Pages

Saturday, November 21, 2026

Regulation of Innovation and Innovative Regulation, Revisited

 


Sixteen years on, the dilemma is familiar. The stakes are not.

In April 2010 I opened a post on my blog with a story about horse manure. In 1898, delegates to the world’s first international urban planning conference in New York were consumed not by housing or land use but by the dung produced by the horses that moved the great cities of the world. The story, as it is usually retold, comes with a forecast from London that by 1950 every street would be buried nine feet deep. The forecast never came true. The internal combustion engine arrived, solved the problem, and a century later became a problem of its own.

My argument then was simple. Every great innovation travels the same road, from saviour to Frankenstein, and in financial services that journey had become alarmingly short. Derivatives built to hedge risk became tools to push risk onto unsuspecting investors. Securitisation deepened debt markets and then ended in a situation where nobody really owned the risk of the underlying asset. I concluded that the two principal challenges of the day were the regulation of innovation and innovative regulation.

Sixteen years later I find myself returning to that post, because artificial intelligence is forcing the same question on us. The issues are different, and I will come to how different they are. But the shape of the dilemma is uncannily familiar.

What has not changed

The first thing that has not changed is human nature. In May 2010, writing about the two bullies in my high school who got away with their transgressions because they won trophies for the school, I quoted John Mack of Morgan Stanley on the cause of the 2008 meltdown: greed, leverage and lax standards. Replace leverage with scale and you have a fair description of the AI race. Firms compete to ship new capabilities every few months. Rob Reich, Nate Persily and Anka Reuel have described the frontier laboratories as “tragic accelerators”: they acknowledge the risks, call for safeguards, and keep pushing capabilities forward anyway, under pressure from rivals at home and from China. When Europe debated delaying its AI rules this year, critics warned that the delay could set off a race to market before the rules bite. As I said at the AI-DPI conference in June, market incentives do not naturally reward careful and transparent development.

The second is opacity. Hank Paulson observed that markets became opaque as structured products grew too complex even for sophisticated investors to understand. The difference, and it is an important one, is that a CDO was complex by design; someone could, in principle, take it apart. A neural network with hundreds of billions of parameters is opaque even to the people who built it.

The third is the argument against innovation itself. In 2010 many voices said the answer to the financial crisis was to curtail innovation. I disagreed then, and I disagree now with those who would freeze AI until it is perfectly understood. Claiming we were insulated from the turmoil because our markets were unsophisticated was like saying I never fell because I never rode. The same logic applies to a country of 1.4 billion people that needs AI to deliver health, credit, education and justice at a scale its institutions cannot otherwise reach. Hibernation is not a strategy. But I must also qualify what I said in 2010. My argument that mistakes trigger better controls assumed the mistakes were survivable. A mis-sold financial product can be compensated. Some AI failures cannot be undone, and for those, learning by falling is not an option.

The fourth is that self-regulation is still asking too much. Nothing in the last three years of AI suggests the technology industry has discovered a virtue that finance lacked. Voluntary commitments are useful. They are not a governance system.

What is genuinely different

Now the differences, because they are real, and a regulator who treats AI as just another derivative will fail.

Speed. The printing press, electricity and the internet each took decades to reshape society. AI capabilities shift in months. A rule written around today’s architecture risks being obsolete by the time it is notified.

The object of regulation is unclear. In finance we knew what we were regulating: a product, an institution, a transaction. With AI it is not even settled what the regulated object is. As I argued in my April essay on the DPDP Act, we cannot yet say whether the law should attach to the training dataset, the trained model weights or the outputs. A right to erasure that makes perfect sense for a customer database becomes close to meaningless once the data has been dissolved into billions of parameters.

General purpose. A credit default swap does one thing. A foundation model can draft a contract, read a scan, write malicious code and tutor a child in the same afternoon. Paulson’s complaint in 2010 was that a regulatory structure organised around traditional business lines had not kept up with the markets, leaving duplication in some areas and gaping holes in others. AI makes that problem far worse, because the same model sits underneath every sector at once. And the risks run through the whole stack, from chips, energy and data to the models and the applications built on them.

The knowledge gap has flipped. In 2010 I argued that regulators must attract people with the experience, domain knowledge and attitude to match or better industry strengths. In frontier AI, the expertise, the compute and the data sit overwhelmingly inside a handful of private laboratories. The regulator does not merely lag the industry. In many cases it cannot see what the industry is doing.

Jurisdiction. Technology moves globally; regulation remains stubbornly national. I made this point recently about digital assets and BRICS. It applies with greater force to AI, where a model trained in one country, fine-tuned in a second and distributed through an app store in a third affects citizens in a fourth.

Some of the harm cannot be undone. This is the difference that worries me most. Reich, Persily and Reuel sort AI harms into four families: malicious use and weaponisation; malfunction and loss of control; societal, economic and environmental damage; and harm to individuals and their rights. Finance gave us plenty of the last two. It never gave us the first two. A model that gives someone real help towards a chemical or biological weapon, or an autonomous agent that takes actions nobody intended across thousands of systems at machine speed, is not a mis-sold product that a fine and a refund can fix. The harm may be irreversible, and it may come from a system that was never deployed for a dangerous purpose at all.

Three answers, and the pendulum

Look at how the major jurisdictions have responded and you see the pendulum I wrote about in July 2010, when I described how auditing swung from judgement to rigid rules and was beginning to swing back towards principles.

Europe went first with a comprehensive, rules-heavy law, and is now retreating. Under the Digital Omnibus agreed this year, obligations for stand-alone high-risk systems have moved to 2 December 2027, and those for AI embedded in regulated products to 2 August 2028. The stated reason is that the standards and guidance needed to comply were not ready. The rulebook arrived before the toolkit.

The United States has gone the other way. A December 2025 executive order set up a Justice Department task force to challenge state AI laws, in pursuit of what it calls a minimally burdensome national framework. An earlier attempt to impose a ten-year moratorium on state AI laws was stripped out by the Senate by 99 votes to 1. This is regulation by litigation.

