Home

Featured Articles

Oct 21, 2026
Rudra Dubey

The Singularity Has Begun: Indian Scientists Must Command AI-Powered Virtual Armies—Before the Window Closes

In October 2026, Anthropic co-founder Jack Clark delivered a stark message to economists and policymakers gathered at the IMF–World Bank Annual Meetings in Bangkok. The technological singularity, the point at which artificial intelligence accelerates scientific and technical progress faster than humans can comfortably follow, has already begun. He cited specific measures: scores on Humanity’s Last Exam, a battery of expert-level questions, rising from roughly 8 per cent at its 2025 launch to 67.7 per cent; FrontierMath, a set of research-grade mathematics problems, climbing from 2 per cent to approximately 94 per cent; and AI systems that not only solved unpublished theoretical physics problems but correctly identified errors in the reference answers supplied by human experts. These benchmark improvements are more than incremental: they indicate that greater-than-human intelligence is beginning to drive discovery in shorter periods.3 Clark’s comparison has particular significance for science and technology policymakers in India: the world’s approximately 10 million scientists will soon direct a billion, or even ten billion, virtual scientists. Intelligence and ideas are no longer scarce. The shortages are in physical laboratories, robotic experimental capacity, high-quality data loops, and the human judgment required to steer vast automated systems productively and safely. Countries that continue to organise science around the individual investigator painstakingly forming and testing hypotheses will be operating at human scale. Those that reorganise around scientists as commanders of virtual armies backed by automated laboratories will operate at machine scale. The productivity gap will compound quickly. In Clark’s assessment and in the view of many frontier researchers, countries have only a few years to adapt, rather than decades. Adding more AI tools to existing scientific work is not enough. The fundamental workflow of science is changing. The Collapse of the Classical Scientific Workflow For most of modern history, the basic unit of scientific progress has been the human researcher or small team. A scientist reads the literature, forms a hypothesis, designs an experiment, performs or supervises the physical work, analyzes the results, and iterates. Instrumentation improved, computational power grew, and collaboration expanded, yet the cognitive and manual bottlenecks remained human. Even large pharmaceutical R&D organisations essentially expanded this model: more scientists, more sequential stages, and more parallel projects, all still limited by human attention and laboratory throughput. That model is now breaking down. The emerging workflow reverses the division of labour: Large groups of specialised AI agents perform the bulk of hypothesis generation, literature synthesis across millions of papers, multi-omics data integration, molecular and protein design, experimental protocol optimisation, failure analysis, and iterative refinement. These agents can operate continuously, explore vast combinatorial spaces, and maintain perfect memory of prior attempts. Physical experimentation shifts toward robotised, closed-loop, self-driving laboratories capable of executing design–build–test–learn cycles with limited human intervention between cycles. High-throughput synthesis, automated cell culture, robotic screening, integrated analytics, and adaptive experimental design allow the virtual armies to receive rapid, high-volume feedback from the real world. The human scientist’s irreplaceable contributions become problem selection and framing, strategic direction of the AI agent groups, critical evaluation and correction of AI outputs, ethical and safety oversight, integration of tacit and contextual knowledge, and translation of validated discoveries into products, services, and clinical or industrial practice. Under this model, a single highly capable researcher can productively supervise the equivalent of hundreds or thousands of expert colleagues while keeping multiple automated laboratory systems running. The effective size of the scientific workforce is no longer limited primarily by the number of trained humans but by the availability of computing resources, high-quality data, automated experimental capacity, and people skilled at directing these systems. This change marks the early singularity in the life sciences and other fields. Not every domain will automate at the same rate. Wet biology remains messier than pure mathematics or certain areas of chemistry. Data quality, distribution shift, rare events, and the need for causal rather than correlative understanding continue to constrain pure in silico progress. Regulatory requirements, manufacturing scale-up, and clinical validation impose additional practical constraints. Human creativity in identifying truly novel questions and in exercising judgment under uncertainty will