Actually I really do like the ideas of using AI to do banking. It irks me incredibly having to look through an email / invoice / WhatsApp for
- the bank account (and then being forced to type in 7 - 11 digits in a continuous fashion because banking apps can't just ignore spaces -
https://www.quora.com/Why-is-it-eas...-of-three-numbers-than-others-longer-actually)
- what the guy prefers to be called vs the business and then the app cuts off most of the name
- the correct reference (invoice number? Account number? My name? My surname? My son's name? My son's name and surname and year and grade but oh then the reference is too long what now fine you guys asked for it your accounting can just handle it)
- finding out where to send out a proof of payment
- double-checking the amount a million times
- accepting the "risks" (looking at you FNB) of paying a third party and thus having to wait 30s on phone to open the FNB app and log in and then I have to acknowledge twice that I understand I'm paying a third party and then having scanning a QR code for some reason.
The idea of telling my banking 'agent' to pay the invoice sounds pretty good.
Security risks with AI are, of course, there - e.g. prompt injections or jailbreaking or an agent breaking its boundaries or of course simply the idea that I'm sharing my financial information with a third party - I doubt Discovery trained their own LLMs.
They may be using open weight models in a third party inference - heck they may even have their own AI cluster running - but unfortunately the article doesn't mention this. They may also just be calling the OpenAI / Anthropic / AWS Bedrock / Azure / Gemini APIs.
Would be nice to find out more about how they did it,
@Jan.