India has chosen a third path. The AI Governance Guidelines released by MeitY in November 2025 are non-binding, rest on seven principles including “innovation over restraint”, and leave binding obligations to existing laws and sector regulators. The RBI’s FREE-AI framework for the financial sector goes a step further and calls for a tolerant supervisory stance towards first-time AI errors by institutions that have put proper safeguards in place.

I have a good deal of sympathy for India’s approach. It also fits a view gaining ground among scholars, that in many respects AI is a normal technology whose impact depends more on how fast it spreads through institutions than on what the frontier model can do. And tolerance for honest first mistakes is what I argued for in 2010. But that tolerance needs a boundary. It is right for a credit model that misprices a loan. It is wrong for a system that could help someone towards a weapon, or that can act on its own across critical infrastructure. On the first kind of risk India’s framework is thoughtful. On the second it says much less, and leans largely on voluntary commitments such as those announced by frontier AI companies at the New Delhi summit in February.

But here is the uncomfortable part. Light touch works only when the hand is strong. A light-touch regime run by a regulator without capacity, data or technical depth is not light touch. It is no touch. The guidelines themselves lean on institutions such as an AI Governance Group and the AI Safety Institute. Whether those bodies get the talent, the budget, the compute and, above all, the courage to act will decide whether the Indian model proves wise or merely permissive. In 2010 I praised SEBI for calling out ULIPs as mutual funds masquerading as insurance. The AI equivalent will be the first time an Indian regulator tells a large and popular AI deployment that it must stop. We have not yet seen that moment.

What innovative regulation looks like for AI

If the challenge in 2010 was innovative regulation, what does that mean for AI? Let me offer seven propositions, drawn from what I saw work while building NSDL and ONDC, and from what I see on the boards of financial institutions grappling with AI today.

First, regulate the function, not the technology. Ask the old questions. Who bears the loss if it fails? Can the decision be challenged? Who is accountable? A loan decision is a loan decision whether a credit officer or a model makes it, and the borrower’s rights should not depend on which.

Second, tier by risk, but admit how hard that is. The obvious approach is to tier by use, with testing, incident reporting and audit for systems that decide on credit, employment, medical treatment, policing or critical infrastructure, and to add a second tier for dangerous capabilities such as weapons uplift or autonomous action at scale. The trouble is that these are not two kinds of system. A bank’s customer-service assistant built on a frontier model carries that model’s full capabilities, whatever the bank intends. The safeguards meant to keep them apart are brittle. Researchers have stripped the safety training from open-weight models in minutes, for less than a dollar of computing time, and a study by CDT and MIT found that even good-faith fine-tuning to improve medical knowledge eroded guardrails in ways that were hard to predict. The line also blurs on purpose: the International AI Safety Report notes that restricting harmful cyber uses can slow the defensive tools that find vulnerabilities first.

Then there is the value chain. High-risk obligations fall on the deployer, the Indian bank or hospital, while knowledge of what the model can do sits with the provider, often abroad, whose terms of use offer few assurances. A deployer cannot certify what it cannot see.

I do not think anyone yet has a clean answer to this, and I would rather say so than pretend. The best available approach is defence in depth: capability testing before release, controls and monitoring at the point of use, incident reporting afterwards, and a duty on model providers to share their test results with those who build on them, with liability shared accordingly. Where those lines should sit ought to be the first assignment for India’s AI Safety Institute, working with RBI, SEBI and the health regulators, rather than something each deployer guesses at alone.

Third, combine rules with principles. My “True and Fair” post argued for a sensible balance of rules and principles, with the weight shifting towards principles. Rules alone will be gamed or will age badly. Principles alone give no certainty. What works is a small number of hard rules on the things that must never happen, and principle-based supervision, backed by board accountability, for everything else.

Fourth, build regulation into the infrastructure. When we built the depository, we did not govern dematerialised shares only by writing a rulebook and sending inspectors; we built the rules into the rails. UPI did the same for payments, the account aggregator framework for consented data sharing, and ONDC for open commerce. AI governance can follow the same techno-legal path: consent artefacts that travel with data, verifiable provenance for content, and logging and audit hooks in public AI platforms. Compliance designed into the architecture is cheaper and far more effective than compliance inspected after the fact.

Fifth, put a review date on every rule. Every AI rule is a bet on how the technology, human behaviour, incentives and institutions will evolve, and some of those bets will be wrong. No AI rule should therefore be permanent by default. We need fixed review cycles, sandboxes with real exit paths, and the willingness to repeal rules that no longer serve their purpose, or to tighten them quickly when a new capability appears. The rules of T20 are not the rules of Test cricket, though both are cricket. The regulator’s job is to watch how the game is changing and modify the rules accordingly.

Sixth, make those who screw up pay up. Tolerance for honest first errors must be matched by real consequences for negligence and repeat failure. Liability is the most efficient regulator ever invented, because it aligns the incentives of the people who understand the system best. Without it, every other measure is advice.

Seventh, do not treat a human click as a safeguard. Many frameworks, ours included, rely on keeping a human in the loop. But requiring a person to approve what an AI agent does assumes that the person can understand the action, catch the error and resist simply approving it. Approval fatigue and the pressure to stay productive turn oversight into rubber-stamping, and a signature that satisfies the rule protects no one. Human oversight has to be designed, staffed and tested, and the human must have both the authority and the time to say no.

The second challenge has changed shape

The other half of my 2010 title, the regulation of innovation, was about something else. I argued that governments must enable and protect disruptive innovations, because incumbents invested in the old system would work overtime to stop them. I gave the example of the oil industry, and recalled the stiff resistance we faced from custodians and registrars when we set up the depository for the Indian capital market.

With AI there is a twist I did not anticipate. The incumbents are now the innovators. The firms at the frontier of AI are among the largest companies the world has seen, and some of them have been among the loudest voices calling for regulation. In 2023 the head of the best-known AI company told the US Senate that a licensing agency might be needed. Licensing can be a genuine safety measure. It can also be the most effective moat ever built, because the cost of compliance falls lightly on a trillion-dollar firm and heavily on a startup in Bengaluru or Jakarta.

