AI’s Diminishing Returns : Avoiding the Overreliance Trap in BFSI

“I find it hard to see, how there can be a good return on investment given, the current math.”

This warning about AI infrastructure economics is not just for tech investors, but it has a direct bearing on banking and financial services, where the current wave of AI adoption risks building systemic fragility under the banner of innovation.

Unlike the earlier generation of AI, which was often developed in-house, fine-tuned on proprietary datasets, and deployed within the secure boundaries of a bank’s own infrastructure, today’s wave of general-purpose AI is of a very different breed. At the centre of this shift are large language models (LLMs) and similar foundation models, which are not built or run locally but unleashed at scale through centralised data centres and software stacks maintained by a handful of technology providers. Every solution built on such infrastructure, be it a chatbot, a fraud detection engine, or a compliance tool, ultimately depends on continuous calls to these central data centres. The implication is that future pricing models will inevitably evolve toward a per-transaction basis, much like how the industry already prices microservices and API-based services.

This distinction is critical for this article because, while traditional AI brought efficiencies without external dependency, general-purpose AI embeds a structural reliance on third-party infrastructure, creating new vectors of systemic risk.

Today, banks and asset managers are racing to reimagine everything, from client servicing to risk management through the lens of general purpose AI. Credit scoring, compliance, fraud detection, portfolio management, back offices in capital markets, and treasury operations, everywhere AI is being plugged in as the new nervous system. The problem is not adoption. The problem is  scale-first adoption, as I will make an attempt to explain.

The Hidden Cost Curve

AI today is being consumed at heavily subsidized prices. The cost of infrastructure and model development is not yet fully passed on to customers — banks included. And the numbers behind the curtain are staggering:

• OpenAI is projected to burn $115 billion through 2029, with annual spend crossing $45 billion by 2028.
• Microsoft, Google, Meta, and Amazon are already spending $20–46 billion each year on AI infrastructure costs, absorbed today, but not indefinitely.
• By 2030, McKinsey estimates $5.2 trillion will be required just for AI optimized data centers, out of a total $6.7 trillion global data center spend.

For perspective, Netflix earns $39 billion a year from 300 million subscribers. At true infrastructure cost, if AI companies start  to charge Netflix of the true costs they invur, they will need almost 3.69 billion payying users to break even on compute. That's nearly half the planet. Now translate that to banking, where global technology spend is $300 to 350 billion a year, and the economic strain becomes obvious. If providers begin passing back even a fraction of their trillion-dollar infrastructure costs, banking margins will come under unprecedented pressure.

The Risk of No Reversal Strategy

Banks historically built or bought their systems - enterprise applications tying front to back, middlewares, proprietary cores, CRM portals etc. Costly, yes, but amortized across decades, predictable, and within their control.

In today’s AI rush, legacy is being decommissioned wholesale. In its place, banks are embedding themselves in AI ecosystems where:

• Exit Strategy is Lost. Once legacy systems are dismantled, there is no easy return.
• AI is Perpetual Rent. Every query, every model call, every retraining cycle incurs an operating cost, priced in proportion to data center and GPU consumption.
• Margins Erode. With data center power demand projected to nearly double to 945 TWh by 2030, costs will rise not just from technology, but from energy and regulation too.

The trap is clear. Banks may soon face an environment where revenue growth plateaus but infrastructure bills compound, leaving no room to claw back profitability. In other words, over reliance on AI to replace complete systems now may backfire when AI service providers start charging for infrastructure and software cost in direct proportion to their spending. Today they don't charge. They are in a customer acquisition spree. A pricing strategy known as market penetration pricing. However once that period is over they will start charging the cost back, else they can't sustain. By that time if most legacy systems have been replaced by AI, the cost of maintaining high cost AI systems with no reversal strategy will lead to lower margins of banks.

The BFSI Lens

To appreciate the systemic exposure, consider how today AI reliance is playing out across different banking verticals:

Retail Banking:

AI is contemplated to become the core engine behind personalized offers, digital onboarding, fraud detection, and customer service chatbots and as well open banking. If the inference cost rises, say from a few cents to several dollars, every customer interaction becomes more expensive. At scale, with millions of daily queries, the economics of retail banking could swing from cost-efficiency to cost burden. Margins in mass-market retail are already razor-thin; absorbing AI surcharges may mean scaling back digital personalization itself.

Corporate Banking:

Corporate banking thrives on relationship management, credit facilities, trade finance, and treasury solutions. AI is being deployed at scale in credit analytics, real-time risk monitoring, and client servicing. But these are compute-heavy applications involving massive data sets. As infrastructure costs rise, banks may face a paradox - scaling AI to improve client experience at the cost of eroding profitability per corporate relationship, especially when corporates demand competitive pricing.