remain essential. Even so, the trend is clear: the fraction of scientific labour that can be productively performed by virtual agents is rising steeply, and the organisations that redesign their work accordingly will pull ahead. Evidence from around the world is already accumulating. Discovery programs built around AI or using it extensively have compressed the time from target identification to a preclinical candidate from the traditional multiyear span to 13–30 months in leading cases. More than 170 AI-originated programs are in the clinic. Early analyses show elevated Phase I success rates for some AI-discovered molecules. Agentic systems are being deployed across literature review, target prioritisation, molecular generation, trial design, and manufacturing optimisation. Self-driving laboratory concepts are moving from prototypes to working systems. Economic projections include tens of billions in annual value from generative and agentic AI in pharmaceuticals, potential double-digit productivity gains in clinical development, and the possibility of dozens of additional novel therapies over a decade. These projections rest on the assumption that the new workflows can expand. India’s Foundations: Substantial, Yet Oriented Toward the Old Model India has both ambition and policy tools. The national vision is to move from leadership by volume in generics, vaccines, and active pharmaceutical ingredients to leadership by value in biologics, biosimilars, complex generics, advanced therapies, and high-performance biomanufacturing, while growing the bioeconomy toward $300 billion by 2030 and substantially higher figures thereafter. Concrete strategies and resources include: Biopharma SHAKTI, a five-year program with a budget of ₹10,000 crore focused on biologics and biosimilars. It includes expansion of the National Institutes of Pharmaceutical Education and Research (NIPER) (three new institutes and upgrades to seven existing ones), creation of a network of 1,000 accredited clinical trial sites, strengthening of the Central Drugs Standard Control Organisation (CDSCO) with specialised scientific capacity, and support for discovery, development, and manufacturing infrastructure. The BioE3 Policy (Biotechnology for Economy, Environment and Employment), which targets high-performance biomanufacturing across six thematic areas, including precision biotherapeutics. Implementation centres on Bio-AI Hubs, Biofoundries, and Biomanufacturing Hubs, with significant public co-funding for private participation and multiple calls already supporting hundreds of projects. The ₹1 lakh crore Research, Development and Innovation scheme, managed by the Anusandhan National Research Foundation (ANRF) is designed to mobilise private capital for deep-tech projects at Technology Readiness Level 4 and above, explicitly including biotechnology, synthetic biology, pharmaceuticals, and AI applications. Early commitments through selected fund managers already run into thousands of crores. Complementary mechanisms such as the Promotion of Research and Innovation in Pharma & MedTech scheme, existing CSIR laboratory networks, DBT/BIRAC incubators and biofoundries (including a national network of biofoundries), GenomeIndia’s population-scale sequencing resource, a large manufacturing base with hundreds of USFDA-compliant facilities, and Global Capability Centres operated by many of the world’s leading life-sciences companies. India has hundreds of thousands of professionals across research, engineering, and manufacturing, plus a growing number of biotech startups. These resources give the country a strong foundation, but they are not yet organised around the singularity workflow. Most training still prepares scientists to be primary hypothesis generators and experimenters. Most laboratory infrastructure remains conventional rather than highly automated and closed-loop. Most funding and promotion metrics still reward traditional outputs. Clinical and translational capacity, while expanding, is not yet large enough for the volume of candidates that machine-scale discovery can generate. India is strengthening capacity within the classical model just as that model is being superseded. Fundamental Changes That Cannot Wait. Don’t Wait Until Rendered Obsolete If the diagnosis of an early singularity is even approximately correct, the required changes go beyond marginal efficiency improvements. They require restructuring how science is organised. First, the professional identity and training of the Indian scientist must change. Doctoral programs, NIPER curricula, CSIR and DBT institutional cultures, and evaluation systems need to treat the ability to direct large groups of AI agents and to operate or design automated laboratory workflows as core competencies. A scientist’s value will increasingly lie in the quality of the problems posed to the virtual army, the rigour with which outputs are examined, the efficiency with which automated experimental capacity is used, and the success with which discoveries are translated into