How to Extract Receipt Data into Excel
for Kakutei Shinkoku

Why Japanese Receipts Break General-Purpose OCR

Conventional receipt OCR operates on a simple assumption: every receipt is structured like a US or European retail receipt — a vertical list of items, one tax rate, a total at the bottom, and a date in MM/DD/YYYY format. A Japanese receipt (領収書) breaks every one of these assumptions.

The most immediate difference is the dual consumption tax rate. Since October 2019, Japan applies a standard 10% rate to most goods and services, and a reduced 8% rate to food (excluding alcohol and dining out) and newspaper subscriptions. On a single receipt from a convenience store, you might see three tax subtotals: 10% taxable (10%対象), 8% taxable (8%対象), and 0% taxable (non-taxable items). Each category carries its own subtotal and its own tax amount — and the description of purpose column needs to indicate which items fall under the reduced rate.

A standard OCR tool trained on US receipts sees a number and labels it "tax." It doesn't split it across two rates because it doesn't know two rates exist. The result: your Excel output has wrong tax totals, which means wrong deductible amounts when you file the final tax return (確定申告, kakutei shinkoku).

The second problem is the date format. Japanese receipts use the era calendar: 令和7年6月15日, not June 15, 2025. The era name changes with each emperor (令和 began in May 2019), so a tool that doesn't understand the era system will either misread the date entirely or output a meaningless string. For a freelancer processing receipts from January through December, every single date field requires manual correction in Excel. Spend 30 seconds fixing each date on 150 receipts, and you've just burned 75 minutes on a single field.

The third problem is structural: many Japanese receipts use a horizontal-then-vertical layout that differs from the strictly vertical flow of Western receipts. The recipient name (宛名) often appears on the right side in a vertical orientation. The issuer's name and registered seal (印鑑) sit at the bottom left. Standard OCR reads top-to-bottom, left-to-right. When text runs vertically, it reads gibberish.

Then there's the handwritten receipt — still common in Japan for services like repairs, consulting, or freelance gigs where the vendor writes a receipt by hand. As noted by a user on Japan's Q&A platform Qiita, testing freee's OCR on handwritten receipts yielded about 20% accuracy on the amount field alone — four digits reduced to two, with only one digit correct. Handwriting recognition on Japanese characters requires models specifically trained on Japanese script, and most general-purpose receipt OCR was trained on printed English.

What a Legally Valid Japanese Receipt Must Contain (NTA Requirements)

Before you extract anything, you need to know what you're looking for. The National Tax Agency (国税庁, NTA) defines specific requirements for a receipt (領収書) to be accepted as proof of expense. Understanding these fields determines which columns you need in your extraction output.

A valid receipt must contain eight essential elements according to NTA guidelines:

Under Japan's Electronic Bookkeeping Act (電子帳簿保存法), in full enforcement since January 2024, any receipt received electronically (email PDF, website download, app screenshot) must be stored as electronic data — you cannot print it out and keep only the paper. The data must be searchable by date, amount, and counterparty, and the system must prevent unauthorized alterations. These requirements apply to all corporations and sole proprietors, regardless of business size, per NTA enforcement guidelines published alongside the 2023 tax reform amendments.

This means if you receive a PDF receipt from Amazon Japan by email, keeping just a printed copy is non-compliant. You need the digital original. An extraction workflow that creates a searchable Excel database of your receipts while preserving the original digital files addresses both the extraction need and the compliance requirement simultaneously.

The Hidden Cost of Manual Receipt Entry During Tax Season

According to the Cabinet Secretariat's estimates, Japan's freelance population reached 4.62 million people — encompassing sole proprietors (個人事業主), one-person company presidents, and parallel workers with side businesses. Small businesses, defined as those with 20 or fewer employees, account for 84.5% of all enterprises in Japan (Ministry of Internal Affairs and Communications, 2014 Economic Census). These numbers mean millions of people file the final tax return (確定申告) each year — and for every one of them, receipts are the most tedious part of the process.

The NTA reports that 23.39 million tax returns were filed for 2024 income, with 13.29 million filed via e-Tax from home. The deadline is fixed: February 16 through March 15 every year. For the two months leading up to this window, freelancers and small business owners across Japan face the same ritual: reconcile a year's worth of receipts against bank statements and credit card records, enter each one into accounting software or a spreadsheet, and pray nothing gets lost.

