Since around May 2026, an AI agent has been doing most of the monthly bookkeeping for my one-person business: working through unprocessed bank and card lines in freee, the cloud accounting service most small businesses in Japan use, drafting invoices, collecting receipt PDFs, matching incoming payments against invoices, and checking withholding-tax differences.
The part that made this work is not that the agent knows tax law. It does not decide anything. A human makes each judgment the first time it comes up; from then on the agent reproduces that judgment against the recorded rule, the evidence, and the prior examples — and stops on anything it has not seen before.
This is my own operation log as of July 2026, not general accounting advice. freee is a Japanese cloud accounting service for businesses that file in Japan; its interface, help center and support are Japanese-only, and the feature names below are my translations of the Japanese UI (automated bookkeeping is 自動で経理, the file box is ファイルボックス). The tax handling described here is specific to Japan. The part that transfers is the operating method, not the tool: decide once, record the rule, stop at the unknown.
Originally published in Japanese on July 20, 2026. This English version was written in September 2026.
Separate the judgment from the work
The first split was between decisions and execution.
Is this card charge a business expense or an owner’s draw? Communications or books? How is revenue subject to withholding tax handled? None of that is left to the agent. I decide the treatment once, then record the account, tax category, item, how the supporting document is obtained, and how to verify the entry afterwards, in docs.
From the next month, the agent works from those docs and from prior examples. What it receives is a reproducible monthly procedure, not a tax judgment.
Blur that line and it gets dangerous. Assume the agent already knows, and it will classify a line item carelessly off the merchant name alone. Keep first-time judgment and recorded rule separate, and the monthly work becomes genuinely delegable.
Working the unprocessed transaction list
The clearest monthly case is the credit card statement. freee’s automated bookkeeping feature lists transactions imported from linked cards and bank accounts; freee’s own help center describes it as bookkeeping based on those imported lines.
A list is not the end of the job. Each line still needs a decision: have I handled this before, does it need a supporting document, which account does it go to?
In my setup the agent picks up only lines that match an existing rule — cloud services, network and phone charges, domains, advertising, books, transport, the recurring things whose treatment is already settled.
Anything unfamiliar is not processed. It does not guess “this is probably a communications expense.” It stops and returns the line as a question.
That turned out to be the important property. What delegated bookkeeping needs is not an agent that processes everything, but one that can tell the difference between what it may process and where it must stop.
Collecting the evidence changed the shape of the job
The heavy part of bookkeeping is not the journal entry. It is finding the invoice PDF, opening the usage statement, pulling a qualified invoice out of a purchase history, saving a screenshot as evidence, putting it into freee’s file box, and attaching it to the transaction.
freee’s file box takes in images, CSVs, and PDFs of receipts and invoices for document management and bookkeeping, and pairs naturally with attaching a file to a transaction.
This is where delegation paid off most. AWS invoices from the billing console. Google Cloud billing invoices. Statement PDFs from network and phone providers. Domain renewals matched against the right domain, amount, and date. Qualified invoices for ebooks and audio services from order history.
Anywhere that needs a login or MFA, a human does it. No credentials are given to the agent. But once a screen is open, where to look, what to save, and which freee transaction to attach it to has become reproducible on the agent side.
With evidence gathered, the rest moves easily. And where evidence cannot be obtained, there is now a clear reason to stop.
Invoicing follows a monthly template
Recurring retainer amounts, the month covered, issue date, payment due date, whether withholding applies, remarks carried over from last month — each client has a shape.
The agent reads last month’s invoice and the docs and assembles the draft of the next one. But creating or editing an invoice in freee changes data, so that step goes through approval as a rule. The agent organizes the content; a human confirms it and executes only what needs executing.
Same division as before: reproduction against an agreed rule is delegated, final judgment and approval are not.
Payment matching and withholding tax
On the revenue side, incoming bank lines are reconciled against invoices and remittance advices. Issuing an invoice is not the end: when payment arrives, the open receivable has to be cleared, and where withholding applies, the gap between invoiced and received amounts has to be accounted for.
In Japan, certain categories of payment to a sole proprietor — professional fees among them — are subject to withholding income tax deducted at source by the payer, so the bank credit is smaller than the invoice by design.
freee’s help center documents clearing unsettled transactions as the operation covering receivable collection and payable settlement, and separately documents how a sole proprietor registers revenue transactions subject to withholding and adjusts the difference between invoiced and received amounts.
The agent reads the bank lines, the invoices, the remittance advices, and prior examples, then checks: which invoice does this payment belong to, do the amounts agree, is the difference the withholding amount, is the evidence PDF attached, and has the unsettled state actually cleared after registration?
Creating a transaction through the API does not count as done. Done means the unprocessed line is gone in freee’s own UI, the transaction is registered, the evidence is linked, and the states agree. Unglamorous, and increasingly important the more you delegate — because “the script ran” and “the books are correct” are different claims.
The thing that mattered most was writing it down
Being able to drive freee’s UI and API matters. But the reason the agent can repeat the same month reliably is that the previous decisions are still written down: how a line was treated, where the evidence comes from, which account, which tax category, what to verify after registration, and where to stop when unsure.
The following month, I no longer try to remember what I did last time — the agent reads the docs and reproduces it. Bookkeeping did not get shorter so much as it stopped depending on my memory. The monthly work moved from recalling to reviewing.
Stopping is the feature
What is frightening about AI bookkeeping is an agent that keeps going: classifying an unknown line plausibly, registering an entry with no evidence, deviating from prior treatment, or guessing at a screen it could not log into.
So the stop conditions are explicit.
- A line with no prior treatment: stop
- Evidence cannot be obtained: stop
- Amounts or dates do not agree: stop
- Any doubt about withholding or tax treatment: stop
- Login or MFA required: return to the human
- Any create, edit, or delete in freee: approval first
The more you delegate, the more you need designed stopping points. For monthly books, stopping cleanly at the unknown beats automating everything.
Bookkeeping became operations design
It used to be data entry: open the line, pick the account, hunt for the document, draft the invoice, clear the payment, check withholding. None of it hard, all of it requiring me to remember the same things every month.
Now the order is: decide once, turn it into a rule, record how the evidence is obtained, let the agent execute in subsequent months, and take back only what it could not resolve. The center of gravity moved from entry to operations design.
I can say the books are handled by an agent — not because the agent got better at tax, but because my rules, evidence, prior examples, and verification steps are written in a form it can read.
Aim for reproducibility, not full automation
Handing tax judgment to an agent with nothing recorded would be reckless, and anything genuinely specialized should be confirmed by a human or a professional. But a monthly cycle contains a lot of already-decided repetition: reviewing unprocessed lines, collecting evidence PDFs, registering to the same accounts as before, building an invoice from last month’s template, reconciling payments, and verifying state afterwards.
That part suits an agent. The goal is not to make it your accountant, but a practitioner that can read your bookkeeping rules. Decide the first time, keep the evidence, write the prior examples down, stop on the unknown, and verify at the end.
The books are handled by an agent. That does not mean the judgment was given away — it means the judgment was recorded and the work made reproducible.
References
The freee help center pages below are in Japanese, and automated checks against them return HTTP 403 (bot protection); they open normally in a browser.
- freee Help Center: Using automated bookkeeping from imported transactions (in Japanese)
- freee Help Center: About the receipt import feature, file box (in Japanese)
- freee Help Center: Clearing unsettled transactions (in Japanese)
- freee Help Center: Registering revenue transactions subject to withholding tax, sole proprietors (in Japanese)