This deep dive expands on the Loop Asia conversation with Abhijit Dey, VP of API Banking at Axis Bank.

Abhijit Dey has a second body of work sitting alongside his job as VP of API Banking at Axis Bank: he's a co-author of Building AI-Powered Financial Products, with Srinivasan Shanmuganathan and David Roldán Martínez. The book's method is unusual for the category — instead of arguing strategy, it documents 58 production AI failures inside a fictional fintech, MoneyGro, each one following the same structure: a specific failure, its quantified business impact, and a targeted fix. A payment system with no contextual intelligence. An AML monitoring system generating so many false positives that real alerts get lost in the noise. A generative AI assistant operating without defined scope boundaries, answering questions it was never supposed to touch. I reviewed it in full here.

The detail worth sitting with is what's absent from that list. None of the 58 failures are about the model being insufficiently capable. Every one of them happens upstream — in the data pipeline feeding the model, in the system integration around it, or in the governance structure that was supposed to catch the failure before a customer did.

The same argument, one layer up

Anyone who's read the API sprawl section of this conversation will recognise the shape of that argument immediately, because it's the same one Dey makes about his own API estate. Sprawl, in his telling, "isn't a failure mode. It's a success mode that went unmanaged" — the direct consequence of teams shipping fast under go-to-market pressure, each individual decision defensible, the aggregate ungoverned. Nobody sets out to build an unaccountable AI feature any more than anybody sets out to build ten thousand APIs covering seven thousand use cases. It happens because the discipline that would have caught it — a live inventory, an honest audit of what's actually in production, a standing willingness to say no — was never built at the same pace as the shipping.

That's not a coincidence of Dey's interests. It's the same execution discipline applied twice, once to the API layer he's spent 15 years building and once to the AI layer now getting bolted on top of it. The book's implicit claim is that AI failure in financial services isn't a new category of problem requiring new governance thinking. It's the old problem — unmanaged growth outrunning the accountability structure — showing up again, faster and with higher stakes, because the thing now making autonomous decisions is a model instead of a hand-written rule.

What this means before you deploy

For a Technical Champion evaluating where to put governance effort on an AI initiative, the MoneyGro case studies are a useful checklist precisely because they're not abstract: contextual intelligence gaps, alert-fatigue failure modes, scope-boundary failures in generative assistants are all things you can audit for in your own stack before they surface in production, the same way Dey audits his API estate against the ratio of APIs to actual use cases rather than waiting for the sprawl to become unmanageable.

For a Business Driver holding an AI mandate with no existing governance framework to point to, the reframe is just as useful. The question a board should be asking isn't "is our model good enough" — it's the question Dey has been answering about APIs for a decade: do we have a live, honest account of what's actually running, who's accountable for it, and the discipline to fix or retire what isn't earning its place. That's a governance and integration question before it's ever a model-selection question, and it's exactly where Blue Connector spends its time with clients who have an AI mandate and no roadmap yet for the infrastructure underneath it.