Grow2.ai replaced a $45–50/month SaaS finance tracker with Finickly — its own multi-company, multi-currency finance platform built with an AI coding agent. First commit on May 15, 2026; the business switched its real books to it on May 20 — five days later. The full eight-phase core took about 2.5 weeks: 326 commits, ~52,800 lines of TypeScript, 838 automated tests, 23,000+ historical transactions migrated. Three people use it daily; serverless hosting costs a few dollars a month.
In May 2026 we did to ourselves what we usually do for clients: replaced a subscription with our own tool. This is an honest case study of building a finance platform with AI — with dates, numbers and the parts that are still not done.
The Challenge: a Good SaaS That Doesn’t Fit the Business
FinMap served the standard bookkeeping fine — income, expenses, reports — at $45–50 per month. As SaaS goes, nothing was wrong with it.
The specifics didn’t fit, though. The business runs several legal entities in three currencies (UAH, USD, EUR). Client licenses carry obligations to vendors that need tracking against actual payments. Invoices live in the CRM and had to be re-typed into the tracker by hand. The cash-flow forecast lived in a spreadsheet next to all of it.
The classic answer — “custom software costs months and tens of thousands of dollars” — stopped being true this spring. So we tested the new math on ourselves: how fast can one founder with an AI coding agent replace a SaaS he pays for?
The Solution: Finickly, a Finance Platform Built Around Our Own Loop
Grow2.ai built Finickly — a self-hosted finance platform that runs the company’s entire money loop:
- Full ledger for 4 legal entities. Multi-currency (UAH/USD/EUR) with exchange rates pulled daily from official sources.
- Salaries. Monthly accrual matrix; marking a salary “paid” books the expense transaction atomically.
- Invoices straight from the CRM. Outbound webhooks — an issued invoice appears in the books without re-typing.
- License-to-vendor commitments. Who paid, what we owe the vendor and the status of every license — computed automatically.
- Weekly cash-flow forecast. A 4–12-week horizon built from planned and recurring payments, plus a Telegram bot that alerts on projected cash gaps.
- Full history on board. A one-off migrator moved 23,000+ transactions going back to 2019 from the previous tracker.
The stack is fully serverless (Cloudflare Workers + D1, React SPA), so the platform costs a few dollars a month to run and needs no ops team.
The Numbers
Metric | Value |
|---|---|
SaaS subscription before | $45–50 / month |
First commit → production books | 5 days (May 15 → May 20, 2026) |
Eight-phase core | ~2.5 weeks |
Commits / TypeScript | 326 / ~52,800 lines |
Automated tests | 838 |
History migrated | 23,000+ transactions (2019–2026) |
Legal entities / daily users | 4 / 3 |
Hosting now | a few $ / month (serverless) |
The timeline is the point of this case. Planning all eight phases took the first five days — and on day five the business already kept its real books in Finickly. The core was done in about two and a half weeks; everything since June is small operational polish driven by daily use: partial invoice payments, salary prefill, visibility of unposted payments.
How You Trust AI Code With Your Money
“An AI wrote 52,800 lines that manage our finances” sounds reckless — until you see the process:
- A written plan per phase. All eight phase plans were drafted in the first five days; each breaks the work into small reviewable tasks.
- Test-driven development. 838 automated tests run against a real serverless database and are written alongside the code, not after it.
- A smoke-checklist after every phase. A human walks the live product: creates transactions, checks balances, tries to break things.
- An end-of-phase code review. Every phase closed with a review pass, and findings became fixes before the next phase started.
This is the opposite of vibe coding: the AI does the typing, the human sets the frame and accepts the result.
What This Is Not
Finickly is single-tenant internal tooling, not a product. There is no multi-tenant isolation, no billing, no onboarding — deliberately. And self-hosted means we are our own vendor: when something breaks, it’s on us.
The economics, honestly: the tooling cost was one Claude Max (x20) subscription — $200 per month, which in the same weeks also served our other projects — plus the founder’s hours for planning, reviews and smoke-tests. The subscription was already in our stack, so the marginal cost of the platform was close to zero. But “AI built it for free” would be a lie.
The takeaway is bigger than one tracker: the threshold for “build your own” has dropped by an order of magnitude. A tool that used to mean months of development and a five-figure budget now costs weeks of part-time attention on top of a subscription you may already have. That doesn’t make every SaaS replaceable — but it makes the question worth asking about each one you pay for.
Which of your subscriptions could a custom AI-built tool replace? Take the 2-minute AI audit or talk to us — we’ll show where custom beats SaaS in your stack.
Case documented by Andrew Maryasov, founder of Grow2.ai — AI agents for business. All numbers come from the project’s git history and the production database, July 2026.