Automated Balance Sheet Reading

An eight-column trial balance isn't a tidy table — it's a document where debit, credit, debtor, creditor, asset, liability, loss and gain are distinguished only by where they fall on the page. Generic OCR returns a list of numbers with no idea which column each one belongs to, and the result is a spreadsheet that has to be reviewed line by line. The problem isn't reading the characters — it's reconstructing the structure.

On top of that, every company names its accounts differently. «Accounts Receivable», «Customers» and «Trade Debtors» may be the same line item, or not, depending on the chart of accounts whoever prepared the balance sheet used.

Extraction has four steps: positional reading of the PDF to group values by column, reconstruction of each account with its balances, assisted mapping where the model proposes which financial-statement line it belongs to, and analyst confirmation, which is saved for future uploads.

1. 1Positional reading. The backend extracts the PDF text along with its coordinates and groups values by column based on where they sit on the page, not the order they appear in the file. That’s what produces exact balances, not approximations.
2. 2Account reconstruction. Each line is rebuilt as an account name plus its balances across the eight columns, respecting subtotals and grouping levels.
3. 3Assisted mapping. The model proposes which financial-statement line each account belongs to. This is the only part where AI is involved: it never reads the numbers, it only suggests the classification.
4. 4Confirmation and learning. The analyst accepts or corrects the suggestion. The correction is saved, so the same account name resolves itself on the next balance sheet.

The distinction in step 3 matters: balances come from the document's positional reading, which is deterministic and reproducible. The language model never calculates or transcribes figures — it only classifies account names, and its proposal always goes through human confirmation.

CreditIQ accepts eight-column and classified trial balances in PDF, income statements and statements of financial position under IFRS, Chile's CMF FECU, and Excel spreadsheets using the company's own chart of accounts.

• Eight-column and classified trial balances, in PDF.
• Income statements and statements of financial position under IFRS.
• FECU, Chile's CMF filing format.
• Excel spreadsheets with the company's own chart of accounts.

Since the analysis relies on financial statements under IFRS — the accounting standard used in more than 140 countries — reading works with balance sheets from any market that applies it, not just Chilean ones.

The system checks that total assets equal liabilities plus equity and that subtotals reconcile against the detail. When something doesn't tie out, it flags it on the data-entry screen instead of carrying the mismatch into the analysis. Amounts are always recorded as positive, and it's the engine that applies the sign based on the account's nature — the most common source of mismatches when loading is done by hand.

The analyst's time stops going into moving figures around. With financial statements already loaded and normalized, the risk analysis and the cash flow projection calculate on data that's consistent across periods and across companies — the condition for comparing a portfolio under the same criteria.