The AI gives you a strong first-pass classification of your transactions — typically catching the vast majority of common business expenses correctly. It is not a substitute for review. You stay in the loop, and uncertain items are surfaced for you to confirm.
How the Classifier Works
- We extract individual transactions from your uploaded PDF — description, amount, date.
- Each transaction is sent to a large language model (Anthropic Claude) along with your industry context and chart of accounts.
- The model returns a category (e.g., "Office Supplies", "Subcontractor Labor", "Owner Draw") and a confidence signal.
- Rules and guards run on top of the model output to catch known edge cases — for example, Zelle transfers, refunds, and healthcare payments.
Where It Flags Uncertainty
Transactions the model is unsure about are written to the "Flagged Items" section of your report. Common reasons include ambiguous merchant names, large round-number transfers, and transactions that could be either personal or business. These are the rows you should review first.
Human in the Loop
- You can reclassify any transaction yourself from the General Ledger view (Professional and Advisory plans).
- Your own corrections improve future runs for the same vendors.
- Your tax professional should review the final report before any tax filing — the AI output is informational, not a substitute for professional judgment.
We do not publish a single accuracy percentage because it varies by industry, document quality, and how clean your source data is. A well-formatted QuickBooks export will produce far cleaner results than a scanned bank statement.