AI can do most of this extraction work, but scanned docs and tax stakes demand a careful human check.

The target fields (revenue, expenses, net income, assets) are consistent, but document layouts vary wildly across 28 clients using different formats, templates, and accounting conventions. Each document requires some structural adaptation, not pure repetition.

The four target line items are well-defined in principle, but accounting labels differ across clients — 'gross revenue' vs. 'total sales' vs. 'net receipts' — and the agent must make judgment calls about which rows map to which fields. Success criteria are clear at the output level but fuzzy at the extraction level.

Excel files are machine-readable and easy to parse; PDFs vary from clean to complex; scanned bank statements require OCR and are the weakest link. An agent needs PDF parsing, OCR, and spreadsheet-writing tools — all available, but quality degrades significantly on scanned inputs.

These figures feed directly into tax planning decisions. A misread net income or missed expense category could cause incorrect tax estimates, client liability, or regulatory exposure. Errors are not immediately obvious and may not surface until filing — making this high-stakes.

Most extraction is mechanical, but edge cases — ambiguous line items, multi-entity statements, unusual accounting treatments — require a consultant's judgment to classify correctly. The agent can do the heavy lifting; a human must validate the output.