The people in this story are made up. The problems are the ones we hear most.
Flowchart
Automatic document classification
- StartClient upload to SharePoint (mixed PDF pile)
- StepIngestion
- DecisionClassification (W-2 / 1099 / bank statement / receipt / brokerage statement / other)otherA personStaff review
- StepPer-type extraction
- DoneSaved to database by client and tax year
February at a five-person accounting firm
Before
- Client upload "scan001.pdf"
- Staff open and rename
- File to folder
- Check checklist by hand
After
- Client upload
- Classified by type
- Stored by client and year
- Checklist shows what's missing
Priya's firm prepares a few hundred individual returns a year. In February the firm's SharePoint site fills with client uploads named "scan001.pdf", "IMG_4471.jpg" and "taxes stuff.pdf". One client sends a single 30-page PDF with a W-2, two 1099s, a mortgage statement and a photo of a daycare receipt, in no particular order.
Before anyone prepares a return, a staff member opens every file, works out what each page is, renames it and moves it into the client's folder. Then they check the year's checklist to see what is still missing. It is the least skilled work in the office, and it eats the first hours of every day in the busiest season.
The problem isn't reading the documents. It's knowing what each one is.
What document classification does
- Unknown document
- Classifier
- W-2
- wage fields
- 1099
- payer and amount
- Brokerage statement
- investments
- Receipt
- credits
Classification answers one question for each document or page: what type is this? Once that is known, everything else follows. A W-2 gets its wage fields pulled out. A brokerage statement goes to the investments checklist. A daycare receipt goes to credits.
Generic tools come with a fixed list of document types. That works for common forms and fails on the ones that make your work different: a local property tax bill, a specific bank's statement layout, your own engagement letter.
Training it on your own documents
- List your types
- Gather real examples
- Train
- Test on unseen files
- Add examples where it's wrong
back to Train
In Document Processing Factory, the classification widget lets you choose the document types you handle and train it on examples of your own files. For Priya, the setup looks like this:
- 01List the types. W-2, 1099-INT, 1099-DIV, 1099-NEC, 1098 mortgage statement, brokerage year-end statement, property tax bill, childcare receipt, and "other".
- 02Gather examples. Pull real examples of each type from last year's client files, including the messy ones: phone photos, crooked scans, multi-page statements.
- 03Train and test. Train the classifier, then run it on a batch it hasn't seen and look at what it got wrong.
- 04Add examples where it struggles. If it confuses 1099-INT with 1099-DIV, add more of both.
- 05Keep an "other" type. Anything it can't place confidently goes to a person, not to a guess.
What happens after sorting
- Typed document
- Extraction for that type
- Saved by client and tax year
- Missing-documents query
Once each document has a type, the rest of the workflow follows on the canvas:
- Extraction pulls the fields for that type: employer and wages from a W-2, payer and interest from a 1099-INT.
- Output saves the results to a database by client and tax year, or sends them to other software through an API call.
- The checklist answers itself. With every document typed and stored by client, "what is this client still missing?" becomes a query instead of a folder search.
Problems it solves, specifically
- Files nobody can findNamed by type
- Sorting instead of preparingSorted on arrival
- Missing documents found in AprilVisible in February
- Unusual formsTrain a new type
- Files named "scan001.pdf" that nobody can find later.
- Staff time spent sorting instead of preparing returns.
- Missing documents discovered in April instead of February.
- New or unusual document types that off-the-shelf tools don't recognize.
The same pattern works outside tax season: insurance agencies sorting claim packets, law offices sorting discovery productions, property managers sorting application bundles.
What to watch while it's new
- Daily sample check
- Note confused types
- Add examples
- Person reviews "other"
Check a sample of classified documents every day for the first weeks. Track which types it confuses and add training examples for those. Keep a person on everything sent to "other". Classification should save sorting time, never hide a document.