So the regulation-of-innovation challenge today cuts both ways. We must protect citizens from the incumbents’ innovations, and we must protect the next generation of innovators from the incumbents’ regulation. I wrote in March that the deeper risk of AI is the concentration of intelligence infrastructure in very few hands. Regulation designed carelessly will accelerate that concentration. Regulation designed well, combined with public digital infrastructure for compute, data and models, can diffuse AI to the edges.

The regulator we need

In 2010 I described the regulator we needed: intellectually capable, courageous enough to take unpopular decisions, morally strong enough not to be swayed by money or influence, and wise enough to get out of the way of genuine game changers. I said such leadership is often available only in limited edition.

The AI age demands one more quality: humility. The regulator must accept that it will often be wrong, and design systems that detect and correct those errors quickly. And it must bring into the room those who bear the costs: the data annotator, the back-office worker, the borrower declined by a model she cannot question. Governance shaped by a few people in a few rooms will serve those few.

The horses did not leave our streets because a regulator banned them. They left because a better technology arrived. Then that technology needed seat belts, traffic lights, emission norms and driving licences, each introduced after the damage became visible, and each late. For everyday uses of AI we can afford some of that lateness, provided we learn quickly and make those who cause harm pay. For the most dangerous capabilities we cannot, because some accidents cannot be counted afterwards. There the seat belts have to be built while the car is still being designed. We will get it far more wrong if we pretend the choice is between innovation and regulation. It never was. The real choice is between clumsy regulation and innovative regulation.

I closed the 2010 post with a line from Arthur C. Clarke. It reads differently now that we have built an intelligence that is not our own.

“It has yet to be proven that intelligence has any survival value.” Arthur C. Clarke


References

From my own writing

1.          T Koshy, “Regulation of Innovation and Innovative Regulation”, Rollingstone Revelations, 25 April 2010

2.          T Koshy, “To be or not to be IV: Challenges of Regulation”, Rollingstone Revelations, 25 May 2010

3.          T Koshy, “Of being True and Fair”, Rollingstone Revelations, 20 July 2010

4.          T Koshy, “Governing the Age of Prediction: Why Digital Public Infrastructure May Define the Future of Freedom”, Rollingstone Revelations, 3 March 2026

5.          T Koshy, “Training Artificial Intelligence Under India’s Data Protection Regime: Navigating the DPDP Act’s Silent Fault Lines”, Rollingstone Revelations, 17 April 2026 (also at tkoshy.me)

6.          T Koshy, “My Speech at the AI-DPI ’26 Conference”, tkoshy.me, June 2026

7.          T Koshy, “For BRICS, the Digital Asset Question Isn’t Which Currency Wins. It’s Who Writes the Rules.”, Rollingstone Revelations, 14 September 2026

Other sources

8.          Rob Reich, Nate Persily and Anka Reuel, “What is AI Governance & Why Is It So Hard?”, AI Gov Lab, 1 October 2026

9.          CMS, “The EU’s Digital Omnibus on AI: What the Provisional Agreement Really Changes”, May 2026

10.      Tech Policy Press, “EU’s AI Act Delays Let High-Risk Systems Dodge Oversight”, April 2026

11.      Ropes & Gray, “Examining the Landscape and Limitations of the Federal Push to Override State AI Regulation”, March 2026

12.      Mintz, note on the December 2025 AI executive order and the Senate’s 99-1 vote

13.      Fortune, “India’s AI Impact Summit closes with the New Delhi Declaration and a $200 billion boost”, 23 February 2026

14.      MeitY, India AI Governance Guidelines, 5 November 2025

15.      Reserve Bank of India, Framework for Responsible and Ethical Enablement of AI (FREE-AI), August 2025

16.      Dmitrii Volkov, “Badllama 3: removing safety finetuning from Llama 3 in minutes”, Palisade Research, July 2024

17.      Center for Democracy & Technology, “New Report Reveals Unexpected Safety Risks from AI Fine-Tuning”

18.      Covington, “International AI Safety Report 2026 Examines AI Capabilities, Risks, and Safeguards”, February 2026

19.      Future of Life Institute, “Can we rely on information sharing?”, October 2023

Tuesday, November 17, 2026

BANKING 2035 | PART 2 OF 2 Building the Foundations

 



Defence, trust and data for India’s AI-era finance, and an agenda for regulators and the ecosystem

Five propositions to argue with

This paper is written to provoke debate among regulators, policymakers, network operators and industry, not to close it. Each proposition is argued in the sections that follow.

•      India’s weakest financial institutions are now one of its biggest systemic risks. In an open, interconnected system, an attacker only needs the least defended door (Section 2).

•      Trust is underpriced until it fails. Markets will not pay for trust they cannot see, so making it visible is a regulatory task before it is a market one (Section 3).

•      India has lost the frontier-model race by orders of magnitude, and should say so plainly. Pretending otherwise misdirects money and talent (Section 4).

•      India’s data is its best AI asset, and its most dangerous one if pooled. The institution that unlocks it must never become the country’s most valuable target (Sections 4 and 5).

•      India’s best export to the Global South will be a method, not a model. Open protocols and governance travel; Indic language models do not (Section 6).

Executive summary

Part 1 of this paper asked who will hold the money and who will hold the customer in 2035. This part asks a prior question: what must India build so that its financial system can be defended, trusted and powered by its own data, whichever scenario arrives? Most of it cannot be built by any single institution.

Defence. AI-assisted attacks reportedly took nearly Rs 300 crore from smaller NBFCs in about two months in 2026, and quantum computing threatens the encryption the system relies on. India has 8,502 registered NBFCs and about 1,457 urban co-operative banks, most without the means to defend themselves alone, and 93% of the institutions surveyed by the RBI rely on third parties for core security functions. Resilience has to be built as a shared capability.

Trust. Customers trust their banks with data more than they trust how fraud is resolved. Markets will not pay for trust they cannot see, so making it visible and machine-readable, for customers and for the AI agents that will act for them, is a task for regulators and standard-setters first.