practice. Promotion, tenure, and grant criteria should reflect these changes. Both institutional culture and technical practices must change, starting immediately in the institutions that train the next generation of researchers. Second, automated and self-driving laboratory capacity must be treated as critical national infrastructure. Virtual scientists without rapid, high-volume physical feedback loops produce limited real-world progress. A substantial portion of capital under Biopharma SHAKTI, BioE3 Biofoundry and Bio-AI Hub programs, and the RDI scheme should prioritise shared, high-throughput robotic platforms, integrated multi-omics and cellular assay systems, adaptive experimental design software, and nationally accessible automated facilities. The objective is enough experimental capacity for thousands of research groups to keep virtual scientists working productively. Scarcity of robot hours and automated wet-lab capacity will otherwise limit Indian scientific output. Third, clinical, regulatory, and translational infrastructure must expand in parallel. Faster preclinical discovery simply shifts the bottleneck to later stages if trial networks, ethics systems, regulatory review capacity, and manufacturing scale-up cannot absorb higher volumes of candidates. The planned expansion to 1,000 clinical trial sites and the strengthening of CDSCO are necessary but must be explicitly designed for higher-throughput, AI-originated pipelines, including greater use of hybrid and decentralised models, real-world evidence, and adaptive designs. Fourth, India must deliberately build sovereign advantages in data and models. Population diversity, clinical volume, and initiatives such as GenomeIndia are strategic resources. They should be converted into high-quality, privacy-preserving, multimodal datasets and domain-adapted foundation models that allow Indian-directed virtual scientists to outperform generic global models on problems of national relevance, including infectious disease, metabolic disorders, affordable advanced biologics, and regionally prevalent conditions. Bio-AI Hubs and public-private data collaborations should prioritise these datasets and models. Fifth, mission design, metrics, and governance across SHAKTI, BioE3, RDI, and related programs must be revised for the new workflow. Funding calls should score proposals on the scale of virtual scientific labour to be directed, the degree of laboratory automation, how closely the design–build–test–learn stages are linked, and the credibility of the path from discovery to practical use. Institutional performance indicators should reward increases in effective experimental capacity rather than solely traditional publication or patent counts. Public-private partnerships should broaden access to automated laboratories and the skills needed to direct AI agents rather than concentrate them in a few elite centres. Nuances, Risks, and the Cost of Gradualism Human insight remains relevant, and not every laboratory must be fully autonomous tomorrow. Biology is complex, and many experiments will still require skilled human hands and eyes. Safety, dual-use, and ethical considerations demand strong human oversight, while manufacturing and clinical requirements impose practical constraints. The transition will be uneven across disciplines and institutions. Premature or poorly governed automation could waste resources or introduce new failure modes. Data biases and model limitations must be actively managed. The greater risk, for the speed and scope, however, lies in gradualism. Treating AI as an incremental productivity tool while preserving the classical organisation of scientific labour will produce only modest gains and is a sure path to obsolescence. Meanwhile, leading groups elsewhere are already multiplying effective research capacity and shortening research cycles. The nations and institutions that first learn to direct virtual scientific armies and automated laboratories effectively will shape the next wave of therapeutics, agricultural solutions, industrial biotechnologies, and scientific tools. Those that delay will find themselves importing both the products and the platforms. India has the policy tools, skilled workforce, manufacturing base, and demographic scale to compete. It must now develop the organisation and infrastructure needed for the new scientific workflow. The scientist is no longer primarily the person who personally constructs and tests hypotheses. The scientist directs vast virtual workforces and keeps the robot laboratories running so that those workforces can deliver validated products and services faster than ever before. Making this transition is the main task facing India’s science ministries, NITI Aayog, DBT, DST, CSIR, ICMR, the leadership of Biopharma SHAKTI, BioE3, and the RDI scheme, and the directors of the nation’s research institutions. The singularity has begun, leaving little time to reorganise Indian science. Delay means accepting a widening gap by choosing not to adapt to the pace of effective scientific power.

