> ## Content Index > Fetch the complete content index at: https://research.finbox.in/llms.txt > Use this file to discover other available public pages before exploring further. # Beyond the AI hype: Warning calls from the GFF stage - URL: https://research.finbox.in/newsletter/the-pattern/beyond-the-ai-hype-warning-calls-from-the-gff-stage/ - Published: 2026-09-11T12:43:05.000Z - Updated: 2026-09-11T12:43:05.000Z - Description: The global fintech fest saw equal amount of AI hype as it did warnings about AI. Documenting the pushback reveals more about the crossroads than a panel would. - Author: Mayank Jain - Tags: #ThePattern, newsletter, FinTech, AI Regulations, RBI, GFF Hello everyone, Welcome to the 219th edition of The Pattern, a weekly newsletter on the latest in finance, technology, and the economy. I was at the Global FinTech Fest this week. RBI Governor Sanjay Malhotra started with the numbers you'd expect at an event like this, India's FinTech sector is now the third largest in the world by funding, the sector raised $2.4 billion last year, and we now have 30 FinTech unicorns. But what actually caught my attention was what he said about customer data: [treat it as a fiduciary responsibility, not a business asset](https://www.deccanchronicle.com/nation/treat-data-as-fiduciary-responsibility-rbi-gov-tells-fintechs-1986469?ref=research.finbox.in). Sounds simple enough, right? Customers give FinTech companies their data for a reason. So, use it for that reason, get the right consent, and keep it safe. Having access to someone's data doesn't mean you get to keep finding new ways to use it. He also talked about the AI risks that FinTechs need to take seriously: opacity, bias, concentration, cybersecurity, and one I think gets less attention than it should: erosion of human judgement. That doesn't mean a model suddenly breaks and everyone notices. Honestly, it's usually much quieter than that. The model makes a bad call, people trust the output anyway, and nobody catches it until that bad call has already turned into a bigger problem. The RBI's been thinking about some version of this for a while now. The RBI's [2025 Digital Lending Directions](https://www.business-standard.com/amp/finance/news/rbi-asks-regulated-entities-to-report-digital-lending-apps-on-cims-125050901500%5F1.html?ref=research.finbox.in) already ask regulated entities to keep algorithmic credit decisions explainable and auditable, even when a vendor builds the model. But Malhotra's point felt bigger than model compliance. He was really talking about the responsibility that comes with having someone's data in the first place. The speech might have faded from memory if two other things hadn't brought the same question back into focus. ### **NPCI wants to know exactly which AI agent just paid your bill** Within the same week, it came out that NPCI is [building a registry to identify and vet AI agents](https://www.brecorder.com/news/40438876/india-plans-ai-registry-as-it-looks-to-roll-out-agentic-payments-sources-say?ref=research.finbox.in) before they're allowed to transact on UPI. This has been building for a while. NPCI had already said the week before that it's preparing to let AI agents make small UPI payments without a human approving every single one, using two things that already quietly exist inside UPI: UPI Circle, which lets you delegate payment authority to someone (or something) else, and Reserve Pay, which lets customers pre-block funds for repeat debits, currently capped at [₹10,000 for 90 days](https://www.brecorder.com/news/40437409?ref=research.finbox.in), a limit that's apparently up for review now, since nobody set it with an AI agent in mind. This isn't coming out of nowhere either. [The RBI's own Payments Vision 2028](https://www.globalgovernmentfinance.com/rbi-payments-vision-2028/?ref=research.finbox.in) already flagged AI, both the generative kind and the agentic kind, as one of the more transformative and riskier shifts heading into the payments system and put anti-fraud tooling and liability frameworks on the list of things to get ahead of rather than react to later. What I find interesting is the order in which this is happening. NPCI isn't giving AI agents the ability to make payments and then asking questions about accountability later. It's building the registry so the system can identify each agent and know what that agent can and can't do. That feels like the same idea Malhotra raised around data, applied to money. Before an AI agent can touch your account, someone needs to know exactly which agent it is and what it's allowed to do. Our co-founder Srijan made this point: automating the work around a loan is very different from [automating the decision