AI finance learning center

Everything here is about one narrow question: what actually happens when software reads a financial document. There is a lot of hand-waving in this space, so these explainers try to be specific about the mechanics, including the parts that do not work reliably.

Start here

• How to read a business bank statement
  The credits that are not income, the debits that are not costs, and the monthly routine that keeps a business account honest.
• What is a settlement report?
  The document that explains the gap between what your customers paid and what your bank actually received.
• Gross, fee and net in a payout
  Three numbers on every payout line, and why recording only the deposit quietly breaks your accounts.
• What is a bank statement?
  Every part of a statement explained, the formats banks issue, and the balance arithmetic that proves nothing is missing.
• What is bank statement analysis?
  The definition, the four stages of the process, and who relies on it: lenders, accountants, and anyone tracking where money actually goes.
• What is cash flow analysis?
  Why a profitable business can still run out of money, and how to calculate net cash flow from statements without double-counting transfers.
• Preparing statements for a loan application
  What underwriters actually check, how many months to supply, and how to review your own statements before they do.
• How to analyze bank statements
  A repeatable seven-step process, from gathering statements to interpreting cash flow, with the reconciliation check that catches errors early.
• How to categorize transactions
  Choosing a category list that survives a year, writing rules instead of tagging rows, and the edge cases no rule can resolve.
• How to convert a PDF bank statement to Excel
  Three routes from PDF to spreadsheet, what each costs you in accuracy, and how to prove the conversion worked.
• What is OCR?
  A plain-English primer on optical character recognition: how it works, why it fails, and what modern OCR really does.
• How AI reads bank statements
  The pipeline behind AI-assisted statement parsing: layout inference, text extraction, hybrid language models, and validation.
• How AI detects expenses
  How automated expense detection works, from receipt capture to duplicate matching, with the failure modes to watch for.
• How AI categorizes transactions
  Where machine learning helps in transaction categorization, where deterministic rules still win, and how the two can coexist.
• How AI finds duplicate expenses
  The techniques behind duplicate detection: fingerprints, fuzzy matching, and cross-account reconciliation.
• How overlapping statements are handled
  Why the same week can arrive in two statements, and what a careful importer does about the transactions inside the overlap.
• How recurring charges are detected
  Merchant normalization, interval consistency and amount tolerance: the signals that separate a subscription from a coincidence.
• How receipts are matched to transactions
  Amount, date, merchant, account and time proximity, and why the confidence of a match should always be visible.
• How to scan a bank statement for unusual transactions
  Build a baseline, then run the amount, merchant, new payee and repeat passes that turn a wall of rows into a short list worth checking.
• How to analyze a credit card statement
  Why a card statement is not a bank statement, and how to keep purchases, refunds, payments, fees and interest from contaminating each other.
• How to extract transactions from a PDF bank statement
  Text layer or OCR, locating the table, mapping columns, joining wrapped rows, and the balance check that proves nothing was dropped.
• How to identify duplicate charges in a statement
  Fingerprints, date windows and merchant normalization, plus the overlapping statement problem that looks identical and needs a different fix.
• How to scan an invoice for errors
  Six checks before an invoice gets paid: arithmetic, quantities and rates, duplicates, tax treatment, payment details and the matching payment.
• How to extract data from a PDF without typing it in
  Four routes from PDF to structured data, what each costs in accuracy, and the verification step that tells you whether the result is usable.

Why "AI reads your statement" is three different things

When a tool says it uses AI to process a bank statement, it is usually describing some combination of three distinct steps, and they fail in completely different ways. Understanding which step is which makes it much easier to judge whether a given tool will work on your documents.

Step one is text recovery. A digitally generated PDF already contains text, positioned on a page. Getting it out is a parsing problem, not an AI problem, and a good parser is close to perfect at it. A scanned PDF contains an image of text, and recovering characters from pixels is optical character recognition. OCR quality depends almost entirely on the scan: resolution, contrast, skew, and whether the page was photographed at an angle. This is the single biggest predictor of whether a statement will process cleanly, and it is why a tool that works beautifully on one bank can fall apart on another.

Step two is structure recovery. Text on a page is not a table. Turning a page of positioned words back into rows of date, description, amount, and balance means inferring column boundaries that were never explicitly encoded. This is where most extraction errors actually come from. Descriptions that wrap onto a second line get read as separate transactions. A running balance column gets confused with the amount column, doubling your totals. Column positions shift between page one and page four because the header block is a different height. Multi-currency statements interleave two amount columns. None of these are exotic edge cases, they show up in ordinary statements from ordinary banks.

Step three is meaning. Deciding that "SQ *BLUE BOTTLE 4471" is a coffee purchase in a Meals category is genuinely a language problem, and it is the step where a language model earns its place. It is also the step where a model will confidently give you a wrong answer with no signal that it did. That is why Karchu applies deterministic rules first and only falls back to a model for descriptions nothing matched, and why every model-assigned category is marked as such so you can review it.

Why confidence scores matter more than accuracy claims

Any tool can quote a headline accuracy number. The number is close to meaningless on its own, because it depends entirely on the document mix it was measured against, and because a 98% accurate extraction of a 400-line statement still leaves eight wrong rows somewhere in the file with nothing pointing at them.

What is actually useful is knowing which specific rows the system was unsure about. A per-field confidence score turns an unbounded review job, checking all 400 rows, into a bounded one, checking the twelve the parser flagged. That is the difference between a tool you can trust with figures going to a lender and a tool you have to fully re-verify by hand, which defeats the purpose.

What none of this can do

Extraction cannot recover information the document never contained. If your statement shows a $4,200 transfer with the description "ONLINE TRANSFER", no model can tell you what it was for. If a page is missing from the PDF, nothing detects the gap except a balance check against the previous period. And no categorization system knows which of your meals were business meals. These stay human judgment calls, and any tool claiming otherwise is guessing.