If your team prepares write-ups from client bank statements, you already know the routine: download the PDF, scroll, re-type transactions into a spreadsheet, categorize each line, tie out the balances, format it for review. For a client with three accounts and a busy statement, that's an afternoon. Multiply by your client list and it's someone's entire month-end.
The entire middle of that process — everything between "statement arrives" and "reviewer opens the workpaper" — can be automated. Here's the pipeline, stage by stage, including the parts most write-ups of this topic skip: where it breaks and how you catch it.
Stage 1: Extraction — getting data out of the PDF
Bank statements come in two flavors. Digitally generated PDFs contain real text — a parser can read the transaction table directly, and extraction is close to lossless. Scanned statements are images and need OCR, which introduces error: a smudged 8 becomes a 3, a line wraps and merges two transactions.
The practical rule: build for digital-first, and route scans through a review step. Extraction should produce a structured table — date, description, amount, direction — plus the statement's own opening balance, closing balance, and transaction count. Keep those three numbers; they're your safety net in Stage 4.
Stage 2: Normalization — making banks agree with each other
Every bank formats differently. Dates appear as 03/04, Mar 4, or 2026-03-04. Debits are negatives at one bank, a separate column at another. Descriptions arrive in uppercase with reference numbers bolted on.
Normalization converts every statement into one canonical schema before anything downstream touches it. This is unglamorous work, but it's what makes the rest of the pipeline bank-agnostic — add a new bank by writing one adapter, not by rebuilding the flow.
Stage 3: Categorization — the judgment layer
This is where teams assume they need heavy AI. Usually they need three layers, applied in order:
- Exact rules first. "STRIPE PAYOUT" is always revenue clearing; the hydro bill is always utilities. Deterministic rules should catch the recurring majority of lines — they're free, instant, and never drift.
- Pattern matching second. Vendor names with changing reference numbers, payroll entries with varying amounts — fuzzy matching against known vendors handles these.
- A language model last, only for the genuinely ambiguous remainder, and always with a confidence flag. A low-confidence categorization should land in a review queue, not silently in the workpaper.
The ordering matters for cost and for trust: the deterministic layers are auditable line-by-line, so the model only ever touches the edge cases a human would also have had to think about.
Stage 4: Verification — tie out before anyone sees it
The non-negotiable stage. Before a workpaper leaves the pipeline, it must pass the same checks a preparer would do by hand:
- Opening balance + sum of transactions = closing balance, per the statement's own figures.
- Extracted transaction count matches the statement's count, where the bank prints one.
- Every line carries a category; anything uncategorized or low-confidence is flagged, never defaulted.
If the tie-out fails, the statement goes to a human with the discrepancy highlighted — the automation's job is to make the exception obvious, not to hide it. This single rule is the difference between an automation reviewers trust and one they quietly re-check line by line, which saves nobody anything.
Stage 5: Output — meet the reviewer where they work
Generate the workpaper in the format your reviewers already use — typically Excel with one sheet per account, a summary tab, and consistent formatting. Include the audit trail: which rule or model categorized each line, and at what confidence. A reviewer who can see why a line was coded reviews in minutes instead of re-deriving everything.
What this actually saves
The honest math: extraction, normalization, and output formatting are near-total time eliminations. Categorization drops from every-line work to exceptions-only. Verification stays human but starts from a pre-checked document. In practice the preparer's role shifts from "type everything, check everything" to "clear the flagged items" — which is where their judgment was needed all along.
If reconciliation write-ups are a meaningful slice of your team's month, this is usually one of the first two or three processes an audit will flag as worth automating. Book a free process audit and we'll tell you whether it's true for yours.