AI can do the heavy lifting here, but tax figures demand a human final check.

The overall pipeline is structurally consistent across 35 clients, but document layouts vary significantly — bank statements from different institutions, hand-formatted P&Ls, and scanned vs. native PDFs all require adaptive parsing logic. This is repeatable in structure but not in execution detail.

The output schema is well-defined (CSV with specific columns including confidence scores), which is favorable. However, categorizing line items as 'deductible expenses' vs. other categories requires tax-domain judgment that isn't always unambiguous from raw document text.

Files are locally stored in a folder hierarchy, so an agent needs filesystem access and OCR tooling (e.g., AWS Textract, Azure Form Recognizer, or open-source alternatives). Excel parsing is straightforward, but scanned PDFs may require high-quality OCR with layout understanding — available but not trivially configured.

Errors in extracted figures flow directly into tax filings, creating IRS liability, client trust damage, and potential professional sanctions for the preparer. A misread income figure or missed deduction is not a minor inconvenience — it's a compliance failure. This makes silent errors especially dangerous.

Most extraction is mechanical, but edge cases — ambiguous expense categories, multi-entity P&Ls, partial-year investment statements — require a tax professional's interpretation. The confidence score column is a smart design choice that surfaces exactly these cases for human review.