My AI Finance Agent Experiment

Miguel A Rios -

Why I Built a Personal Finance AI Agent

Let me be honest—I wanted to learn about this AI agent hype, and I needed a practical use case that would actually benefit me. For the past few years, I've been managing my finances using Tiller, which feeds all my bank transactions into Google Sheets. It's been great for tracking my expenses and understanding my savings rate to help me meet my financial goals.

I wanted to take my finances to the next level and get actionable insights from all this data. I kept thinking: what if I could build something that not only analyzed my spending patterns but actually told me what to focus on each month? That's when I decided to dive into building my first AI agent.

What My Finance AI Agent Actually Does

Here's the thing—my agent doesn't try to do everything. It has one clear job: analyze my transactions and give me 3 SMART goals to focus on for the current month. That's it. Simple, actionable, and valuable.

But here's what makes it special: it has autonomy to decide if and when it should alert me based on the data. It's not just running on a schedule sending me the same report every time. It's actually thinking about whether the insights are worth my attention.

How It Works Behind the Scenes

The process is surprisingly straightforward:

  1. Data Collection: It automatically pulls my Tiller transaction data directly from Google Sheets
  2. Analysis: It calculates key financial metrics based on my transaction patterns—things like category spending trends, unusual purchases, and budget variance
  3. Intelligence: It feeds this processed data to Claude Sonnet 4 with carefully crafted prompts designed to generate actionable insights
  4. Memory: It keeps track of previous notifications to avoid sending me the same advice repeatedly
  5. Decision Making: Here's the key part—it makes autonomous decisions about when insights are actually worth sharing with me

The agent doesn't just mindlessly process data. It evaluates whether what it found is significant enough to interrupt my day.

Technical Architecture (And Why I Chose It)

I built this entire system on n8n, a low-code/no-code platform that I'm self-hosting on my personal server. Here's why I made this choice.

I wanted to learn AI agent fundamentals before diving deep into LangChain, LangGraph, or building something from scratch. There's no doubt in my mind that those other tools are probably better for scale and security, but I needed a working prototype that would teach me the basics while delivering real value.

Here's my current tech stack:

  • Platform: n8n (self-hosted on my personal server, though the cloud version would work fine)
  • LLM: Claude Sonnet 4 for all the reasoning and insight generation
  • Database: Google Sheets (keeping it simple and leveraging my existing Tiller setup)
  • Trigger: Currently a scheduled cron job, but I'm planning to add the ability for my agent to schedule itself to run when it sees fit

The beauty of n8n is that I could focus on the agent logic and prompting rather than getting bogged down in infrastructure. I got a working prototype in days, not weeks.

The Real Challenges (And Expensive Lessons)

Cost Control & Token Optimization

This was my most expensive lesson. My first version of this AI agent was using the LLM to basically do everything. I was feeding it the entire spreadsheet and expecting it to analyze, calculate, and generate insights all at once.

Reality check: this was costing me around $0.50 per day just to run, and the token usage was outrageous.

I had to step back and do some research. After talking with ChatGPT and Claude (yes, I used AI to help me build better AI), I learned that summarization and pre-processing are absolutely key. Now I do all the calculations myself—things like spending totals, category breakdowns, and trend analysis—before feeding the summarized data to the LLM.

The result? A 66% reduction in token count and much more consistent outputs.

The lesson here is crucial: anything you can do to help your agent out—whether it's calculations, data formatting, or preprocessing—will help you in the long run, not just with cost savings but with consistency. LLMs are powerful, but they're not perfect calculators.

API Limits & Rate Management

API rate limits are real, and you can hit them fast. This was probably my greatest technical challenge. Thankfully, the Anthropic team graciously allowed me to access a higher tier to finish this work, but for most people starting out, you need to be very careful.

I also learned that you can quickly accrue costs if you're not careful with error handling. A runaway loop or poorly designed retry logic can burn through your budget in hours.

Agent vs. Automation Balance

Here's something I didn't expect to struggle with: understanding the difference between an AI agent and an AI automation workflow.

My first attempts were essentially giving the AI so many detailed instructions that it turned into a glorified automation script. I had to learn that the key difference between an agent and a workflow is that you want the agent to have some autonomy, with instruction and guardrails in place.

Good prompting techniques became absolutely critical for getting consistent results. And yes, AI agent building is a LOT of trial and error. Don't let anyone tell you otherwise.

Data Security Comfort Level

Transaction data is deeply personal, and I had to make some trade-offs here. I'm already committed to using Google Sheets through Tiller, and that's about as far as I'm willing to go in terms of sharing financial data with cloud services.

For a prototype and learning experience, I decided this was an acceptable risk, but it's definitely something to consider carefully for your own situation.

Results So Far & What's Next

I'm currently in the testing phase to measure how much actual value this agent delivers. The early signs are promising—instead of staring at spreadsheets trying to figure out where to focus my financial attention, I get 3 clear, actionable goals each month.

Here's what I'm planning to add next:

  • Self-scheduling capability: Let the agent decide when to run based on new transaction patterns or market events
  • More sophisticated analysis: Seasonal spending patterns, goal tracking, and predictive insights
  • Other use cases: I want to experiment with AI agents for other areas of my life and work

The real test will be whether this actually changes my financial behavior over the next few months.

Should You Build Your Own?

If you're on the AI bandwagon and want to upskill, then I think building a simple agent is a must! I use AI to help me make my life easier, and an AI agent is just the next evolution of that.

Who This Is For:

  • Developers and tech enthusiasts who want hands-on AI agent experience
  • People already using tools like Tiller, YNAB, or detailed spreadsheet budgeting
  • Anyone comfortable with APIs and basic automation concepts
  • Folks who learn better by building than by reading tutorials

Start Here If You're Interested:

  • Pick a personal problem you actually have—don't build in a vacuum! The best learning comes from solving real problems
  • Start with n8n or similar low-code platforms to learn the concepts before diving into complex frameworks
  • Budget for LLM costs—start small and monitor closely. My $0.50/day mistake was educational but expensive
  • Expect lots of iteration—my first version was terrible and expensive, and that's completely normal

Not Ready to Build? Try This Instead:

  • Use existing AI tools like ChatGPT or Claude to analyze exported transaction data manually
  • Set up simple automation with Zapier + AI for basic financial insights
  • Start with manual prompting to understand what insights are actually valuable before automating

The Bottom Line:

Building this taught me more about practical AI applications than any tutorial or course could. Even if you never use the final product in production, the learning experience is absolutely worth it. You'll understand the limitations, costs, and possibilities of AI agents in a way that only comes from hands-on experience.

The future of personal productivity isn't just using AI tools—it's building AI agents that work for you, your specific needs, and your unique situation. This project was my first step into that future, and I'm excited to see where it leads.

What aspects of personal finance do you think would benefit most from AI assistance? Have you experimented with building your own AI agents? Let me know in the comments—I'd love to hear about your experiences!