Audit document extraction for working-paper evidence

An audit runs on evidence: trial balances, bank confirmations, GL extracts, supporting schedules. Most of that evidence arrives as PDFs and scans that have to be transcribed into working papers by hand, a slow, error-prone step that delays the whole engagement. Dynamite Docs turns the source documents into structured rows so the tie-out starts from clean data instead of a rekey.

Try this workflow free or view pricing.

Evidence transcription is the part of an audit that should not need a human

Every client delivers trial balances, bank confirmations, and schedules in a different format, and every one has to be keyed into the working papers.
Rekeying a trial balance or confirmation is where transposed digits and dropped schedules hide, and those errors surface as tie-out differences and review points.
Confirmations arrive as scans and photos with no text layer, so standard PDF tools read nothing.
Month-end and year-end audits pile up source documents, and the transcription backlog delays fieldwork and the close.

How Dynamite Docs handles audit source documents

01 — Drop the source evidence in

Trial balances, bank confirmations, GL extracts, and supporting schedules. Clean digital PDFs parse deterministically from their embedded text; scans and photos are read by the vision model you choose.

02 — Infer the evidence schema

Account codes, descriptions, periods, opening/closing balances, movements, and totals are pulled from each layout, no per-client templates.

03 — Check the low-confidence rows

Every value carries a confidence score. You check the low-confidence rows, usually a balance or a confirmation amount, and each correction becomes a pattern the next schedule reuses.

04 — Export into your working papers

CSV, JSON, Excel, Word, or Sheets, or push rows with the API and webhooks into your audit workflow. Evidence stays structured and attributable.

Document types this workflow handles

Trial balances — the primary working-paper input
Bank confirmations — client statements and bank responses
General ledger extracts — GL runs and movement reports
Supporting schedules — the evidence behind each assertion

What gets extracted from audit source documents

Inferred per layout. A typical trial balance or confirmation run yields:

Account code — 1100-100
Account description — Cash — operating
Period — FY2026 / Q3
Opening balance — $48,200.00
Movements / activity — +$12,400.00
Closing balance — $60,600.00
Bank / institution — First Meridian Bank
Confirmation amount — $60,600.00
Entity / client ref — CLIENT-0142
Document / evidence ID — TB-2026-04

A realistic example

A client trial balance extracts to rows like these, ready for the tie-out and working papers:

Account	Description	Opening	Movements	Closing
1100-100	Cash — operating	$48,200.00	+$12,400.00	$60,600.00
1200-200	Accounts receivable	$112,000.00	−$8,300.00	$103,700.00
2100-300	Accounts payable	−$74,500.00	−$3,900.00	−$78,400.00
5000-400	Revenue — services	$0.00	+$340,000.00	$340,000.00

What good audit evidence looks like in practice

The audit workflow is built on evidence, and evidence is only useful when it is readable and attributable. A trial balance is a list of accounts with balances; a bank confirmation is a small form with one number that matters; a GL extract is a long table of transactions. All three arrive as documents, and all three have to end up as rows you can tie out and cite. The extraction output, clean rows with confidence scores, is what turns a folder of client PDFs into a working-paper set.

Accuracy is the point, not the speed. A transposed digit in a trial balance or confirmation hides in plain sight and shows up later as a tie-out difference or a review note. That is exactly why confidence-scored review matters more here than in a high-speed back-office batch. You verify the handful of low-confidence values, usually a closing balance or a confirmation amount, instead of trusting every field blindly.

Attribution is the other requirement. Each extracted row should stay linked to its source document so the working paper can point back to the evidence. Dynamite Docs keeps documents and their extracted rows together in folders, so the audit trail stays intact and exportable with the numbers.

Choose the document your team transcribes most. For most audits that is the trial balance. Run a real client’s file, review the low-confidence balances, and let the remembered layout make the same client’s next period extract cleanly. The same pattern extends to confirmations and schedules as you roll the workflow across the evidence set.

Why bring-your-own-AI matters for audit evidence

Audit evidence is confidential and under professional obligations. It should never be routed to a model your firm did not approve. Bring your own keys, lock processing to no-training providers in Sensitive Documents mode, or run fully local with Ollama, so client data stays under your control.

Audit volume varies wildly by season. Owning the model choice means you scale cost to the engagement, deterministic parsing for clean digital trial balances for free, your own vision key only for the confirmations and scans that need it.

Your per-client layout corrections and evidence attributes are remembered as patterns. The knowledge accumulates in your workspace and exports with you. The working-paper structure stays yours, not a vendor’s.

Questions, answered

Can it read a scanned bank confirmation?

Yes. Scans and photos with no text layer are read by the vision model you choose, using the same schema inference as digital PDFs. The confirmation amount surfaces as a confidence-scored field for your review.

Does it tie out a trial balance for me?

No. Tie-out and analytical review are your audit procedures. Dynamite Docs extracts the balances accurately and flags low-confidence values, so your tie-out runs on clean, attributable rows instead of a rekey.

How do I keep evidence linked to source documents?

Documents and their extracted rows live together in the workspace. Exports keep row references intact, and the API and webhooks let you push both the data and its evidence ID into your audit system.

Is client data safe for an audit firm?

Documents are encrypted in transit and at rest, never used to train foundation models, and provider keys are encrypted. Sensitive Documents mode blocks training-permitted providers by default, and local or self-hosted options keep data on your infrastructure.