> ??! You've got a system of unpredictable reliability and you're going to let it handle money?

Yes. Recall phone banking? I imagine it like that. Potentially via a chat service like Whatsapp or signal.

You'd start with simple standard tasks and low limits, if course. But the risk and value can be directly quantified in online banking.

I remember how phone banking relied on the honesty of individual, accountable employees. How do you hold AI accountable? You give it low limits, write off the mistakes and keep scaling and ignoring failure even at the million-dollar-mistake level?

Banks indeed look at things in terms of risk mitigation, but this is silly. You will not see a credible bank incorporate LLM-based features because it is a genuinely enormous attack surface with marginal ROI. Any logical developer would create an ELIZA-style on-rails system instead, and it would never misbehave like ChatGPT.

Some banking still runs on COBOL, probably.