Data. India is far behind in frontier models, by orders of magnitude in money and with the world’s largest net outflow of AI talent. Its advantage lies in the data generated by public rails and by new sector stacks in agriculture, health, energy and lending. But registries are not yet training data. A federated institution for financial AI could change that, if it is designed so that it does not become the country’s most valuable target.

Talent and the Global South. India’s engineering base and its experience of population-scale systems suit AI use cases for ordinary citizens. Its most valuable export to the Global South is likely to be open methods and governance rather than models.

The agenda. Section 7 sets out seven proposals for regulators, network operators and industry bodies. Section 9 poses questions this paper does not fully answer.

1. The context: four scenarios for 2035

Part 1 identifies three forces reshaping banking: agentic AI, tokenisation under the Finternet vision set out by Agustín Carstens and Nandan Nilekani, and stablecoins. It crosses two uncertainties: whether money stays in regulated bank form or fragments into stablecoins and other tokens, and whether customers keep deciding inside bank channels or delegate to AI agents and open-network apps. That gives the four scenarios in Figure 1.

Title: Scenario frame - Description: Two by two matrix of four 2035 banking scenarios by money layer and customer interface

Figure 1. Scenario frame for 2035, from Part 1: money layer by customer interface.

On current policy, Part 1 judges India most likely to move toward Scenario B, in which public rails and AI agents own the customer interface and banks supply regulated balance sheets. That makes the foundations in this part more pressing in India than elsewhere. An open, agent-driven system is only as strong as its weakest connected institution, only as trusted as its least visible standard, and only as intelligent as the data it can safely use.

2. Defence as a precondition for growth

The threat has already reached the weakest links. In October 2026, The Economic Times reported that AI-assisted cyberattacks had taken nearly Rs 300 crore from smaller NBFCs over about two months, with around 150 complaints and FIRs filed in Mumbai, Delhi, Chennai and other cities. Attackers reportedly spent 10 to 12 days studying each target, broke in through internet-facing applications and exposed APIs, took over credentials for payment gateways and partner banks, and moved funds to mule accounts through IMPS and NEFT. In one Hyderabad case, an NBFC lost more than Rs 3.3 crore through 183 unauthorised transfers in 87 minutes; police were able to block only about Rs 40 lakh.

AI raises the speed of attack; quantum threatens the foundations. The RBI’s June 2026 Financial Stability Report found AI-enabled cyberattacks to be the leading perceived risk among Indian financial institutions for the year ahead, with most rating their own preparedness as developing or intermediate rather than mature. Quantum computing adds a slower but deeper threat. A sufficiently powerful quantum computer would break the public-key encryption that protects payments, authentication and stored data. Because adversaries can copy encrypted data today and decrypt it later, the risk begins before such a machine exists. The G7 Cyber Expert Group’s January 2026 roadmap sets 2035 as the overall target for moving the financial sector to quantum-resistant cryptography, with the most critical systems prioritised for around 2030 to 2032.

India has started, mostly at the top. The National Quantum Mission’s task force published a quantum-safe roadmap in February 2026 that lists financial services among priority sectors. In May 2026 the RBI set up an expert committee on a Quantum Secure and Adaptive Financial Ecosystem, chaired by Professor Anil Prabhakar of IIT Madras with members from SBI, NPCI, DST, MeitY and DSCI, to review the sector’s cryptographic inventory and propose a roadmap. The finance ministry has reportedly asked banks to develop a sector-wide strategy, with an SBI-led working group examining the transition. On fraud, the India Digital Payment Intelligence Corporation, set up in consultation with the RBI to share real-time fraud intelligence, was running in pilot with eight banks in September 2026.

The gap is in the long tail. India has 8,502 registered NBFCs and about 1,457 urban co-operative banks, alongside small finance banks, payments banks and the fintech partners connected to all of them. The RBI’s July 2026 directions on cybersecurity and technology risk now cover all of these entity types, scaled by size and layer, but rules do not create capability. In the RBI’s survey, 93% of institutions relied partly or heavily on third parties for security functions such as monitoring, cloud security and threat intelligence, and 81% spent less than 5% of revenue on IT in 2025-26. That creates two linked risks: weak entry points for attackers, and concentration in a few vendors whose failure would hit many institutions at once. In a system joined by open APIs, UPI and Account Aggregator flows, a breach at a small lender can quickly become a problem for its partner bank and for the rails it uses. As RBI Deputy Governor Shirish Chandra Murmu said in September 2026, quantum resilience is “an ecosystem capability, not an institutional one”.

What follows. Security can no longer be treated as a cost centre that slows growth. In an open, agent-driven system, an institution that cannot show it is secure will increasingly be unable to connect, and will not be chosen. Large banks can largely fund their own defence. For the long tail, the realistic answer is shared capability rather than thousands of separate efforts: sector-level security operations and threat intelligence, a common standard for cryptographic inventories, certified quantum-safe components from vendors, and a minimum security floor for any entity that connects to public rails. Section 7 sets these out.

3. Trust as a standard, not a slogan

Convenience has outrun trust. India’s digital finance was built to remove friction, and it has succeeded. Trust has largely been treated as a given, something the regulator guarantees in the background. Customers notice the gap. In an EY survey of 2,030 Indian banking customers published in May 2026, 82% said they trust their bank with their data, but only 73% believed fraud is resolved effectively.

Markets rarely pay for trust they cannot see. A trust premium is plausible, but it will not appear on its own. Customers cannot judge the security of an app or the conduct of a lender until something goes wrong, so trust tends to be underpriced until a failure and then overpriced in a panic. A premium becomes real only when trust is made visible and verifiable through common standards. India already has early building blocks. The RBI required banks to move to the exclusive bank.in internet domain, registered through IDRBT, by October 2025, with fin.in planned for other financial entities, so that customers can tell a genuine site from a fake. Regulated lenders must report their digital lending apps to the RBI so that borrowers can check whether an app is linked to a regulated entity. The Account Aggregator framework built consent into the architecture rather than the fine print. The DPDP Rules, notified in November 2025, make obligations on consent, security safeguards and breach notification enforceable from May 2027, with penalties of up to Rs 250 crore.

