$340 bllion annual growth, $23 trillion at risk: Globally banks stuck in AI’s demo room

Ninety-four point one per cent of India's banking and financial services firms claim to use artificial intelligence to improve operational efficiency. Only 19.1 per cent of those same firms track whether it is actually contributing to revenue. That gap - between adoption and accountability - is not an Indian problem. It is a global one. And it sits at the centre of what may be the most consequential technology deployment in the history of modern finance.
Globally, one in every four banks is actively using AI to gain competitive advantage, with the rest still marooned in fragmented, inconclusive pilots, according to the Boston Consulting Group, cited in a 2026 Backbase industry report. The World Economic Forum projects total AI spending in financial services will reach $97 billion by 2027. McKinsey puts the potential annual value addition to global banking at $340 billion. And yet, the majority of the world's financial institutions cannot demonstrate - in measurable terms - what their AI investments are actually delivering.
The revolution, it turns out, is real. Proving its value is another matter entirely.

THE MEASUREMENT PARADOX

Take a look at India's BFSI sector, that has moved faster on AI adoption than almost any other regulated industry globally. Seventy-eight per cent of Indian financial firms are already exploring generative AI tools. Forty-seven per cent have multiple AI use cases live in production, according to a November 2025 report by EY and the Confederation of Indian Industry. AI-related mentions in private sector bank annual reports grew six-fold between FY2015-16 and FY2022-23. In public sector banks, the growth was threefold - significant, but half the pace, constrained by slower procurement cycles and tighter compliance frameworks, according to the RBI's October 2024 Bulletin. By June 2023, eleven of twelve public sector banks had deployed AI-powered customer service chatbots, per a 2025 analysis in the International Journal for Multidisciplinary Research.
But the measurement data tells a structurally different story. Only 19.1 per cent of BFSI firms track AI's contribution to revenue. More than a third - 36.8 per cent - do not measure AI's revenue impact at all. And 44.1 per cent are unsure whether such tracking even exists within their own organisations, according to a March 2026 survey of 32 firms covering approximately 1.3 lakh employees by KNOLSKAPE. Only 47.1 per cent measure return on investment from AI initiatives. Only 57.4 per cent link AI learning programmes to business performance outcomes.
The workforce powering these deployments is also largely unprepared. Only 31 per cent of India's BFSI workforce is ready for the digital and AI transition, the same KNOLSKAPE survey found. The sector's top workforce development priority for 2026 is leadership capability, followed immediately by AI readiness - in an industry simultaneously claiming 94.1 per cent AI adoption. That is not an irony. That is a warning sign.

THE DEMO ROOM PROBLEM - WHY PILOTS STALL

Understanding why the gap between adoption and accountability persists requires looking inside the institutions attempting the transition. Rohit Yadava, Chief Operating Officer of enterprise AI firm Aziro, has worked directly within BFSI and audit deployments and identifies the failure point precisely.
“In live BFSI and audit environments, AI agents are being used best in narrow, rules-influenced workflows where there is high volume, repetitive review, and a need for faster turnaround,” Rohit said. “In lending, agents can extract borrower data from documents, cross-check it against internal systems, identify missing information, and prepare a case summary for an underwriter. In claims, they can classify claim types, detect mismatches across forms and supporting evidence, and route exceptions for manual review. In audit functions, agents are useful for collecting evidence, mapping records to controls, summarising deviations, and helping teams prioritise what to investigate first.”
The capability, Rohit Yadav is clear, already exists. What fails is the infrastructure around it.
“The shift from pilots to real deployments happens when enterprises stop asking whether the model works and start asking whether the workflow works,” he said. “In the pilot stage, teams often test AI on sample data, isolated tasks, or controlled prompts. That is useful, but it does not reflect production reality. Real deployment means the AI has to work with messy inputs, variable document quality, system dependencies, approval logic, and business accountability.”
He identifies five non-negotiable requirements for reliable AI deployment at scale: clean integration with enterprise systems including loan platforms, ERP tools, document stores, and audit infrastructure; strong data grounding so agents operate from approved internal sources rather than generic model memory; orchestration logic defining what the agent can do, when to seek additional information, and when to escalate to a human; full observability covering logs, prompts, outputs, latency tracking, and escalation points; and governance encompassing access control, versioning, compliance capability, and rollback.
“At scale, the real challenge is not generating responses,” he added. “It is maintaining consistency, traceability, and acceptable performance across thousands of transactions and edge cases.”

