Abhijit Dey has been a regular speaker at apidays Singapore for several years. When he asked me to review the book he co-authored with Srinivasan Shanmuganathan and David Roldán Martínez, I said yes — both because I trust his judgement and because AI in financial services is precisely what The Loop Asia exists to examine.
I want to be honest with you about how I read it. I tend to move quickly past high-level strategic introductions and go straight to the specific: what did teams actually try, what broke, and what fixed it. That is the lens I brought to this book, and it is the lens through which I think it is most useful. Read that way, it rewards the effort.
The case studies are the book
The core of "Building AI-Powered Financial Products" is a set of 58 recurring case studies built around a composite fintech called MoneyGro, running from Chapter 4 through to Chapter 15. Each follows the same structure: a specific AI failure in a production financial services context, what broke and by how much, the specific fix, and what changed afterward. A few examples, paraphrased:
A payment system with no contextual intelligence treats every transaction identically regardless of user history, device, or location. When a payment fails, it returns a generic error with no recovery path. Abandonment rates on declined transactions run above 60%. The fix redesigns around context-aware orchestration, progressive trust for verified users, and a failure-resilient flow where every decline produces a specific recovery path rather than a dead end.
An AML monitoring system generating over 90% false positives overwhelms the compliance team supposed to review them. Most alerts are never meaningfully investigated. The fix introduces dynamic behavioural baselines via machine learning, with a retraining loop fed by the compliance team's own confirm-and-reject decisions. The authors note plainly that getting false positives below 20% — the level mature programmes achieve — typically takes twelve to eighteen months of supervised feedback. Sub-50% is a starting point.
A generative AI customer service assistant is given an open-ended mandate to help with financial questions. Within two weeks it is generating investment and tax advice requiring a licensed advisor. Compliance shuts it down. The fix sets explicit scope boundaries enforced in the system prompt and in output filtering, with a hard redirect to human agents for anything outside scope.
None of these are edge cases. They are among the most common ways AI initiatives fail in production, documented with enough specificity to assess your own exposure.
The pattern worth naming
Across all 58 vignettes, a consistent pattern emerges: every failure is upstream of the AI model itself.
The payment abandonment problem was not a model problem — it was a missing feedback loop between transaction outcome data and the routing layer. The AML false positive problem was a data problem — static rules with no mechanism to learn from outcomes. The generative AI assistant failure was a governance problem — no scope definition, no output filtering, no escalation path.
In each case the problem existed before the model was deployed: in the data that fed it, the integration connecting it, or the governance that should have bounded it. The MoneyGro fixes reflect this: redesigning objective functions, building feedback loops, adding governed knowledge bases, shifting from static rules to adaptive ones. These are data, integration, and governance decisions — not model decisions.
This is the book's most important practical contribution.
How to read it
The book's most valuable content is in the MoneyGro case studies — 58 of them, each clearly labelled, each following the same pitfall-and-fix structure. Read those first, chapter by chapter, before engaging with the surrounding material. It takes about two hours and gives you a working taxonomy of production AI failure in financial services.
The most execution-dense chapters are 10 through 15, covering generative AI, roadmaps, team transformation, and autonomous AI agents. The agentic AI governance material in the final section is the most operationally specific I have encountered in practitioner-facing publishing.
Building AI-Powered Financial Products is published by Packt and is available on Amazon.
Blue Connector helps organisations build the data and integration foundations that make AI initiatives work. If you want to understand your readiness before committing budget: www.blueconnector.co
Conversations on technology leadership across Asia-Pacific
The Loop Asia covers AI, APIs, and digital transformation with practitioners doing the work. Subscribe for new episodes and articles.
Subscribe to The Loop Asia