AI agents change the economics of trust. A person rarely reads a security certificate; an AI agent can check every one, every time. If trust signals are standardised and machine-readable, agents acting for customers can route business to institutions that meet them. That is how a trust premium could be earned in practice: through lower acquisition costs, stickier deposits and cheaper funding. The global payment networks are moving in this direction. In September 2026, Visa, Mastercard and Ant International announced work on an interoperable Know Your Agent framework so that AI agents can be identified and verified across networks. The logic runs both ways: institutions must also be able to verify the agents that act on customers’ behalf.

What a financial trust standard could include. Verified identity for domains, apps and agents; clear disclosure of consent and data use; security certification against a published baseline; published service and fraud-resolution metrics; and defined liability when something fails, following the “liability follows control” principle NPCI is weighing for agent payments. Certification would need to be independent, tiered by size and re-tested regularly, so that it signals current practice rather than a one-time audit.

4. India’s data advantage: an institution for financial AI

India is far behind in frontier models. The gap is one of orders of magnitude, not a few years. The IndiaAI Mission has a budget of Rs 10,372 crore over five years, roughly $1.2 billion in total. Amazon, Microsoft, Alphabet and Meta together plan about $725 billion of capital spending in 2026 alone, most of it on AI: roughly 3,000 times India’s annual public outlay. Private capital does not close the gap; the 2026 Stanford AI Index puts US private AI investment at $285.9 billion in 2025, against $12.4 billion for second-placed China. Talent tells a similar story. India has the world’s second-largest pool of AI researchers, about 50,000, yet recorded the largest net outflow of AI talent of any country in the same report. An earlier MacroPolo study found that 12% of elite AI researchers took their first degree in India, but almost none of them worked there.

The sovereign models are a hedge, not a frontier challenge. The IndiaAI Mission has selected 12 organisations, including Sarvam AI and BharatGen, to build models trained on Indian data and languages, supported by more than 38,000 GPUs offered at subsidised rates. Early results show the distance still to travel: Sarvam’s 2025 model, Sarvam-M, had 24 billion parameters, was built on an open model from France’s Mistral, and drew 23 downloads in its first two days. Later models have improved, but they are optimised for Indian languages and cost rather than for competing with the leading general-purpose systems. They are valuable as insurance and as a base for local adaptation; they should not be mistaken for parity.

The realistic advantage lies in applications and diffusion. Value from AI increasingly comes from applying models to specific, high-quality data, and India’s public rails generate such data at a scale few countries match: more than 25 billion UPI transactions a month, consented financial data through the Account Aggregator network, GST filings, credit bureau records and ONDC commerce flows. For finance, smaller specialised models trained on this data, for credit, fraud and risk, are where India can compete, because the advantage lies in the data rather than in compute.

But an applications strategy depends on someone else’s models. Indian institutions will run largely on closed models from US firms or on open-weight models from the US and China. That creates three risks. Pricing and access are set abroad and could change with licence terms or export controls. Sensitive financial data may have to leave Indian control to be processed. And models tuned to other populations may perform worse on Indian borrowers, which in credit is a question of fairness and accuracy, not only of language. The realistic sovereign goal for finance is therefore control over data, evaluation and deployment, rather than ownership of frontier models.

The data exists but is fragmented. It sits with banks, NBFCs, bureaus, payment apps, tax authorities and networks, under different rules and formats. AIKosh, the national AI data platform, held 7,541 datasets across 20 sectors in February 2026, but its public documentation names no finance datasets. The RBI’s FREE-AI committee recommended shared infrastructure to democratise access to data and compute, and the development of indigenous financial-sector AI models. In October 2026 MeitY published a handbook on the Data Empowerment and Protection Architecture (DEPA) for AI training. It proposes that models travel to the data in secure environments rather than data being pooled, and gives credit models trained on de-identified GST and utility payment data as an example.

The case for an institution. India built UPI through NPCI, a not-for-profit owned by banks, and built Account Aggregator through licensed operators with Sahamati as their self-regulatory body. A comparable institution for financial AI could do three things no single firm can. It could set common standards for how financial data is described, labelled and de-identified. It could run secure environments in which approved models are trained and tested on data that never leaves its custodians. And it could certify the resulting models for fairness and performance before they are used in lending, fraud or risk decisions. The prize is not a single Indian chatbot. It is credit, fraud and risk models trained on Indian behaviour that work for thin-file borrowers and small businesses, and that institutions of every size can use, including those that could never build such models alone. That is how India could lead in the diffusion of AI across finance.

The design risks are real. The DEPA handbook itself acknowledges that removing names does not prevent re-identification, and its proposal to go beyond individual consent for training sits uneasily with India’s consent-led data protection law. A central pool of financial data would also be one of the most valuable targets in the country, which links this issue directly to Section 2. The institution should therefore be federated rather than centralised, consent-respecting by default, governed with representation for citizens as well as industry, and held to the highest security tier from the start.

5. Sector stacks: data foundations for use cases

India is building data rails sector by sector. After identity, payments and financial data, India is now building comparable infrastructure in other sectors. Agri Stack had created more than 10.31 crore farmer IDs by August 2026, linked to land records, and its Digital Crop Survey covered more than 31.3 crore plots in Rabi 2025-26; states keep ownership of their data under a federated design. The Ayushman Bharat Digital Mission had issued more than 93 crore health accounts, with over 104 crore health records linked, by July 2026, and its National Health Claims Exchange aims to standardise insurance claims the way UPI standardised payments. The India Energy Stack, led by REC under the Ministry of Power, reviewed its final report in October 2026 after pilots in six utilities across five states, including peer-to-peer energy trading and energy credentials held in DigiLocker. In lending, the Unified Lending Interface already connects 64 lenders to more than 136 data services.

