foundation vision models plus a cheap edge box plus a commodity camera collapse that: you train it on a handful of sample parts, no code, and it flags a defect it has never seen. Entry cost falls from fifty lakh to five. Sell the qualification, not the software — a ten-lakh box that turns an unauditable line into one an Apple or a Foxconn will certify is not a tool purchase, it's an export contract the factory couldn't win before. Ripik and Frinks are early; against seventy-three million units, the field is almost empty.
Problem 06The knowledge locked behind a language
In India the door to a better life is locked, and the key is a language almost nobody has.
About one in ten Indians speaks English with any real fluency, and nearly everything worth reaching — medicine, law, engineering, the higher courts, the white-collar job — is conducted in it. The Constitution itself mandates English in the High Courts and the Supreme Court. Fluency carries a thirty-four percent wage premium, roughly the return on finishing secondary school. This is not a preference gap. It is a wall, and a billion people stand on the far side of it, locked out of mobility not by ability but by vocabulary.
The generic version of this is "build a vernacular chatbot," and the field is thick with them. Not the interesting cut. The interesting cut is that the entire corpus of high-value knowledge — the textbook, the case law, the clinical protocol — never crossed the language line, because the only tools that could carry it across used to mangle it, and a mistranslated dosage is worse than no translation at all.
Models can finally translate and tutor over dense professional text at near-zero cost — but, honestly, they are measurably worse in Indian languages and hallucinate more in exactly the low-resource tongues that need them most. That is not solved, which is the whole opportunity. Pick one credential where the answer is verifiable — Hindi-medium NEET prep, a nursing licence, a paralegal exam — so the answer key catches the model's lies. Own the checked, exam-aligned corpus and the human-in-the-loop that makes vernacular output trustworthy enough to certify on. Sell mobility — a job, a licence, a degree — not translation. The moat is the verified corpus, precisely because the raw model still can't be trusted.
Problem 07The scam economy eating the next billion
The same five hundred million people we just brought online are being robbed, in their own languages, at industrial scale.
Indians lost something like 22,800 crore rupees to cyber fraud in 2024 — up more than two hundred percent in a single year. Fake police officers place video calls and put victims under "digital arrest." Investment scams, loan-app blackmail, and now voice clones of your own son asking for money. Nearly half of Indian adults say they have already encountered a deepfake scam, about twice the global rate, and more than half of victims never report it, because who exactly would they tell.
The missing institution here is the most basic one of all: a cop at the scene of the crime. Rural police have no cyber units to trace the money; the courts can't process the volume; there is no beat officer standing between a first-time internet user and a professional fraud network wielding better AI than the bank.
Catching a scam in the moment means understanding the intent of a live call in code-switched Hindi and Tamil, and spotting a cloned voice — both now possible with small multilingual models running on the phone itself, privately, for almost nothing. Keyword blocklists never had a chance; they flag numbers, not manipulation. The least-defended users won't download a security app, so distribute through the telco and the handset — an on-device, vernacular scam detector baked into the eight-thousand-rupee Android phone. Google's own scam detection launched English-only and Pixel-only, which tells you exactly how wide the vernacular gap is. This is preventive policing, privatized and embedded in the device, because the public version was never going to arrive in time.
Problem 08The old age nobody is planning for
India is about to get old before it gets rich, and it has built almost nothing to catch the fall.
The over-sixties roughly double from about 150 million today to 350 million by 2050; somewhere around 2046 the old outnumber the children. The safety net is a fiction — nearly four in five elderly Indians have no pension, most have no health insurance, and the country crosses into deep aging at a fraction of the income China had when it did. The institution that used to absorb all this, the joint family, is dissolving as the children move to Bengaluru and Dubai and the parents stay behind in the town.
So exploding demand for care meets collapsing supply, with no pension system and fewer than five thousand geriatricians standing behind it.
What made remote care impossible was the gap between a worried child a thousand kilometres away and a non-English-speaking parent who cannot use an app. A vernacular voice companion the parent can simply talk to — one that notices the slurred speech, the skipped medication, the three quiet days of low mood — paired with an agent that coordinates local help and reports back, closes that gap for the first time. And the wedge is unusually clean: the person who pays is not the elder but the guilty, distant, employed child, whose willingness to pay is close to unlimited. Emoha and KITES are building the ops-heavy version; the AI-native triangle — companion for the parent, coordinator in the middle, dashboard for the child — is barely started.
