Section 01 — Context
The client is a mid-Atlantic accounting firm: eight partners, twenty-two staff, roughly two hundred small-business clients on monthly retainers. The bottleneck, which they had named correctly going in, was vendor-bill processing — the clients’ AP documents arriving by email, by portal, by photograph from a phone, in a hundred subtly different formats per quarter.
Eighty partner-and-senior hours per quarter were going to a task neither partners nor seniors should be doing.
Section 02 — Approach
We did not propose replacing the firm’s existing review discipline. Partners take responsibility for the books they sign. The job was to push the work AI can do — read the document, propose the entries, flag the unusual ones — upstream of where the partners look.
The boundary, drawn at audit:
- AI may classify a document, extract line items, propose a coding, and reconcile against historical vendor patterns.
- AI may not post entries above a configurable dollar threshold, post entries against a flagged vendor, or close a period.
- All sensitive-vendor entries route to the assigned partner regardless of confidence. The partner sees the document, the proposed entry, and a one-line “what’s unusual here” note.
Section 03 — System
Documents flow in through a single intake address. A normalization step converts photographs and PDFs into a consistent textual representation. A classification pass identifies document type. An extraction pass produces a structured proposal — vendor, date, line items, suggested account codings — with a confidence per field.
A second AI pass, separate from extraction, performs sanity checks: does the total reconcile, does the vendor match recent entries, is the coding consistent with how this client codes similar items historically. It is a small but important detail: extraction and validation are different jobs and should not share a prompt.
High-confidence entries within threshold post directly to QuickBooks and appear on a daily digest the assigned senior reviews. Exceptions — low confidence, sensitive vendor, threshold breach, validation flag — drop into a Linear queue with everything the reviewing partner needs to decide in one glance.
Section 04 — Outcome
Cycle time per vendor bill fell 62% in the first quarter. The exception rate settled around 1.4% — meaning roughly ninety-eight in a hundred documents clear automatically, and the remaining two arrive in front of a partner with the work already done.
The eighty hours of partner-and-senior time returned per quarter were redirected to advisory work, which the firm bills at materially higher rates. The economic effect of the engagement, in other words, was not the labor savings — it was what the firm could sell with the time back.
What we’d do differently: the validation pass should have shipped on day one. We added it in week two after seeing the kinds of mistakes the extraction pass would make on novel layouts. Treat it as a foundational component, not an enhancement.