At a Backbase Engage event in London, Marcus Martinez from Microsoft, Kanishka Bhattacharya from Bain & Company, and Dave Murphy from Publicis Sapient gave an unfiltered read on where AI actually stands inside banks today.

The verdict: AI is working, but only in specific places and for specific reasons. The banks leading in the AI era are the ones who understand the why behind AI, and not the ones chasing the broadest possible definition of "AI adoption."

This blog summarizes the main takeaways from the panel discussion, moderated by Backbase's CMO Tim Rutten. Click here to listen to the full conversation.

Where real AI impact in banking is happening right now

Ask what's actually in production, and the pattern is clear: generative AI is delivering the fastest returns in compliance monitoring, trade finance, and legacy code - anywhere banks have been drowning in documents, voice data, and manual review.

Marcus Martinez, Industry Advisor at Microsoft, called out two use cases gaining the most traction:

1) Complaints monitoring: regulation typically requires banks to sample around 3% of customer calls for quality and compliance review. With generative AI, banks can now analyze 100% of that data, catching issues that would never have surfaced in a 3% sample.

2) Cross-border trade finance compliance: checking transactions against the specific regulations of multiple jurisdictions simultaneously once took hours. With well-crafted prompts, it now takes seconds.

Dave Murphy, Head of Financial Services - EMEA & APAC at Publicis Sapient, added a less-discussed unlock: legacy code comprehension. Banks are sitting on core systems written in assembler, COBOL, Pick, and other languages most engineers have never encountered. Large language models can now read that code and translate it into human-readable functional specifications - giving CIOs a credible path to modernization that didn't exist a few years ago.

Murphy described this as an enormous unlock, because with that translation in hand, banks can finally begin addressing technical debt in a structured way.

Murphy also flagged a concrete signal of how AI is already changing delivery economics. The standard software development unit in banking today is an 8-10 person team. He expects that to compress to 3-4 people plus a set of software agents within 12 months - a projection he grounded in what's already happening inside a 24,000-person transformation firm.

Why AI adoption in banking keeps stalling

The honest answer to "why isn't AI scaling faster in banking" isn't the models. Kanishka Bhattacharya, Partner at Bain & Company, has an AI and machine learning background and teaches in the field. Based on his experience, the blockers are consistent across every institution he works with:

1) Leadership: A shortage of senior technology leadership who have actually built AI systems at scale - not theorized about them, built them.

2) Foundations: Weak data and technology foundations, the product of chronic underinvestment over the past two decades in both the platforms and the people who know how to use them.

3) Ownership: The inability to find individuals who can own the change end to end, from technical architecture through to business adoption.

That underinvestment doesn't evaporate because a new model gets announced. Bhattacharya's three blockers reflect a deeper structural reality across the industry. Fragmented records, siloed product systems, and inconsistent data definitions degrade model performance and produce outputs that neither business users nor regulators can trust. That's why so many well-funded AI pilots in banking never make it to production.

The temptation, Bhattacharya noted, is what he called "the lazy way out" - relying on off-the-shelf commercial AI stacks and calling it a strategy. His view: it's not just a technology problem, it's a business transformation problem. For organizations that haven't built a habit of systematic change, the complexity compounds rather than simplifies.

Continue reading: 7 reasons AI pilots fail in banking - and how to fix them

Where banks are actually deploying AI - and where they're holding back

Most live AI deployments in banking today are internally focused, such as back-office efficiency, compliance monitoring, and developer productivity. Banks are cautious about customer-facing deployment, and with good reason: regulatory clarity on non-deterministic agent behavior doesn't yet exist.

There still needs to be a human in the loop, and that's the right position for now. The picture, however, varies sharply by geography and risk appetite, as innovation tends to run faster in emerging markets.

Murphy described a client in Thailand that deployed outbound AI voice agents to engage customers with overdue loan payments - a move he said would face significant risk appetite headwinds from major UK banks. He referenced the CEO of Lloyds Banking Group, who has consistently pointed to East Asia as the place to watch first for customer-facing AI deployment.

Call centers are the clearest near-term opportunity for customer-facing AI. Martinez pointed to how quickly voice AI quality is improving - today it's genuinely hard to tell from a human. Some institutions already have AI agents listening alongside human agents in production, learning before going live.

On whether customers will accept being served by an AI agent: Martinez says it follows familiarity. Putting a credit card number into Amazon felt alarming in the late 1990s, and is now automatic.

The agentic banking question: how close is it really?

All three panelists agreed that AI agents - software that takes action on a customer's behalf, completes transactions, and negotiates with other systems - represent the next significant shift. The disagreement is on timing, not direction.

Bhattacharya grounded his view in what he's watching with e-commerce platforms - businesses that built their entire growth model around website and app experiences. He expects 20% of their transaction volume to shift to agent-mediated interactions by the end of next year, or within 18 months. His point: if it happens there first, banks face the same dynamic shortly after.

Murphy's framing cuts to the core governance challenge. AI agents making decisions without deterministic rules create explainability problems that regulators haven't finished working through. His position: keep a human in the loop now, build confidence in how agents make decisions, then extend autonomy progressively as that confidence is earned.

Scaling AI in banking responsibly: what actually works

Murphy was direct on how banks get the most out of AI: treating it as a smarter autocomplete delivers maybe 10-15% improvement. The real unlock comes from rethinking processes from the start, with agents making decisions rather than just assisting humans in making them. That requires genuine reinvention, not a patch on top of an existing workflow.

Bhattacharya's practical advice for organizations still finding their footing: map the end-to-end process first. The value of AI isn't individual productivity - it's the ability to redesign an entire workflow from intake to resolution, collapsing what currently requires logging into five, six, or seven different systems into a single governed flow. Until you harmonize the process, you won't see the real returns.

Martinez's advice was more personal: use AI to find the flaws in your own judgment. Run your proposals, your analyses, your decisions through a model, ask it to score them and explain why, then ask it to improve them. That habit, applied consistently across an organization, compounds faster than most leaders expect. He also made the case for prompt engineering as a genuine skill - drawing a direct parallel to knowing Excel in the early 2000s, which was a differentiating capability and is now just table stakes.

Continue reading: A decade of disruption and what comes next - insights from a global fintech leader

The real cost of AI in banking

McKinsey's Global Banking Annual Review 2025 finds that agentic AI could lower operational costs by 20% or more, equivalent to 9-15% of operating profits - though most of those gains will eventually be competed away as the technology becomes standard across the industry.

Running AI models isn't where the money goes. Bhattacharya broke it down plainly: a human agent call costs $20-$30 to handle. A well-designed AI interaction costs around $0.20-$0.40, and resolves the query around 60-70% of the time. The cost of the technology, in his view, is the smallest item on the bill.

The structural costs are talent, organizational change, and leadership capable of driving the transformation. Engineers who've built AI systems at scale are scarce, and every employer is competing for the same people. Those costs won't compress quickly, and banks that start building that capability now are buying time that later movers will struggle to recover.

What separating AI hype from reality actually requires

The panel's closing message was consistent.

Bhattacharya: start with something simple, do it end to end, and do it with urgency. Pace matters more than perfection at the start.

Murphy: if it feels like an assistant making marginal improvements, you're missing the point - lean into reinvention.

Martinez: be curious, learn to prompt well, and use AI to challenge your own thinking rather than just validate it.

The banks figuring this out early won't just run more efficiently. They'll build learning systems that compound over time. The architecture you put in place now determines what's possible two years from now, and that's a strategy argument as much as a technology one.

This blog summarizes the main takeaways from the panel discussion, moderated by Backbase's CMO Tim Rutten. Click here to listen to the full conversation.