What Finman's AI does today (checked 23 October 2026)The AI CFO chat answers questions from your own accounts, budgets, goals, debts and transactions. The same model writes a monthly summary, spending insights and a savings plan, and reads receipt photos. It is an OpenAI GPT model reached through Finman's own ChatGPT account. Transaction categories come from merchant codes, keywords and your own rules.
AI categorizes a transaction by predicting its category from features — the merchant string, amount, frequency, the account, and crucially your own past corrections — rather than by matching the merchant name against a fixed lookup table. The defining property of a real system is that it improves when you correct it; a fixed rule table does not.
The signals a model actually uses
- Merchant string — normalised (e.g. "SQ *BLUE BOTTLE" Blue Bottle), the strongest single signal.
- Amount and sign — a $4 debit and a $1,400 debit at the same merchant are often different categories.
- Recurrence — same merchant, similar amount, regular interval implies a subscription or bill.
- Your history — how *you* have categorized similar transactions before, which personalises the model to your life.
- Account context — a charge on a business card vs a personal card shifts the likely category.
Why "learns from corrections" is the test
A rule-based categorizer maps strings to categories and never changes. A learning categorizer treats your correction as a labelled training signal, so the next similar transaction is classified your way. If you re-categorize the same merchant three times and it keeps getting it wrong, it is not AI — it is a lookup table with an AI label.
Where it still fails
- Generic processors (PayPal, Square) hide the real merchant, so the model is guessing from amount and history.
- Ambiguous merchants that legitimately span categories (a superstore that sells groceries and electronics).
- Cold start: a brand-new user with no correction history gets population-average guesses until the model personalises.
A well-designed system handles this by exposing a confidence level and flagging low-confidence items for review rather than silently guessing — silent confident errors are worse than asking.
Frequently Asked Questions
How does AI categorize bank transactions?
It predicts a category from features — normalised merchant name, amount and sign, recurrence pattern, account context, and your own past corrections — and updates from your corrections so similar future transactions are classified your way. This is fundamentally different from a fixed merchant-to-category lookup table.
Why does my finance app keep miscategorizing the same merchant?
Because it is likely rule-based, not learning. A real AI categorizer treats your correction as training data and stops repeating the mistake; a lookup-table app will repeat it indefinitely.
How accurate is AI transaction categorization?
High for clear merchants and after the model has learned your corrections; weaker for generic payment processors (PayPal/Square) that mask the merchant and for genuinely ambiguous merchants. Good systems show a confidence level and flag uncertain items instead of guessing silently.
Does Finman learn from my corrections?
Yes — Finman’s categorization uses your corrections as a signal so similar future transactions are classified the way you intend, and it falls back gracefully when confidence is low.
See categorization that actually learns
Import a month of transactions and correct one merchant twice — watch it stick.
Try Finman FreeRelated reading: AI Personal Finance Guide · Best AI Budgeting App · Is AI Safe for Finance?