Language is the layer that makes these usable by everyone. AI4Bharat at IIT Madras, set up as the Nilekani Centre at AI4Bharat in 2022 with Rs 36 crore from Nandan and Rohini Nilekani, and supported with a total of Rs 70 crore from Nandan Nilekani by 2025, builds open-source speech, translation and text models for India’s 22 scheduled languages. Its datasets include 15,000 hours of transcribed speech from more than 400 districts, and its models are used in Bhashini, the national language platform. For a farmer applying for credit or a patient filing a claim, a voice interface in their own language may matter more than any frontier model.

Each stack opens a financial use case. Farmer IDs linked to land records and crop data can support faster, cheaper crop loans and insurance for borrowers that banks have struggled to underwrite. Health records and a common claims exchange can make health insurance cheaper to process and easier to price. Energy credentials and peer-to-peer trading can support lending for rooftop solar and other small renewable assets. Combined with ULI and the Account Aggregator network, these stacks could let a lender assess a farmer, a clinic or a small solar installer from verified data, with consent, rather than from collateral alone.

Registries are not yet training data. Most of these stacks are identity, registry and exchange layers. Their value for AI depends on data quality, how often records are updated, whether consent covers model training, and whether states and departments will share across silos. Linked health records still average little more than one per account, which shows reach rather than depth, and a farmer ID tied to out-of-date land records can mislead a credit model. The institution proposed in Section 4 would need to work across these stacks, not only within finance, with the same federated and consent-respecting design.

6. From India’s engineers to the Global South

India’s advantage is building at population scale and low cost. India has about 1,900 global capability centres employing roughly 2.5 million professionals, the world’s second-largest pool of AI researchers, and the highest AI skill penetration on LinkedIn’s index according to the 2026 Stanford AI Index. It also has a decade of experience building population-scale systems that work on basic phones, in many languages and at near-zero cost per transaction. That combination suits the use cases that matter most to ordinary citizens: credit for first-time borrowers, crop and health insurance, benefit delivery and grievance redress.

But the talent is not yet where the use cases are. The same capability-centre study found a 43% shortfall in AI, data and analytics roles. One industry estimate puts India’s 2026 need at about 1.4 million AI professionals against a pool of about 400,000. Much of India’s engineering capacity serves global enterprises rather than Indian citizens, and India loses more AI talent than any other country. Turning engineers into builders of public-interest use cases needs open data foundations, public procurement that pays for outcomes, and institutions such as AI4Bharat that keep talent working on Indian problems.

The Global South is the natural market. India’s digital public infrastructure is already travelling: in July 2026 the Ministry of External Affairs said 25 countries were piloting components of India’s stack. At the India AI Impact Summit in February 2026, the first in the series held in the Global South, 92 countries and organisations endorsed the New Delhi Declaration, and a Global AI Impact Commons was launched with more than 80 deployment stories from over 30 countries. Countries that share India’s conditions, with many languages, large informal economies and thin credit histories, need affordable, open applications they can adapt and govern themselves.

What India can offer, and what it cannot. India cannot offer the Global South frontier models. It can offer open protocols, reference applications, methods for building language data, and governance models that keep data under national control. For finance, that means open lending and payment rails, credit models that work for thin-file borrowers, and the trust and security standards proposed in Section 7. India’s Indic language models will not transfer directly to other countries, but the open, public-good approach behind them can. The offer competes with Chinese infrastructure packages and with global technology platforms, and it will succeed only if it is cheaper to adopt, easier to govern and proven at home first.

7. An agenda for regulators and the ecosystem

Several of the moves in Part 1’s board agenda cannot be made by individual institutions. They need public or collective action, and most can build on institutions India already has. Table 1 sets out seven proposals for discussion.

Proposal

What it would do

Who could lead

Shared cyber defence for smaller lenders

Security monitoring, threat intelligence and incident response offered as a service to NBFCs, co-operative banks and small finance banks, priced by size

RBI with industry bodies, building on CSIRT-Fin and IDPIC

A quantum-readiness standard for finance

A common cryptographic inventory format, certified quantum-safe components and staged deadlines in line with the G7 2030 to 2035 horizon

RBI quantum committee with DST and CERT-In

A security floor for connecting to public rails

Any entity connecting to UPI, Account Aggregator, ULI or ONDC meets a published, tested baseline and is re-certified regularly

Network operators and their self-regulatory bodies, with the RBI

A financial trust standard

Machine-readable trust signals for institutions, apps and agents; published fraud-resolution metrics; agent identity aligned with global Know Your Agent work

RBI with self-regulatory organisations; IDRBT for identity

Liability rules for agent-initiated transactions

Settle who bears the loss when an AI agent acts wrongly, before agent payments scale

RBI and NPCI

An institution for financial AI data

Federated, DEPA-based secure training environments, common data standards and model certification for finance

RBI and Reserve Bank Innovation Hub with MeitY’s IndiaAI Mission; industry-owned on the NPCI model

Open use cases for the common citizen and the Global South

Open reference applications for credit, insurance and benefits built on Agri Stack, ABDM, the India Energy Stack and ULI, packaged for adoption by partner countries; funding for public-interest AI talent

MeitY’s IndiaAI Mission and the Ministry of External Affairs, with AI4Bharat, network operators and industry

Table 1. Proposals for regulators, network operators and industry bodies.

India’s experience with UPI and Account Aggregator suggests a pattern: public standards and shared infrastructure, with private institutions competing on top. The same pattern can work for defence, trust and data. The risk of not acting collectively is that the gains of AI and the Finternet accrue to the largest banks and platforms, while the losses fall on the long tail of smaller institutions and their customers.

8. Signposts to watch

Table 2 lists signals that would show whether these foundations are being built, or whether the risks are growing faster. Part 1 sets out the market signposts for the scenarios themselves.