Oct 21, 2026
Siddhesh Prabhugaonkar

UPI in the Era of Agentic Payments

UPI (Unified Payment Interface) is a groundbreaking innovation from India. It is a real-time payment protocol that enables billions of transactions for India’s massive population using merely a smartphone and a bank account. It is not only immensely popular but has also inspired similar technologies elsewhere, such as Brazil’s PIX and the USA’s FedNow. UPI regularly breaks its own records for maximum transactions per month, which currently stand at 20 billion per month. UPI is not just a financial innovation success story; it is also the world's most successful technology for financial inclusion. It has enabled 100s of millions of Indians who otherwise may not have access to make digital payments seamlessly. UPI’s success is not limited to India. Other countries have launched similar standards and they are also interoperable in some cases, such as Singapore’s PayNow and India’s UPI. As agentic commerce becomes more popular, payment methods that can operate in harmony with agentic AI interfaces are going to be increasingly more important. Traditional payment methods like credit cards are already adapting to this change. Both VISA and Mastercard have released their own SDKs for agentic commerce. Stablecoins are also popular in this context. This article, however, explores challenges and opportunities surrounding UPI for agentic commerce. What is Agentic Commerce? Agentic commerce is e-commerce driven by agentic AI, which can autonomously buy and sell items without the direct involvement of human beings. E-commerce traditionally relied heavily on humans for critical decision-making and, more specifically, payment-related decisions. This was because the traditional algorithms were not designed to deal with uncertainty and dynamic environments. However, modern AI is getting better and is at a stage where AI can suggest products to buy, and humans can trust it enough to let AI buy them automatically. Additionally, there is also a change in how users access information and shopping ideas. In the past, they relied on visiting different websites of merchants that showcased the products. Now it might be in an AI chat window. What are Agentic Payments? The unique nature of agentic commerce means there is no website where the user is expected to make the payment. Here, an AI algorithm can act on behalf of a user to make purchase decisions and also make payments. To do this, AI needs to have access to the user's payment methods such as credit cards and authorise charges on the user's behalf. Additionally, the entire process needs to have sufficient protection so that the AI does not overspend or get tricked into making payments the user might not have authorised. This is an evolving technology where many major players such as PayPal, Mastercard, Visa etc. have provided their own SDKs to AI developers and Google has created its Agentic Payment Protocol (AP2), an open protocol to manage agentic payments with credit cards and stablecoins. UPI with Agentic Payments One of the core features of the UPI payment system is that it relies on the user's smartphone and a bank account-linked SIM card. The real-time nature of payments also means the payments are a one-step process that cannot be cancelled once completed. Eventually, UPI also introduced UPI AutoPay and UPI Circle. Autopay is a mandate system where a merchant can charge the bank account for recurring payments. UPI Circle allows users to delegate access to other users (typically children or elderly family members) who can pay from their smartphone without having access to the bank account. Specialised constraints can be placed on such payments. UPI Circle, even if meant for other human users, has potential to also play a big role in agentic payments. Providing constrained access to specific AI Agents on a user’s behalf who can make full (payment that goes through immediately) or partial (where the account owner must approve the payment) AP2, which is emerging as a widely adopted industry standard for agentic payments, has UPI support on its roadmap even though it is not supported yet. AP2 uses a system of cryptographically signed Mandates (Intent, Cart, and Payment Mandates). These mandates provide the "trust layer" that allows an AI agent to prove it has your permission to initiate a UPI transaction. For UPI, the digital wallet will serve as the credential provider and the wallet will ultimately handle the final payment authorisation through the push payment model. The push payment model is where the user initiates the payment, which is different from the pull model where the merchant requests the payment. Razorpay, one of India’s major payment companies, also has unveiled its own agentic payment solution in partnership with the National Payments Corporation of India (NPCI), an Indian body that oversees the development of UPI, and ChatGPT. This solution uses UPI Circle and UPI Reserve Pay features to enable conversational AI agents to manage payments on behalf of human users through a human-in-the-loop system. Conclusion McKinsey estimates that global agentic commerce payments could reach up to $5T by 2030. India is a country growing economically and demographically and would be an important part of any technological progress. UPI, being India’s foundational payments infrastructure, has enough potential to benefit from agentic payments. Its UPI Circle infra already provides basic building blocks to achieve this and industry-leading protocols like AP2 are on its roadmap. Though the lack of credit card-like tokenisation and facilitation would certainly impact its potential, its wider reach and near ubiquity in India would mean it will overcome those shortcomings nonetheless.