A month of supplier bills, drafted in Xero. By nobody.
A Hong Kong retail & e-commerce group used to hand-key every supplier bill into Xero at month-end. Now an automation reads each bill, checks the numbers, and leaves a review-ready draft — their accountant just approves.
Every month, the same stack of bills.
Photos and PDFs of supplier bills pile up. Someone opens Xero and keys each one in by hand — the supplier, the date, every line, the right account code — then checks the totals tie out. It's careful, repetitive work that eats a day and quietly invites small errors. Classic manual work. A good candidate for the right-sized kind of automation.
AI reads the bill. Code checks the money. A human signs off.
No black box. The AI does one job — turning a messy document into structured data. Everything that touches a number is plain, deterministic code that flags problems instead of hiding them. And nothing lands in the ledger without the accountant's approval. AI only where it earns its place.
Six steps — and a loop that keeps it sharp.
Left to right is one bill's journey. The last step is what makes month two better than month one.
Drop the bill
PDF, photo, or Excel — however the supplier sends it.
AI reads it
Runs on your own machine. Pulls supplier, dates, line items and suggested account codes.
Code checks the math
Totals must reconcile. Mismatches get flagged — numbers are never invented or edited.
Draft lands in Xero
Created as a Draft, with the source bill attached. Nothing is posted.
Accountant approves
Eyeballs it, fixes anything, approves. The human stays in charge.
It learns
Each correction becomes context for next time — sharper per supplier, month over month.
The loop: every correction from step 05 flows back into step 02. Next month, on that supplier's bills, it needs fewer of them. The model isn't retrained — it's given a growing memory of how your books should read.
Automation you'd actually trust with the books.
Human-in-the-loop
Every draft waits for your accountant. The automation proposes; a person decides.
Deterministic on the numbers
Math is checked by code, not guessed by a model. Discrepancies get flagged, not buried.
It gets better
Corrections become memory. Accuracy climbs per company and per supplier over time.
Yours, and low-cost
Runs locally on your own hardware. No per-token bill, and the financial data stays with you.
Early, and honest.
bills in the first pilot batch matched the reviewer on type, supplier and total — including one read straight from a phone photo.
numbers invented. On a real bill, the math check caught a discrepancy a model would have glossed over.
per-token API cost. The reading model runs on a MacBook — no cloud bill, no data leaving the desk.
It's early, and the sample is small — which is exactly why we measure accuracy on every real batch and only widen scope when the numbers earn it. Same discovery-first discipline as every build.
This is how every Pawtomation build starts.
Map the real process. Find the highest-leverage, lowest-risk slice. Build the simplest thing that works — automation before AI, a person on the calls that matter. Scoped and priced from a fixed Automation Audit, not a guess. Not an AI agency — an operator who builds, with an accountant's eye on the numbers.
Got a process that looks like this?
If a corner of your operation is still done by hand, the Audit is where we work out what's worth automating — and, just as often, what isn't.
Book an Audit →Fixed fee, credited toward a build · HK-based · Operator-led