Problem 09The address that isn't there
Most Indians don't have an address a machine can find. They have directions.
"Behind the temple, past the third turn, ask for Kumar." Four-fifths of Indian addresses are landmark-relative, and the average one points some four hundred metres from the actual door. Bad addressing costs the country somewhere between ten and fourteen billion dollars a year; a fifth of e-commerce deliveries fail on the first attempt, each failure roughly doubling the cost of the order. A slum of thousands can share a single address — which is another way of saying its residents have no spatial identity the formal system can see. The missing institution is a national street-addressing authority, the dull civic layer the West built a century ago and India never did.
Here I'll be honest: this is more a data-and-adoption problem than a pure-AI one. But the AI part is real — a model can now turn "behind the temple" into a precise coordinate by learning from a billion past deliveries, and read informal settlements straight off satellite imagery. Delhivery built exactly this from its own shipment exhaust; the government's DIGIPIN is a parallel attempt. Whoever owns the messy-text-to-doorstep dataset owns the layer — and the honest truth is the largest logistics player and the state are already building it, so this is a race, not an empty field. The opening is the neutral, shared address API everyone else can rent, if someone independent gets there first.
Problem 10The epidemic the country can't see
India finds out about most of its outbreaks late, because it can barely see them.
The formal surveillance system captures a rounding error of what is actually happening. For dengue, by one careful estimate, it records about a third of one percent of real cases — an undercount of roughly two hundred and eighty to one. A disease has to send someone to a facility, get lab-confirmed, and be typed into a form before the state registers it, by which point transmission is days or weeks ahead. The institution meant to be the country's early-warning system is watching a signal that arrives too late to act on.
The early signal does exist — it's just scattered and unstructured: pharmacy sales spiking, symptom searches, a cluster of vernacular posts, wastewater. Fusing that into an alert used to be impossible; a model that reads thirteen languages of messy text and flags the anomaly makes it a real-time capability. BlueDot famously flagged Wuhan six days before the WHO, on almost no money.
And now the most honest caveat in this whole list: the buyer is weak. The state is already building this in-house through a nonprofit and will not pay a startup for it. The only genuine venture wedge is selling the forecast sideways — to the pharma company positioning antivirals, the insurer bracing for a claims surge, the hospital chain planning its beds. This one is mostly a public good, and I would tell a founder that to their face. It is on the list because it is a real, life-and-death institutional void — not because it is the easiest company to build. Some problems earn their place by mattering.
Line the ten up and the pattern is impossible to miss. A registry that doesn't register. A regulator that criminalizes. A clinic with no clinician. An extension service that extends to no one. A police force that can't see the crime. Each is the same failure wearing different clothes: India built the form of a modern state — the office, the ledger, the sanctioned post — and never got the function to reliably show up behind it.
For seventy-five years the assumption was that you fix this the way every rich country did: slowly, over generations, brick by institutional brick, training the bureaucrats to run it. India tried. It is genuinely hard, and we are not close.
That assumption is the thing that broke.
India has leapfrogged before. It skipped landlines for mobile, skipped card rails for UPI. But those were infrastructure, and infrastructure is the easy leap. This one is heavier: a chance to skip the hardest, slowest thing a poor country ever builds — a working state — and install its function on rails we already laid. The demographic clock says the window is about a decade, which is roughly how long we have to lay the pipes before the country is old.
One caution, because I don't trust a clean story. A machine that performs an institution's function is not the same as an institution. It has the output and not the accountability. When the AI registry is wrong, whose land is it? When the model in the informal doctor's ear misses the sign, who answers for the death? A functioning state is not merely competent — it is answerable, and a model is not, yet. The work is not to hand the poor a black box in place of the institution that failed them. It is to give them the institution's function, with a human still on the hook. Get that wrong and you have only built a faster way to be unaccountable to the people who can least afford it.
But get it right, and the arithmetic of this country changes at the root.
India was never short on talent, and lately not on capital. It was short on institutions that work. For the first time in its history, that is a problem you can build your way out of.