Evolution Exchange Finance Podcast

The episode surfaces a handful of genuinely useful practitioner observations - human-in-the-loop as a design parameter rather than a fallback, the semantic model gap between physical database attributes and business terminology, and prompt-change governance as an unresolved risk - but most of the runtime is consumed by agreement-chaining and generic statements about data quality and trust that add little for an experienced B2B operator.

“the right framing is like this human in the loop is a design parameter, and it is not a fallback”

“teams that succeed, they build the evals before they build the features they design for the edge cases and the hard cases”

The mild reframing that traditional ML has been succeeding for years while generative AI is the new failure surface is a useful distinction, and the prompt-change governance question in regulated environments is underexplored in most AI discourse, but the bulk of the conversation recycles standard AI-implementation warnings with no contrarian or first-principles angle.

“we've been running machine learning motors for like decades now... Last decade maybe we started with 10 parameters or something 10 years ago and now we are up to a thousand”

“whenever there is a change in the system prompt, who has to sign off this prompt change? Does this change in the system prompt trigger a new model risk validation?”

All three guests are genuine domain practitioners - a lead data and AI engineer at a bank with 20 years in IT, a product leader with eight years in financial services, and a data-pipeline department head at a large Norwegian insurer - giving the conversation credible grounding, though none are in C-suite roles or widely recognised as domain authorities.

“I'm working as a lead data and AI engineer at Nadia Bank. I have been working in the field of data and machine learning and data science for the past 15 years”

“the last 20 years or so I've been working in the data area with the data pipeline... for one of the largest insurances in Norway”

The episode offers isolated concrete touches - the 30% concentration-risk query example, the AUM/database-attribute naming problem, and legacy code written 25 - 30 years ago with retired authors - but there are no hard metrics, no named projects, no outcome data, and most claims stay at the level of 'many organizations are facing this right now.'

“show me the clients with the concentration risk above 30% in technology sector”

“you could call assets under management as AUM or the total assets under management... the underlying attributes on the database side could be named in a different way”

The host moves the conversation through a reasonable structure and occasionally sets up follow-on topics well, but questions are broad and pre-announced ('another area of failure'), there is almost no pushback on vague claims, and the panellists spend most of the episode agreeing with each other, leaving productive tension unexplored.

“what's fundamentally different between a demo environment and uh, live production?”

“are banks measuring the right things?”