Signpost

Points toward

Status, November 2026

RBI’s quantum committee publishes a roadmap with deadlines

Cross-cutting: defence

Committee set up May 2026; report due six months after its first meeting

Further AI-enabled attacks on lenders connected to public rails

Cross-cutting: defence and trust

Nearly Rs 300 crore reported lost by smaller NBFCs in about two months, October 2026

DPDP obligations on consent, security and breach notice become enforceable

Cross-cutting: trust

Main obligations from May 2027

A secure environment for training financial AI on Indian data goes live

Cross-cutting: data; supports B

DEPA for AI training handbook published October 2026

Agri Stack or ABDM data opened, with consent, for credit and insurance underwriting

Cross-cutting: data; supports B

10.31 crore farmer IDs (August 2026); 93 crore health accounts (July 2026)

Table 2. Signposts for defence, trust and data.

9. Questions for discussion

These questions do not have settled answers. They are offered to regulators, policymakers, network operators, industry bodies and civil society to argue over.

•      Should a lender that cannot meet a minimum security standard be allowed to connect to UPI or the Account Aggregator network at all, knowing that excluding it may also cut off credit to its borrowers?

•      Who should pay for shared cyber defence for more than 8,500 NBFCs and 1,400 urban co-operative banks: the institutions, the networks they use, or the state?

•      Can individual consent survive AI training at population scale, or does India need a new form of collective consent? Who speaks for the citizen whose data trains the model?

•      Should India keep funding sovereign foundation models, or redirect more of that money to data, evaluation and applications?

•      Is a national institution for financial AI data a public good or a honeypot? What governance would make you trust it with your own data?

•      What would make a Global South country choose India’s open approach over a Chinese infrastructure package or a global technology platform?

•      If trust becomes machine-readable, will institutions compete on being trustworthy, or on gaming the signals?

10. Limitations and next steps

•      Several figures come from secondary reporting of primary documents (RBI, Sahamati, NPCI, BIS, BCG-FICCI-IBA).

•      Section 2 relies partly on news reporting of recent cyber incidents and policy announcements, some of which are not yet confirmed in official documents.

•      The proposals in Section 7 are outlines; each would need its own design, governance and cost assessment.

•      Sections 5 and 6 go beyond banking. They are included because sector data and engineering talent will shape financial use cases, but they are not examined in depth, and some talent estimates come from commercial reports.

•      Other emerging-market DPI systems, including Indonesia, are treated briefly. The scenario analysis and its supporting evidence are in Part 1.

About the author

T Koshy is the founding Managing Director and CEO of India’s Open Network for Digital Commerce (ONDC), and serves on the Board of Commissioners of the Indonesia Open Network (ION), a Beckn protocol-based digital public infrastructure initiative modelled on ONDC.

He is a Director on the boards of CRIF High Mark, Aditya Birla Housing Finance, Perfios Account Aggregation Services and Last Mile Care.

His career spans NSDL, EY, Projects with the World Bank and the Asian Development Bank, as well as several national digital transformation programmes in India. He is a distinguished alumnus of IIM Bangalore.

"The art of war teaches us to rely not on the likelihood of the enemy's not coming, but on our own readiness to receive him; not on the chance of his not attacking, but rather on the fact that we have made our position unassailable." Sun Tzu, The Art of War, Chapter VIII 

Disclosure: the author’s roles at a credit information company, an Account Aggregator, a housing finance company and an open commerce network relate directly to several proposals in this part, including the security floor for public rails, the trust standard and the financial AI data institution. The views expressed are his own and do not represent those of any organisation with which he is associated.

Sources

1.      Carstens, A. and Nilekani, N., Finternet: the financial system for the future, BIS Working Paper 1178, April 2024. https://www.bis.org/publ/work1178.pdf

2.      Reserve Bank of India, Report of the FREE-AI Committee, press release, August 2025. https://www.rbi.org.in/Scripts/BS_PressReleaseDisplay.aspx?prid=61018

3.      Forkast, India eyes AI-driven UPI payments as NPCI weighs agent protocol, September 2026. https://forkast.news/india-eyes-ai-driven-upi-payments-as-npci-weighs-agent-protocol/

4.      Free Press Journal, UPI transactions hit 25.47 billion in October, November 2026. https://www.freepressjournal.in/business/upi-transactions-hit-2547-billion-in-october-value-soars-to-3153-lakh-crore

5.      IANS, India’s Account Aggregator ecosystem drives Rs 3.82 lakh crore loans in FY26 (Sahamati), August 2026. https://ianslive.in/indias-account-aggregator-ecosystem-drives-rs-382-lakh-crore-loans-in-fy26--20260827133606

6.      MediaNama, Unified Lending Interface: 64 lenders, 136 data services, January 2026. https://www.medianama.com/2026/01/223-united-lending-interface-64-lenders-136-data-services/

7.      Inventiva, citing The Economic Times, AI-powered cyberattacks siphon nearly Rs 300 crore from small NBFCs, October 2026. https://www.inventiva.co.in/trends/ai-powered-cyberattacks-siphon-%E2%82%B9300-crore-from-small-nbfcs-as-rbi-flags-risks-to-banks/

8.      Siasat, Hyderabad NBFC loses Rs 3.3 crore in 87-minute cyberattack, October 2026. https://www.siasat.com/hyderabad-nbfc-loses-rs-3-3-crore-in-87-minute-cyberattack-3562826/

9.      Outlook Money, AI-enabled cyber threats seen as biggest risk in the next 12 months: RBI Financial Stability Report, July 2026. https://www.outlookmoney.com/personal-finance/ai-enabled-cyber-threats-seen-as-biggest-risk-in-the-next-12-months-rbi

10.   K S & K, RBI Cybersecurity, Technology Risk, Resilience and Assurance Framework, August 2026. https://ksandk.com/newsletter/rbi-cybersecurity-technology-risk-resilience-framework-2026/

11.   G7 Cyber Expert Group, Quantum Roadmap for the financial sector, January 2026. https://home.treasury.gov/system/files/136/G7-CEG-Quantum-Roadmap.pdf

12.   The Quantum Insider, India reveals national plan for quantum-safe security, February 2026. https://thequantuminsider.com/2026/02/09/india-reveals-national-plan-for-quantum-safe-security/

13.   Retail Banker International, Indian central bank sets up panel on quantum technology for financial sector, May 2026. https://www.retailbankerinternational.com/news/india-quantum-technology-financial-sector-panel/

