I Built an AI Analyst Using My Own Bank Statements
I’ve been thinking a lot about what parts of my job I’d actually be happy to delegate to AI.
After 13 years at Google/YouTube in Finance & Analytics, I know the pattern well. For every strategic insight that shifts a decision, there are dozens of requests that look like this:
“What was our revenue last quarter?”
“Can you break that down by region?”
“How does that compare to last year?”
These questions are important. They enable decisions and surface trends. But answering them manually? That’s where analyst time goes to die.
So I wanted to see: could I build an AI that handles these recurring queries automatically?
Why I used my own bank statements
I wasn’t about to experiment with proprietary company data, so I used what I had: 5+ credit cards, 4+ bank accounts, multiple years of transactions.
The fundamentals are identical to corporate finance: structured data, recurring queries, need for accuracy.
Plus, if I screwed it up, I’d only embarrass myself in front of my husband, not my entire team.
The vision
An AI analyst you can talk to in plain English:
“What did I spend on groceries last month?”
“How does my restaurant spending this year compare to last year?”
“What subscriptions am I paying for?”
No SQL or even spreadsheet maneuvering required. No waiting for analyst bandwidth. Just answers.
The real test
After weeks of building, I let my husband try it.
Within five minutes, he asked: “What subscriptions are we paying for?”
The AI agent listed them all. Including three we’d completely forgotten about and didn’t need anymore.
We canceled them immediately. Saved $47/month.
What I learned
Building this taught me three critical things:
Why AI should write code, not do math
Why data cleaning is still 50% of the work (and always will be)
What privacy actually means when you’re sending financial data to LLMs
I’m going to break down each of these in detail over the next few weeks.
But here’s the punchline: this absolutely works. The technology to democratize data access exists. The gap isn’t the AI but it’s the implementation.
Finance teams have the data. They have the questions. They just need the translation layer.
Next week: I’ll share why my first attempt failed spectacularly, and the counterintuitive fix that made it work.
What parts of your analytics work feel most repetitive? Reply and let me know. I’m curious if the bottlenecks I’m seeing are universal.