// blueprint

A bookkeeping firm, six people.

Two people spend every morning chasing documents and typing numbers into the system.

// what they typedsix-person bookkeeping firm. clients email statements and photos of receipts, some of it comes by text. two of my people spend every morning chasing missing documents and typing numbers into our system. month-end is brutal.

// read

The typing is the visible cost and the chasing is the expensive one. The work is document intake, and the part you cannot automate is the judgment at the end of it.

// the system

A scheduled job reads the intake inbox, classifies each attachment, and extracts the fields with a confidence score attached to every one.
Anything above your confidence line is filed and queued. Anything below it goes to a person, with the document and the extracted values side by side.
The chase runs itself. A missing December statement produces a drafted message, sent on a schedule you set, tracked, and escalated to a person after the second try.
A nightly reconciliation compares what was filed against the source records and reports only the differences, each linked to the document it came from.
Nothing posts to the books without a person approving it.

// buy vs build

Extraction is a buy. Dext, Hubdoc and Rossum already do document capture, and one of them probably covers most of your document types. Verify two things: that the API is on the plan you would buy, and what happens to the documents they cannot read.

The build is everything around it: the confidence line, the exception queue, the chase, and the reconciliation that ties a filed number back to its document. That part names your firm rather than a category.

// cost band

Build $24,000–$33,500, and less if you keep it to intake first. Running it is $110–$275/mo plus $1,500/mo if we operate it.

// what we'd watch

A model that is 95% right on extraction is useless at close if nobody knows which 5% is wrong. The confidence score and the exception queue are the product, not the extraction.
Client documents are sensitive. The pipeline runs in your accounts, and the retention rule gets written down before anything is built.
Month-end is a spike. The system gets sized and tested against your worst week, not your average one.

// next

Scope one workflow, prove it through a real close, then expand. The first conversation is free.

This is a worked example on a composite business, drawn by the same desk that runs on the architect. Describe your own situation there and it will draw yours.

studio@anchor163.com // copied

// more architect services ai automation cost build vs buy ai automation