How AI Bookkeeper works

Step 1: Connect a bank feed, or upload a document

Everything starts with two inputs: your bank feed and your paperwork. Connect a bank account and transactions sync in automatically — no CSV exports, no manual entry. Upload bills, receipts, and invoices as PDFs or images, straight from wherever they land.

Signing up takes an email and password, or a Google account. Once a bank feed is connected, it keeps syncing on its own; documents you can drop in whenever they arrive.

Step 2: The AI first pass

The AI does the first pass on everything, and it scores its own certainty as it goes. Every bank transaction that syncs in gets categorized, and every categorization gets a confidence score attached. The score is not decoration — it decides what happens next in the pipeline.

Uploaded documents go through OCR plus a language model that reads the file the way a bookkeeper would: it extracts the vendor, totals, tax, dates, and individual line items as discrete fields, not one undifferentiated blob of text. A crumpled receipt photo and a clean PDF invoice come out the other side as the same structured data.

Step 3: Provenance — click any field, see where it came from

Every extracted value links back to the exact spot on the original document it was read from. Click the vendor name, the total, the tax amount, or any line item, and the source document opens with that region highlighted — a bounding box drawn around the precise place the value came from.

This is the part most AI bookkeeping tools skip, and it is the reason we describe the product as the AI bookkeeper that shows its work. An extraction you cannot trace is an extraction you have to either trust blindly or re-check by hand, which defeats the point of automating it. With click-to-highlight provenance, verifying a field takes a glance instead of a hunt through the PDF.

It matters most when something looks off. If a total seems wrong, you do not debug the AI — you click the field, look at the highlighted region, and see in seconds whether the document really says that or the extraction missed.

Step 4: The review queue — exceptions only

You only review what the AI is not sure about. Items that score below the confidence threshold route to a human review queue; high-confidence items post without needing your attention. The work in front of you is the exceptions, not the whole pile.

In the queue, you approve or reject each item and can attach a memo explaining the call. Every action — every approval, rejection, and memo — lands in an append-only audit trail, so there is always a record of who decided what, and why.

Step 5: The ledger underneath

Underneath it all is a real double-entry ledger, not a spreadsheet with formulas. Approved entries post as proper double-entry records, which means the books balance by construction and stand up to an accountant's scrutiny.

Each client's books are fully isolated. Separate organizations and businesses each have their own ledger with per-business access control, so a bookkeeper managing multiple clients never has one client's data anywhere near another's. Combined with the append-only audit trail, the result is books where every number has a source and every change has a record.

What's next

QuickBooks Online sync is in development. When it ships, the ledger described above will sync with QBO rather than living alongside it.

We are early-stage, and we would rather say that plainly than pretend otherwise. Pricing will be published at launch — no sales calls required.