Let's quantify what manual entry actually costs. The average manual data entry speed for receipt data is roughly 45 seconds per receipt if you are typing directly into a spreadsheet — finding the date, converting it from era format, typing the vendor name, entering the amount, splitting the tax, reading the description of purpose (但し書き), and guessing the expense category. At 150 receipts per year (fewer than 3 per week), that's 112.5 minutes of pure data entry — just under two hours. But the real cost is higher.

Each receipt requires context switching: the one from Yoshinoya is a thermal slip with faded text, the one from a client lunch is a handwritten receipt with barely legible kanji, the one from Amazon Japan is a PDF with the amount buried in a confirmation email. The mental switching cost between formats adds an estimated 15–20 seconds per receipt. A receipt with era-date conversion, dual tax splitting, and handwritten vendor name can take 90+ seconds. At that rate, 150 receipts consume most of a workday — work you're not billing a client for.

And this doesn't account for the error cost. The NTA's 2024 filing data shows 8.24 million of the 13.29 million e-Tax returns were filed by taxpayers themselves — meaning over 8 million people are doing their own accounting without a licensed tax accountant (税理士). The error rate in expense classification alone, when done manually without automated validation, can result in missed deductions or audit triggers. During a tax audit, the NTA can request receipts going back 7 years for Blue Return filers and 5 years for White Return — a retrieval task that's nearly impossible if your records exist only as a shoebox of faded paper slips.

How AI Extraction Handles Japanese Receipts Differently from Template OCR

The core distinction between traditional OCR and AI-powered extraction isn't about accuracy percentage — it's about how each approach finds data on the page. And on Japanese receipts, this distinction determines whether you get usable output or a blank spreadsheet.

Template-based OCR works by coordinate matching. You define a zone on the page where the total amount appears on Store A's receipt format. The software looks in that zone every time. When Store B uses a different layout — different coordinates for the total — the template breaks. For Japanese receipts, where every shop, restaurant, and taxi company uses a different format, template-based OCR requires a separate template per vendor. A freelancer with receipts from 40 different vendors needs 40 templates. This is why template OCR fails at scale.

Column-name extraction — the approach used by AI vision models — works differently. Instead of telling the tool where on the page to look, you tell it what you want: the date, the vendor name, the total amount, the description of purpose (但し書き), the 10% tax subtotal, the 8% tax subtotal. You type these as column names before uploading. The AI reads the entire document, understands what each field means semantically, and locates the corresponding value regardless of where it appears on the page. A date in the top-right corner of one receipt and in the middle of another both resolve to the "date" column because the AI recognizes both as dates.

This is the mechanism that makes it possible to batch-process 50 receipts from 50 different stores in one upload — no templates, no per-vendor configuration, no coordinate-drawing. The extraction logic is format-independent because it's meaning-based, not position-based.

For Japanese receipts specifically, this approach handles the three format-level challenges that break template OCR:

• Dual tax rates: You create two columns — 消費税10%対象額 and 消費税8%対象額 — and the AI separates the totals by recognizing the tax rate labels next to each subtotal, not by assuming a single tax field.
• Era dates: You can use a computed column — a feature that performs calculations during extraction rather than requiring post-processing in Excel. Define a column like "Western date (Convert 令和 year to Western by adding 2018)" and the AI outputs 2025-06-15 instead of 令和7年6月15日. No manual conversion step.
• Description of purpose extraction: The AI reads the description field and extracts it as text. You can pair this with an inferred column — a column where the AI makes a judgment call based on document content. For example, a column named Category (Options: 飲食費/交通費/消耗品費/通信費/その他) tells the AI to read the description, determine the expense category, and fill in the appropriate value. A receipt with "お食事代として" gets categorized as dining expenses automatically — the extraction and classification happen in a single pass.

Step-by-Step: From a Stack of Japanese Receipts to a Tax-Ready Spreadsheet

Here's the exact workflow that turns a year-end pile of receipts into a single, sorted Excel file you can use directly for the final tax return (確定申告) — whether you're filing via freee, Yayoi (弥生), MoneyForward, or handing a spreadsheet to your licensed tax accountant (税理士).

