Automate receipt entry in an accounting firm, with exceptions
A Fukuoka tax accountant had AI run weekly receipt sorting, reading and import, holding back only flagged items for one email. No tool cost or time saved is given.
Yuji Ido, a tax accountant with a 12-person firm in Hakata, Fukuoka, kept an existing receipt-reading tool that had no correction function and no direct import. He had AI check and correct its output from the outside. The first AI-designed workflow still relied on staff sorting receipts from invoices by hand. The second sorted automatically but waited for his approval every week. The third runs late on Sunday night: staff only scan receipts into a set location, then AI runs the in-house sorting software, reads, inspects and imports. Items are held back only if the account title is unclear, the date falls outside the accounting period, the invoice registration number is unreadable, a duplicate is possible, the amount does not match the card statement, or personal use is possible. Those are sent to him in one weekly email.
How it went
As reported by the author, human input is now needed in two places only: staff placing scanned receipts, and the owner replying to the weekly exceptions email. Unflagged items import without approval. The post gives no figures for hours saved, error rates or cost, and describes the result as the current setup, not a measured outcome.
Worth knowing
The post does not name the AI model, the receipt tool or the accounting software, and gives no cost or accuracy figures. The sorting software was built in-house with AI, which not every firm can copy. Importing without approval puts the duty to check on the firm, so the exception rules need testing and spot checks. The Japanese invoice registration number rule is specific to Japan.
Try it yourself
Paste into ChatGPT, Claude or another AI assistant and fill in the square brackets.
Act as a bookkeeping workflow reviewer for [type of practice]. Our receipt tool is [tool name] and our accounting software is [software name]. Step 1: list what each tool can already do before suggesting any manual step. Step 2: design a weekly run where staff only scan receipts into [folder], and the AI sorts, reads, checks and prepares the import. Step 3: write the exception rules that hold an item back, for example unclear account, date outside [accounting period], unreadable [tax registration number], possible duplicate, amount not matching [card or bank statement], possible personal use. Step 4: draft one weekly email that lists only the held items, with a suggested answer for each. Do not import anything yet. Show me the workflow, rules and a sample email for review first, then test on [number] past receipts and spot-check the results before going live.
Source: note.com · Oct 28, 2026. Summarised in our own words; read the original for the full account.
Useful example?
Worked ideas for accountancy practices
Checked prices, setup steps and what each one is worth.
Fathom Pro for 30 clients is about $533£401€477A$770C$754 a month ($450 plus $17 for each of 5 extra companies). Editable AI commentary and branded reports.
Copilot recaps client meetings in Teams with who said what and the actions, checked by the partner. From about $43£32€38A$61C$60 a month for 2 users, plus Microsoft 365.
A third-generation izakaya owner in rural Japan had Claude enter cash receipts into a ledger. Date errors were fixed by adding one line to a written rule sheet.
An Iowa grocer photographs wholesaler catalog pages and his staff schedule and sends them to Meta's Muse agent. The ad now takes about 15 minutes instead of two hours, he says.
A multi-store salon operator listed 9 tasks for Claude Code. The first one built, auto-updating hairstyle posts, wasn't on the list. Login and API blocks stalled the sales check.
A small business owner on Hacker News says Codex now handles tax, invoice and insurance admin once outsourced or neglected, with a human check and an accountant's sign-off.