itself](https://www.finbox.in/newsletter/thepattern/why-lenders-should-bet-on-ai-based-automation-not-decisions?ref=research.finbox.in), and I think the same argument is showing up here on the payments side. Automate as much of the process as you want. Just don't lose track of who's holding the pen. ### **The deputy governor said the same thing to lenders, without meaning to** The third story seemed unrelated to the first two, at least at first. Deputy governor Shirish Chandra Murmu spoke to NBFCs and HFCs at a CII summit, and most of it was about [funding, not AI](https://www.business-standard.com/finance/news/rbi-s-murmu-urges-nbfcs-hfcs-to-diversify-funding-strengthen-liquidity-126090301058%5F1.html?ref=research.finbox.in). NBFC credit now sits at 16.7% of nominal GDP, up from 15.9% a year ago, and works out to roughly 27% of what scheduled commercial banks lend, up from 26%. His point was that NBFCs need to rely less on short-term wholesale funding, help build a deeper corporate bond market, and treat securitisation as actual risk transfer with real skin in the game, not just a liquidity crutch. Then, almost in passing, he added that lenders should be using AI and machine learning to catch early signs of borrower stress, and that growth should never come at the cost of underwriting standards. That's not really a new idea, and it's close to something our co-founder and CEO Rajat pointed out recently, that [AI in lending doesn't actually win on capability, it wins on orchestration](https://research.finbox.in/blog/ai-in-lending-doesnt-win-on-capability-it-wins-on-orchestration/). Trusting a model isn't what separates the lenders who do well from the ones who don't, governing it properly is. Alternative data and faster scoring don't count for much on their own, you also need to be able to explain to an auditor, later, why the model decided what it decided. A faster approval doesn't mean anything if you can't defend it once the loan actually goes bad. ### **What I took away from all three** Put Malhotra, NPCI, and Murmu next to each other, and they're pointing at the same thing from three different rooms. Move as fast as you want, just be able to explain what happened afterward, whether someone's asking about data, a payment, or a credit decision. That seems to matter more to the RBI right now than the growth numbers do. Here's what I'd actually sit down and check this week, if I were running risk or credit at a lending shop right now: - **Pick one model in your stack that's making a call a person used to sign off on.** Underwriting, fraud, pricing, it doesn't matter which. Could you reconstruct exactly why it made that specific decision, on a specific file, if someone asked tomorrow? Not roughly, actually reconstruct it. - **Check where consent ends in your data flow.** Malhotra's fiduciary language wasn't abstract; it was aimed at exactly this: data collected for one purpose quietly getting reused for another because nobody built a wall between the two. - **If you're anywhere near agent-assisted journeys touching payments or disbursals**, get ahead of the accountability question before a regulator asks it for you. NPCI clearly isn't waiting to be asked. - **Look at where your growth is fastest right now and** ask whether your stress-testing and early-warning signals are keeping pace with it, or still running on last year's assumptions. None of these have a comfortable answer for most lending shops today. That's the point of writing them down. **Reading list** - [NPCI's AI agent registry for UPI](https://www.brecorder.com/news/40438876/india-plans-ai-registry-as-it-looks-to-roll-out-agentic-payments-sources-say?ref=research.finbox.in) - [RBI governor Malhotra on treating data as a fiduciary responsibility](https://www.deccanchronicle.com/nation/treat-data-as-fiduciary-responsibility-rbi-gov-tells-fintechs-1986469?ref=research.finbox.in) - [Deputy governor Murmu's remarks to NBFCs and HFCs](https://www.business-standard.com/finance/news/rbi-s-murmu-urges-nbfcs-hfcs-to-diversify-funding-strengthen-liquidity-126090301058%5F1.html?ref=research.finbox.in) - [RBI's Payments Vision 2028](https://www.globalgovernmentfinance.com/rbi-payments-vision-2028/?ref=research.finbox.in) - [Why lenders should bet on AI-based automation, not decisions](https://www.finbox.in/newsletter/thepattern/why-lenders-should-bet-on-ai-based-automation-not-decisions?ref=research.finbox.in) - [AI in lending doesn't win on capability. It wins on orchestration.](https://research.finbox.in/blog/ai-in-lending-doesnt-win-on-capability-it-wins-on-orchestration/) Thank you for reading. If you liked this edition, forward it to your friends, peers, and colleagues. Cheers, Mayank *All opinions expressed are my own and do not necessarily reflect the views of FinBox or its promoters.*