THE TRUST GAP

Sajal Gupta, Co-founder and Director of AI infrastructure firm Fexo, frames the structural problem in terms of what he calls an existential requirement - not a technology choice.
“For the BFSI and audit sectors, AI is not a luxury - it is an existential requirement,” Sajal Gupta said. “Finance, especially in India, is now governed by thousands of distinct document formats - from Annual Reports, GST and ITRs to varied bank statements to internal databases to regulatory policy documents - all of which must be reconciled. Relying on human reviewers for millions of data points leads to fatigue-driven errors, which, in a regulated environment, are viewed as a legal catastrophe."
He points to the market pressure compounding that internal failure risk. “While traditional due diligence used to take weeks, the market now demands that initial assessments happen in hours or days to remain competitive,” he said. “Generative AI is expected to add up to $340 billion in annual value to the global banking sector. The transition to AI agents is no longer a fancy technology to adopt - it is the new imperative required to keep costs low enough to compete with digital-first fintechs.”
The reason most institutions are failing to make that transition, Sajal Gupta argues, is that they are attempting it with the wrong tools. “The last two years were mostly experimentation and piloting of AI in different workflows and use cases,” he said. “Now, a transition from experimental AI pilots to agentic ecosystems is in progress - moving away from simple text generation toward autonomous, multi-step workflow execution.”
Bridging what he calls the Trust Gap - the distance between a system that performs in a demo and one that can be relied upon in a regulated workflow - requires four specific architectural layers. First, a unified, governed data foundation with automated quality checks and digital provenance tracking, so that every data point's origin, transformation, and usage is auditable. Second, a shift from large cloud-based language models toward smaller, specialised models - typically under 14 billion parameters - that run fully on-premise, keeping sensitive data within institutional firewalls and complying with the RBI framework, GDPR, and the EU AI Act. Third, a multi-agent orchestration mesh in which supervisor agents coordinate specialised sub-agents handling distinct tasks such as AML investigation, treasury functions, or retail support. Fourth, deterministic guardrails - what Gupta describes as a Triangulation Engine - that act as a compliance check on probabilistic AI outputs before they reach a regulated decision.
“The Trust Gap is the primary barrier preventing experimental AI from becoming operational reality in regulated finance,” Gupta said. “Production systems use a Triangulation Engine - deterministic rules acting as a guardian, ensuring every output is auditable and legally defensible. Organisations that use specialised evaluation and governance tools move nearly six times more AI systems to production.”

INDIA'S REGULATORY ARCHITECTURE

India's financial regulator has moved to get ahead of the deployment wave. In August 2025, the Reserve Bank of India’s Committee on Financial Sector Regulation of Emerging and Evolving AI - the FREE-AI Committee - published India’s first comprehensive AI governance blueprint for the financial sector, addressing IT outsourcing risk, technology governance, and operational resilience.
The framework identifies four AI risk categories: model hallucinations, cybersecurity vulnerabilities, data privacy breaches, and threats to systemic financial stability. It mandates a Right to Explanation - requiring that any automated financial decision carry a plain-language justification accessible to the consumer, a provision Sajal Gupta specifically identifies as something only generative AI can reliably produce at scale. The framework also requires 100 per cent transaction monitoring and continuous reporting, formally ending sampling-based compliance. KPMG's September 2025 analysis noted India had established binding AI governance in financial services ahead of many developed markets.

THE GLOBAL RACE

The competitive context gives the urgency a sharper edge. In the United States, generative AI adoption in commercial banking surged from 10 per cent in production in 2023 to 47 per cent by 2025 - a near five-fold acceleration in two years. US AI banking spend reached $19 billion in 2024, forecast to hit $54 billion by 2028.
China has moved at even greater scale. The country's core AI industry reached 1.2 trillion yuan - approximately $173.9 billion - in 2025, with over 6,200 active AI companies, according to China's Ministry of Industry and Information Technology in a March 2026 release. Total AI investment reached between $84 billion and $98 billion in 2025, up 48 per cent year-on-year, per Bank of America analysis. China's AI banking automation market stood at $3.43 billion in 2025, projected to reach $23.8 billion by 2033, according to Grand View Research. But up to 80 per cent of China's newly built AI compute infrastructure was reportedly idle as of March 2025 - a striking overcapacity concern flagged explicitly by the United States Federal Reserve in an October 2025 report.
Europe has chosen governance over velocity. The EU AI Act - the world’s first comprehensive AI regulation - came into force in 2025, with prohibited systems banned from February. In November 2025, the European Parliament voted 426 to 182 against introducing new AI-specific financial regulations, directing regulators to apply existing law instead. The European Parliament's ECON Committee identified the same four AI risk categories as India's RBI: hallucinations, cybersecurity, data privacy, and systemic stability - a convergence that signals emerging global regulatory consensus.
India sits in a position that is simultaneously promising and precarious. It has headline adoption numbers that match or exceed the United States. It has a binding regulatory framework that rivals Europe's in ambition. It has practitioners deploying agentic systems in live lending, audit, and claims environments right now. But it carries an 81 per cent measurement gap, a 31 per cent workforce readiness figure, and a public sector banking system advancing at half the pace of its private counterparts.

THE $23 TRILLION QUESTION

McKinsey's Global Banking Annual Review 2025 put a precise number on what is at stake for any institution that fails to close that gap. Agentic AI could disrupt $23 trillion in global consumer deposits currently sitting in near-zero rate accounts, enabling digital-first challengers to compete directly for retail deposits that traditional banks have long treated as permanently captive.
Venture investment by banks into AI companies declined 17 per cent in 2025, according to the Evident Insights Banking AI Index - a recalibration from hype-driven spending toward disciplined deployment. But the leader-laggard divide is accelerating: the top ten banks on the Evident AI Index increased their AI scores at more than double the rate of the average institution in 2025. AI-focused product launches across more than 10,000 banks and fintechs between late 2022 and mid-2025 represented only approximately three per cent of all product launches globally, per McKinsey Panorama - suggesting that even within an industry claiming transformation, the actual deployment footprint remains narrow.
Rohit Yadava's framing may be the most honest summary of where the industry actually stands. “Trust is built when the AI becomes predictable to operate,” he said. “Enterprises adopt AI more confidently when it behaves like a governed system inside the process - not like an impressive but loosely controlled layer on top of it.”
That is the distance most of the world’s banks still have to travel. The money is committed. The technology is capable. The regulatory frameworks are arriving. What remains is the harder work - making AI not just deployable, but provable, measurable, and trusted at the scale the $340 billion promise actually requires.