Investment Banking

Here, AI is making waves in deal sourcing, valuation modeling, trading strategies, compliance, portfolio management and as well back office. But investment banking economics are highly sensitive to cost of capital. Spending millions annually on AI driven trading algorithms or compliance engines will only be justified if returns outpace costs. The risk is that the “arms race” in AI-driven trading forces every investment bank to spend heavily, not for competitive advantage, but just to keep parity. A textbook case of margin compression.

Transaction Banking

The plumbing of global finance - payments, clearing and trade finance, is increasingly embedding AI for fraud detection, sanctions screening, and exception handling. These are high volume businesses. A rise in per-transaction AI cost, multiplied across billions of transactions, could quickly wipe out margins. Transaction banking depends on efficiency at scale. If AI costs scale linearly with volume, profitability may erode at precisely the moment when clients demand lower fees and faster settlement.

History Rhymes

We have seen this before.

• In 2000, trillions were sunk into fiber optic cables before demand caught up.
• In 2014, shale drillers pursued volume at the expense of economics. The US primarily was chasing shale exploration disrupting oil pricing in the middle east. But now no more, because it was soon evident that the cost refining shale oil laden with sand is economically a disaster.
• Not so long back, banking witnessed a strikingly similar arc with the adoption of Robotic Process Automation. The mantra was simple: automate whatever could be automated. What was missing, however, was the deeper question of what should be automated once the cost-to-value equation was fully considered. In the rush to scale, banks soon realised that the run costs of maintaining armies of bots often exceeded the original cost of the processes themselves. Ironically, what had begun as a quest for efficiency ended in diseconomy. The correction came later, with institutions quietly running “bot decommissioning” programmes. An attempt to unwind what had once been heralded as transformation.
• A parallel is unfolding even now with migration to cloud at scale. While early business cases were often built on the promise of cost savings, the narrative today has shifted. Banks now find themselves justifying the move on grounds of agility, scalability, and resilience—anything but cost. The paradox is telling. Unlike RPA, where decommissioning could reverse the excesses, moving off cloud is far more difficult. Once physical infrastructure is wound down and data centres shuttered, the exit door narrows. What began as a cost play now locks institutions into a structural dependency, with little room for retreat. Each followed the same script. Adoption was mistaken for profitability, and scale was mistaken for sustainability. AI in BFSI today risks becoming the next verse in that rhyme

The Banking Parallel

Global banks currently generate an average 11-12% RoE (Bain, 2023). But when you line that against AI’s future cost curve, the fragility shows.

By 2027, training a single state-of-the-art model could exceed $1 billion. By 2028, $76 billion annually will be needed just to operate generative AI data center servers. AI companies, as Sequoia Capital has warned, may need $600 billion in annual revenues simply to justify their infrastructure investments.

Where will that revenue come from? From industries most reliant on AI. BFSI sits at the top of that list. If banks have fully replaced legacy systems with AI by then, they will bear the brunt, forced to absorb rising infrastructure costs while margins sink below cost of capital.

A Smarter Path Forward

The way out is not abandoning AI, but it is abandoning scale-first AI approach. Instead the BFSI sector should prioritize:

1. Adaptation-first systems: Lean AI tuned to specific solutions (e.g., fraud detection, AML checks, KYC workflows etc) rather than brute-force general intelligence AI where the complete dependency shifts to AI data centers.
2. Hybrid architectures: Retain critical legacy cores, layering AI selectively to preserve optionality.
3. Cost-to-revenue discipline: Deploy AI where there is measurable payback in risk reduction, compliance savings, or new business. Not for vanity transformation.
4. Energy-efficient partners: Collaborate with providers who build sustainable architectures, not just scale barns.

In summary, AI will undoubtedly shape the future of banking. But unless adoption is grounded in economics, the sector risks replacing one form of legacy lock-in with another. Only far more expensive, and far harder to unwind.

Over time, the inevitable question will arise as who bears this burden?

Finally, from Banks to Customers: The Silent Cost Transfer

The final, and perhaps most overlooked, dimension of this debate is the potential transfer of AI costs downstream to customers. If banks find themselves locked into expensive, transaction-based AI infrastructure, history suggests that much of this burden will not remain on balance sheets alone. Inevitably, it will seep into pricing through higher transaction fees, convenience charges, or subtle cost adjustments spread across millions of accounts.

And here lies the subtle paradox. Think of our own behaviour. How often do we pause to check the platform fee or the convenience fee for quick commerce? In many ways, our choices are now conditioned more by convenience than by rational calculation. Would banking customers behave any differently? Would they accept the hidden cost of AI, folded neatly into transaction charges and service fees, simply because the experience feels seamless?

This possibility is may be, the industry’s brightest silver lining. The risk of thinner margins, being invisibly transferred from institutions to customers, in exchange for convenience that few will stop to question.