|
BUILD WITH AI |
BUILD WITH AI · PART 1 OF 5
How to build a client document extraction workflow with an AI assistant
The workflow replaces manual re-keying of client statements - and the prompt below is the part most firms skip writing down.
Most accounting teams lose time not on analysis but on intake: a client sends a PDF bank statement or a scanned lease schedule, and someone on staff manually types figures into a workpaper or a spreadsheet. A general-purpose AI assistant can move that extraction step from a half-hour task to a two-minute one, provided the firm builds a repeatable prompt and keeps a human in the verification seat.
BUILD WITH AI
Client document extraction, step by step
This workflow replaces manual re-keying of client-supplied statements by having an AI assistant produce a structured draft extraction that a preparer verifies in a single pass. Run it on one document type at a time until the prompt is stable for that format.
1 |
Confirm the source document has a selectable text layer; if it is a scanned image, run it through an OCR tool first and save the output as a searchable PDF. |
2 |
Open your AI assistant and upload the document or paste the relevant pages as text. |
3 |
Paste this prompt exactly, substituting the bracketed items for your document: 'You are an accounting data extraction assistant. From the attached [bank statement / lease schedule / trial balance], extract all line items into a table with these columns: Line item label | Period | Amount | Currency | Source page number. |
4 |
Review the uncertainty flags first; any row marked uncertain must be verified directly against the source before the figure enters a workpaper. |
5 |
For unflagged rows, spot-check at least three line items by tracing each amount back to the page number the assistant cited. |
6 |
Confirm that the extracted totals agree with any summary totals printed in the source document; a discrepancy here means at least one line item was missed or misread. |
7 |
Paste the verified table into your workpaper template or reconciliation schedule and note in the workpaper that the source extraction was AI-assisted and human-reviewed, with the reviewer's initials and date. |
The failure mode is subtle. AI assistants read text layers in PDFs well but struggle with scanned images, handwritten amendments, and multi-column statements where totals belong to the wrong row after extraction. The assistant will return a confident-looking table even when a column has shifted. That is exactly the condition a staff reviewer must be trained to catch - not just spot-checking the math, but confirming that each extracted figure maps to the correct label in the source document.
The workflow below is designed to produce a draft extraction that a preparer can verify in one pass rather than rebuild from scratch. The prompt asks the assistant to flag any cell where it was uncertain, which surfaces ambiguity rather than burying it. Firms that add a second column for source-page citations find that verification time drops further, because the reviewer knows exactly where to look in the original file.
Once the extraction step is reliable, the same structured output can feed directly into workpaper templates or reconciliation schedules without a second round of keying. The practical limit is document quality: anything below a clean digital PDF should be run through an OCR tool first, or handled manually. A workflow that silently produces wrong numbers on degraded documents is worse than no workflow at all.
WORKED EXAMPLE
In practice
A three-month bank statement for a small business client, covering the period just ended, arrives as a digital PDF. The preparer needs a line-item table of all credits and debits to populate a cash reconciliation workpaper for the current month.
What came back. The assistant returned a 47-row table covering all three months, with source page numbers in every row. Two rows were flagged as uncertain: one where a description was partially cut off at the page margin, and one where a deposit amount appeared twice in adjacent rows and the assistant could not confirm whether the second entry was a duplicate or a separate transaction. The running balance in row 31 did not match the printed statement balance for that date, indicating one line item had been skipped.
How it was checked. The preparer opened the source PDF, located the skipped line item on page 4 using the source-page column, added it to the table, and confirmed the running balance then agreed to the printed statement figure before importing the table into the reconciliation workpaper.
A constructed example. The prompt is usable as written; the figures show the shape of a result, not a measured one.
WHEN TO USE IT
| WHEN NOT TO
|
WHAT TO TAKE FROM THIS
| Check whether your assistant reads the PDF text layer or only the image before relying on it for scanned documents. | |
| Ask the assistant to flag uncertain extractions explicitly - do not assume a clean table means a correct one. | |
| Pilot on one document type at a time; a prompt tuned for bank statements will not transfer cleanly to lease schedules. |
SPONSORED
QUESTIONS THIS ANSWERS
Can I use any AI assistant for client document extraction, or does it have to be a specific product?
Any general-purpose assistant that accepts file uploads or pasted text will work for the extraction step. The Financial Cents guide to Claude for accounting notes that Claude supports document summarization and data extraction; other assistants with similar file-handling capabilities follow the same pattern. The prompt and review steps in this workflow are product-agnostic.
What document types does this workflow handle reliably?
Clean digital PDFs with a selectable text layer extract reliably. Scanned images, handwritten notes, and documents with merged or multi-column tables require OCR pre-processing or manual handling. Run a pilot on your most common document type before rolling the workflow out broadly.
How do I know if the assistant got the numbers wrong?
Ask the assistant to include a source-page citation for each extracted figure and to flag any cell where it was uncertain. Then spot-check flagged cells and at least one unflagged row against the original document. A total that does not foot to the statement's own summary line is the fastest signal that something shifted during extraction.
SOURCES
Where this comes from
The career portal does the rest.
GO DEEPER
Go deeper
IN THIS SERIES
Previously: How to use an AI assistant to prep client advisory meetings faster
Next: Build with AI: bank transaction categorization (coming)
An original workflow written for practitioners. Replicate it in a sandbox first; nothing here replaces your review.
How accountants are using AI, automation and smarter workflows to close faster, audit cleaner, and free up time for real work.
Accounting Stack · Audit Friendly Data · Accounting & Finance Jobs
Audit Friendly · modernaccounting.ai