The problem
Every client sends a different stack: prior-year returns, W-2s, 1099s, K-1s, brokerage statements, and a phone photo of a 1098. In busy season, staff retyping that stack is the bottleneck in front of every preparer.
What goes wrong
- The firm trusts the extraction blind. A 1099 total is read wrong, nobody checks it against the source, and the error reaches the return.
- Every document goes to review. The queue takes as long as keying did, so staff approve on sight and the review step stops meaning anything.
- Client files go into a public AI tool with no consent and no log. When a client asks who saw their return, the firm has no answer.
The approach
- Start with 50 to 200 of your own client documents under a mutual NDA, and agree the document types, the fields, the checks, and the thresholds in writing before anything is built.
- Check every field before a person sees it. The 1099-B totals tie to the statement summary. The partner on the K-1 is your client. The tax year is the one you are preparing.
- Put the review queue where your staff already work, with the source page beside the fields. Start with every record in review, and let records that pass every check skip it only after the shadow weeks.
- Run it in accounts you own, log every step per document, and send nothing to a public AI tool without your written consent.
When it works
When the same document types arrive every season, you know the fields you need from each, and a named manager owns the review queue and the thresholds.
Why intake is the busy-season bottleneck
A CPA firm bills for judgment and spends busy season on sorting. A client uploads last year's return, a dozen 1099s, three brokerage statements, and a phone photo of a 1098. A staff accountant opens each one, finds the number, and keys it into the tax software and the workpaper. The preparer cannot start until that is done, and one pile at the front of the line holds up the whole return.
Document intake fixes what was actually broken: a person was the parser. The pipeline reads what the client sent, pulls the fields, checks them, and queues a workpaper for a yes. Your staff review instead of retype.
What the pipeline should do
Read
Every document type you agreed on, from the upload link, the firm inbox, or a shared folder. PDF, scan, or photo. The pipeline tells a W-2 from a 1099-INT from a K-1, splits a packet into its parts, and pulls the fields each form carries. Every field shows a confidence level your staff can see.
Check
Extraction is a guess with a score. Checks catch the wrong guesses. Beyond the ties and the tax year, W-2 withholding has to be less than wages, and a payer the client did not have last year is flagged. A field that fails a check is marked, with the reason, before anyone opens it.
Queue
Approve, edit, or reject. Auto-posted fields are marked so the preparer sees which numbers no person has checked. Every correction is logged and feeds the accuracy check.
Chase
The pipeline tracks each client's checklist against what arrived and drafts the request for anything missing. A staff member approves it before it goes to the client.
How we build it
First comes the sample. We sign a mutual NDA before any file moves. You send 50 to 200 real client documents, and we agree the document types, the fields, the checks, and the thresholds in writing.
Next is extraction in shadow mode. Your staff see records in the queue. Nothing posts. We measure accuracy per field against their corrections.
Then approved records reach the workpaper beside your current process, so you can compare.
Last is switch-over: thresholds set from the shadow weeks, the runbook, and the first monthly report at month end.
When it does not work
It does not work for the one-off engagement with documents nobody has seen before, and it does not replace the reviewer. It gets the reviewer a workpaper instead of a pile.
Where to start
If intake is the first thing to fix, start with Document Intake. The same pipeline reads invoices for accounts payable (AP) desks, as in document intake automation, and loan documents for banks and lenders. If you are not sure, ask Roo, Webaroo's AI: Could AI read our clients' tax documents into workpapers?
Book 15 minutes with Connor