How AI actually categorizes a transaction
A look under the hood at how Tryizzy classifies a bank line, scores its confidence, and decides when to ask you a question.
When a transaction lands from your bank, it arrives as a messy string, an amount, and a date. Turning that into a correct accounting entry is the core of what we do. Here is roughly how it works, without the jargon.
Step one: clean the signal
Raw bank descriptions are noisy. The same coffee shop might appear three different ways across two banks. We normalize the merchant, strip processor prefixes, and match it against vendors we have seen before, both yours and the broader pattern across similar businesses.
Step two: propose a category
With a clean merchant and amount, the model proposes a chart of accounts entry. A recurring SaaS charge looks different from a one time equipment purchase, and a Stripe payout is revenue net of fees, not a single expense. The model has seen enough of these to be right most of the time.
Confidence is a first class output
Every proposal comes with a confidence score. This is the part that makes the system trustworthy. A 99 percent confident coffee purchase posts automatically. A 61 percent confident transfer that might be owner draw or might be a loan repayment does not. It becomes a question.
Step three: ask, do not guess
When confidence is below threshold, the transaction goes to your clarifications inbox with a short, specific question. You answer in a sentence. The model learns your answer, applies it, and uses it to raise confidence on similar transactions in the future.
- High confidence posts automatically and is reconciled.
- Medium confidence posts but is flagged for a quick review.
- Low confidence becomes a clarification before it touches your books.
The goal is not a model that is always right. It is a model that knows when it might be wrong and asks instead of guessing.
Why this beats rules
Old school bookkeeping software relies on rules you build by hand: if the description contains this, categorize it as that. Rules are brittle and they never cover the long tail. A model that proposes, scores, and asks handles the messy middle that rules cannot, and it gets better the more you use it.
