I work in AI in financial services. I’ve built traditional ML systems — fraud detection, AML, credit scoring — and I’ve spent the last two years evaluating where gen AI fits. The honest answer is: not as many places as the vendor slides suggest.
Here’s the list of industries where I’d say gen AI is genuinely transformative, not just incrementally useful:
Software development. This is the clearest case. A junior developer with Cursor or Copilot today produces what a mid-level developer produced two years ago. It’s not a demo. It’s how software actually gets written now.
Content and media. Entire categories of freelance work have been repriced. Marketing copy, image generation, first-draft video. Whether this is “transformative” or just “cheaper” is debatable, but the economics have genuinely shifted.
Legal. Document review, contract analysis, case research. Work that took associates hundreds of hours now takes minutes. Law firms are restructuring around this.
Education. A good LLM tutor is better than no tutor, and most people had no tutor. The accessibility gain is real.
That’s roughly the list. Notice the pattern — gen AI is transformative where the work is language-heavy, judgment-light, and high-volume. Summarise this, draft that, find the needle in this haystack of text.
Now here’s the uncomfortable part for my industry. Financial services is where the work requires domain-specific reasoning with real consequences. And that’s exactly where gen AI is weakest. The FS use cases I see deployed right now are mostly efficiency plays — customer service chatbots that are marginally better than the old ones, document summarisation that saves time but isn’t transformative, “copilots” for relationship managers that are mostly demos.
The genuinely promising FS use cases — regulatory compliance automation, institutional knowledge management, decision support for complex judgment calls — are real but early. Nobody knows yet which will stick and which will be remembered like blockchain-in-banking circa 2018.
Meanwhile, the traditional ML systems I built years ago — the fraud models, the AML alert prioritisation, the credit scoring — are still the workhorses. Still in production. Still catching real criminals and making real decisions. They’re not glamorous, but they work.
The gap between “gen AI demo” and “gen AI in production at a bank” is enormous, and it’s filled with regulation, explainability requirements, data governance, and the simple fact that banks can’t afford to be wrong in the ways LLMs are sometimes wrong.
My job isn’t to sell gen AI. It’s to know which use cases are real and which are theatre. That distinction — the honest assessment of what actually works — is, ironically, the most valuable thing an AI practitioner can offer right now. Everyone can build a demo. Almost nobody can tell you whether it should go to production.