14.   Reserve Bank of India, Trusted Innovation: Shaping the Future of Finance, speech by Deputy Governor Shirish Chandra Murmu, 11 September 2026. https://www.rbi.org.in/Scripts/BS_SpeechesView.aspx?Id=1579

15.   The420.in, Government directs banks to develop quantum-safe systems and AI security framework, October 2026. https://the420.in/india-banks-quantum-cybersecurity-ai-framework/

16.   Free Press Journal, RBI speeds up IDPIC rollout to combat digital payment frauds, September 2026. https://www.freepressjournal.in/business/rbi-speeds-up-idpic-rollout-to-combat-digital-payment-frauds-pilot-running-with-8-banks

17.   Business Standard, UCBs decline to 1,457 in FY25, December 2025. https://www.business-standard.com/industry/banking/ucbs-decline-to-1457-in-fy25-125122901022_1.html

18.   NewKerala, Customers trust banks with data, but only fully satisfied with fraud resolution: EY report, May 2026. https://www.newkerala.com/news/a/customers-trust-banks-data-only-fully-satisfied-fraud-782.htm

19.   Business Standard, RBI asks banks to complete migration to bank.in domain by October 31, 2025, April 2025. https://www.business-standard.com/finance/news/rbi-asks-banks-to-complete-migration-to-bank-in-domain-by-october-31-2025-125042201515_1.html

20.   MediaNama, RBI digital lending apps centralised directory, May 2025. https://www.medianama.com/2025/05/223-rbi-digital-lending-apps-centralised-directory/

21.   Mondaq, India’s Digital Personal Data Protection Act: phased rollout and key compliance milestones, November 2025. https://webiis08.mondaq.com/india/data-protection/1706878/indias-digital-personal-data-protection-act-phased-rollout-and-key-compliance-milestones

22.   Compliance Hub, India DPDP consent manager and phase deadlines, 2026. https://compliancehub.wiki/india-dpdp-consent-manager-november-2026-phase-two-deadline-compliance/

23.   PYMNTS, Visa and Mastercard team with Ant on Know Your Agent framework, September 2026. https://www.pymnts.com/cybersecurity/2026/visa-mastercard-team-with-ant-know-your-agent-framework/

24.   Press Information Bureau, IndiaAI Mission backgrounder, February 2026. https://static.pib.gov.in/WriteReadData/specificdocs/documents/2026/feb/doc2026210784401.pdf

25.   YourStory, RBI panel suggests measures to mitigate AI risks in financial sector (FREE-AI), August 2025. https://yourstory.com/2025/08/rbi-panel-suggests-measures-mitigate-ai-risks-financial-sector

26.   ThePrint, Govt handbook for sharing hospital and bank data to train AI models, November 2026. https://theprint.in/india/governance/govt-handbook-for-sharing-hospital-bank-data-to-train-ai-models-says-individual-consent-impractical/3089636/

27.   Tom’s Hardware, Big Tech’s AI spending plans reach $725 billion, April 2026. https://tomshardware.com/tech-industry/big-tech/big-techs-ai-spending-plans-reach-725-billion

28.   ThePrint, India leads in AI talent but also brain-drain anxiety, says Stanford’s AI Index report, April 2026. https://theprint.in/india/governance/india-leads-in-ai-talent-but-also-brain-drain-anxiety-says-stanfords-ai-index-report/2909479/

29.   ThePrint, 12 per cent of elite AI researchers come from India but almost none work here (MacroPolo Global AI Talent Tracker), June 2020. https://theprint.in/tech/12-per-cent-of-elite-ai-researchers-come-from-india-but-almost-none-work-here-says-data/439049

30.   Outlook Business, Sarvam’s Indic AI model: hype, hope and the hunt for tech sovereignty, May 2025. https://www.outlookbusiness.com/news/sarvams-indic-ai-model-hype-hope-and-the-hunt-for-tech-sovereignty

31.   IIT Madras, IIT Madras launches Nilekani Centre at AI4Bharat, July 2022. https://www.iitm.ac.in/happenings/press-releases-and-coverages/iit-madras-launches-nilekani-centre-ai4bharat-promote

32.   Entrepreneur India, Nandan Nilekani increases support to AI4Bharat with total grant of INR 70 crore, June 2025. https://india.entrepreneur.com/?p=84828

33.   BioVoice News, India creates over 10.31 crore Farmer IDs under AgriStack, August 2026. https://biovoicenews.com/india-10-31-crore-farmer-ids-agristack/

34.   News On AIR, Over 104 crore health records linked to over 93 crore ABHA accounts, July 2026. https://newsonair.gov.in/over-104-crore-health-records-linked-to-over-93-crore-abha-accounts-says-govt/

35.   Telangana Today, National Health Claims Exchange will simplify health insurance claims processing, March 2026. https://telanganatoday.com/national-health-claims-exchange-will-simplify-health-insurance-claims-processing-across-india-says-nha-ceo

36.   T&D India, India Energy Stack taskforce reviews final report, sets out pathway for national rollout, October 2026. https://www.tndindia.com/india-energy-stack-taskforce-reviews-final-report-sets-out-pathway-for-national-rollout/amp/

37.   Storyboard18, India’s 1,900 GCCs face 18 to 43% talent shortfall in key tech roles (Quess Corp), February 2026. https://www.storyboard18.com/how-it-works/indias-1900-gccs-face-18-43-talent-shortfall-in-key-tech-roles-quess-corp-90866.htm

38.   UNI India, AI talent demand gap (BinQle Talent Intelligence Report 2026), October 2026. https://www.uniindia.com/business-economy/ai-talent-demand-gap-india/208921

39.   Open Source For You, MeitY and MEA drive global rollout as 25 nations pilot India’s DPI, July 2026. https://www.opensourceforu.com/2026/07/meity-mea-drive-global-rollout-as-25-nations-pilot-indias-dpi/

40.   Press Information Bureau, India AI Impact Summit 2026 outcomes, 2026. https://pib.gov.in/PressReleasePage.aspx?PRID=2238195