Y Combinator published their latest request for startups, and one area they’d want to see startups from is called AI for Personal Finance. I make a personal finance tool. This post is a brainstorm where AI could enhance it, and the conclusion is it’s better if AI is involved on a higher level.
What
Personal finance tools as I understand them should be very different for people in different financial situations. E.g. what they say on the YC page about investing and tax optimization usually comes after you have your own contingency fund in place. I make a tool to help get there. It takes time and effort, and actions you likely only need to do on this stage. Roughly, this involves
- figuring out how you spend money,
- finding which parts, if any, you want to cut to put this money into the contingency fund,
- plan your budget and track how well you can stick to it,
- correct your budget, repeat until the fund is formed and your habits don’t need to change.
This is a bit like forming your eating and activity habits. You don’t need to count calories all the time, but you need to understand what you eat and how you should be acting.
Expense tracking apps usually give you a way to attach categories to your transactions and run some statistics on this info. Some apps also let you plan your spendings beforehand and see how close the plan is to the reality. Recognising behaviour patterns, planning, and plan adjustments are on you.
Have you tried rubbing AI on it?
Pattern recognition is AI’s second (or first) name. Even with very little info like an amount spent, category and time, behaviour patterns can be found. If you add calendar and location, more things could be derived. While it could recognise things that don’t exist aka hallucinate, you could quickly correct it and not follow recommendations that don’t make sense.
While people are different, there is probably a trick or fifty that will work on most, and there are probably behavioural antipatterns that are better be fixed. This reformulates the task from clustering to classification. People did this even before LLMs with all sorts of playbooks, especially management. Why not compiling such a playbook for everyday life and make AI looking for them?
Several rounds of budgeting and your habits enhancements follow. AI could help you tracking how well the budget is executed, what works and what doesn’t. With that in mind, a bit of time thinking about yourself and your situation and where you want to be, you’re going to figure it out. Hold on.
There is a profession for it
People are so bad in figuring out themselves they need specialists to help. Those specialists are called psychologists or therapists. And for them this information could be way more useful. Need to figure out patterns that make people unhappy? Here are some suggestions related to money stress, but also there could be activity charts and sleep schedule and diet summary. Oh, and calendar too.
I wonder if people go to therapists because their life is a mess but don’t ever mention how they eat or sleep or spend money because e.g. everyone lives like that. If therapists had a tool to access this info, and optionally find something that resembles certain behaviour, that might be more helpful than giving the same information to people themselves.
For that to happen, personal finance tools (and health can calendar and whatnot) need to export data to where it’s processed. Today every chat interface to LLM supports MCP which is a way to call external tools. Given there are apps that have MCP servers and a decent prompt, therapists probably don’t need another LLM wrapper, existing are enough. Which doesn’t mean those things won’t appear.
Could AI replace therapists? Eh, I’m not qualified to judge from either side, but if I project my programmer experience, AI copilots work way better than AI replacements so far.
Any implications for Within Means?
No promise. Within Means was intended as a tool that can be done and taken care of on weekends. Making an MCP server and seeing what different LLMs can derive from my behaviour looks doable in a weekend.
This post on hackernews: https://news.ycombinator.com/item?id=44217809
Mastodon: https://mastodon.online/@smagin/114648650668786649