Building a Financial Planning MVP: 30 Days of AI-Assisted Development, $127 Spent, and the Lessons That Actually Mattered

2026-01-09 10:24:38

TL;DR: I spent a month building a financial advisor tool for founders using AI-assisted coding. Burned $127 in credits, made nearly every mistake possible, and ended up with a $50/month validation from one founder. The real lesson: AI excels at speed but struggles with precision. Less turned out to be more than I ever expected.

The Problem Worth Solving

I’ve worked with founders for years. I’ve watched the same scene play out repeatedly: a VC asks “what if churn drops 2%?” and the founder’s face goes blank. His answer lives somewhere in a 47-tab Excel nightmare. The meeting momentum dies. The founder loses hours rebuilding formulas. Cells break. Circular references crash everything.

The core frustration I kept hearing: “I built a financial model once. When they asked for a single scenario change, I had to rebuild the entire thing.”

Most early-stage startups still use spreadsheets. Most founders despise it. So I decided to test if AI could help them escape this trap.

Building Without the Blueprint: The First Two Weeks

Week 1: How Optimism Gets Expensive

I dove in convinced this would take 2-3 weeks. I’d seen the AI influencers make it look trivial on social media, right?

My initial roadmap looked like:

AI-powered financial cockpit with real-time syncing
QuickBooks and Stripe integration built-in
Scenario planning with investor-ready exports
Everything in weeks, not months

Reality had other plans.

The Cost of Vague Instructions

My first mistake was treating the AI agent like it could multitask. I fired off three requests while it was still working on the previous one:

“Make the dashboard cleaner”
“Add dark mode”
“Fix the calculation bug”

The AI absorbed all of them simultaneously, got confused, and created something that did none of them well. That cost me 6 rollbacks, 3 hours of debugging, and $23 in credits. I could have saved that entire expense by simply waiting.

The UI That Broke Everything

I asked the AI to “add night mode.” It proceeded to make 47 changes. The result: white text on white backgrounds, invisible buttons, a complete interface collapse. Spending three days matching fonts and backgrounds taught me that UI complexity scales faster than expected.

The Magic Discovery

Then I found the phrase that changed everything: “Don’t make any changes without confirming your understanding with me.”

This single instruction could have saved me $50+. It forced the AI to explain its approach before executing, catching misunderstandings before they burned credits.

Week 2: When Travel Slows Down Progress

Building from airport lounges in Japan taught me humbling lessons:

Hotel WiFi + Replit development = constant frustration
Debugging TypeScript errors on mobile is almost impossible
The rollback button becomes your closest friend

I’d chosen TypeScript thinking it was the “modern choice.” Bad call. It’s a language I don’t really understand. When financial formulas got complex, I spent more time fighting syntax than building features. Example: a simple runway calculation took 2 hours because TypeScript kept complaining about type mismatches.

Note to future builders: Pick a language you actually understand. The learning tax isn’t worth it when you’re prototyping.

By day 15, Replit credits were hemorrhaging. Week 1 cost $34. Week 2 cost $93. Each iteration—change, test, rollback, try again—drained $2-5. I had to set a new rule: $40 per week maximum, or stop and rethink why I’m burning through so much.

The Moment Everything Changed: User Feedback Week

Day 17: Hunting for Testers

I posted in founder Slack channels: “Building a financial planning tool that doesn’t suck. Need critical feedback.”

Crickets.

But I persisted. Eventually, one friend and two founders agreed to test. Their feedback was brutal and eye-opening.

Day 18-20: The Humbling Truth

Issue #1: My calculations were wrong by 20%

A founder’s customer acquisition cost showed $47 when it should have been $58.75. That margin of error could have tanked their Series A pitch. The culprit: I’d asked MistralAI to “calculate customer acquisition cost” with vague instructions. The AI made assumptions about methodology. Sometimes it interpreted “churn” as monthly; other times as annual. Consistency evaporated.

Issue #2: Larger models crashed the export feature

Anything with >50 rows caused memory overflow.

Issue #3: The core feature was buried

Founders wanted runway calculation most. I’d buried it three screens deep. They had to navigate through five pages just to find what they needed.

The 6-Hour Debugging Session

The LTV/CAC calculations stayed consistently wrong. Six hours of tracing revealed the problem: MistralAI was interpreting “monthly churn” as “annual churn” in some scenarios and vice versa in others. When I asked for “customer lifetime value,” it made hidden assumptions.

Bad prompt: Calculate LTV

Good prompt:

This page may contain third-party content, which is provided for information purposes only (not representations/warranties) and should not be considered as an endorsement of its views by Gate, nor as financial or professional advice. See Disclaimer for details.