Step 1: Digitize Everything First

Before any extraction, every receipt needs to be in digital format. For paper receipts, use your phone camera — place the receipt flat on a dark surface, ensure the full receipt is visible (vendor name through total), and avoid hand shadows. For thermal paper receipts that have already started to fade (common after 6–12 months of storage), take the photo now — the image will not get better with time. For email receipts and PDFs from online purchases, save them to a single folder before uploading.

Name your files sensibly — a convention like YYYY-MM-DD_Vendor.pdf (e.g., 2025-06-15_Yoshinoya.pdf) makes each file findable later. If you have too many to rename manually, most extraction tools preserve the original filename in the output spreadsheet, giving you a traceable link from each data row back to its source document.

Step 2: Define Your Columns Once

This is the step that determines the quality of your output. Instead of extracting every field on every receipt, extract exactly the fields you need for your tax return. For a typical freelancer or sole proprietor filing the Blue Return (青色申告), the columns might be:

For White Return (白色申告) filers, the requirements are simpler — you don't need the granular tax breakdown since White Return doesn't support the Blue Return special deductions. A basic column set of Date, Vendor, Amount, and Category is typically sufficient.

The key advantage of defining columns this way: once you set them up, they work for every receipt you upload — from a 7-Eleven thermal slip to a handwritten receipt from a freelance interpreter.

Step 3: Upload in One Batch and Extract

Upload all your receipt files at once — 30, 80, 150, whatever you've accumulated. The AI processes each one, locates the values matching your defined columns, and populates a spreadsheet where each row is one receipt and each column is one of the fields you specified. Processing takes a few seconds per page: a 100-receipt batch completes in minutes, not hours.

The output includes a source-file reference for each row, so you can click from the spreadsheet back to the original receipt image if you need to verify a value. This traceability is critical for tax audit readiness — it satisfies the NTA's requirement that records be searchable by date, amount, and counterparty under the Electronic Bookkeeping Act (電子帳簿保存法).

Step 4: Validate the Output Before Filing

No extraction tool achieves 100% accuracy on every receipt — that's a reality of document processing, not a shortcoming of any specific tool. The goal of batch extraction is to reduce the manual work from "type every single field" to "spot-check outliers." Here's a two-minute validation routine:

1. Sort the output by Amount (descending). Visually scan the top 10 rows — these are your largest expenses and carry the highest audit risk. Verify each one against its original receipt.
2. Scan the Category column for any "other" classifications — the AI uses this catch-all when it can't confidently determine the category. Reclassify these manually.
3. Sort by Date and verify no entries fall outside the current tax year (January 1–December 31). A receipt from January 2026 doesn't belong in your 2025 filing.
4. For Blue Return filers: check that the sum of the 10% taxable subtotal + 8% taxable subtotal + consumption tax roughly equals the total amount. A discrepancy flags a potential extraction error.

Structuring Your Output for Blue Return vs. White Return Filing

The choice between Blue Return (青色申告) and White Return (白色申告) affects both your tax outcome and your receipt-processing strategy. Understanding the difference helps you design your extraction columns to capture the right data for your filing type.

Blue Return offers a special deduction of up to ¥650,000 for filers who submit electronically via e-Tax and maintain double-entry bookkeeping (複式簿記). Starting from 2027, the maximum deduction increases to ¥750,000 for those who combine e-Tax with "excellent electronic records" — but paper filers will be limited to just ¥100,000. Blue Return also allows loss carryforward for up to three years and lets you deduct salaries paid to family employees who work in your business. The trade-off: you must maintain a general ledger (総勘定元帳), journal (仕訳帳), and balance sheet, and retain all receipts and records for 7 years.

White Return is simpler — cash-basis records, no balance sheet required, receipt retention for 5 years — but offers no special deduction and no loss carryforward. For a sole proprietor with ¥4 million in business income, Blue Return's ¥650,000 deduction at the marginal tax rate (20% for income between ¥3.3M and ¥6.95M) means roughly ¥130,000 in tax savings — real money for a solo operation.

What does this mean for your extraction workflow? Blue Return filers need granular data — splitting each receipt by tax rate, tracking payment method (cash vs. card), and mapping to the correct expense account category (勘定科目) for journal entries. The output spreadsheet becomes the source data for your Blue Return Financial Statement (青色申告決算書). White Return filers need less detail — vendor, date, amount, and a broad category are sufficient for the Statement of Earnings and Expenses (収支内訳書).

If you're using freee, Yayoi, or MoneyForward, the process works differently: these platforms can auto-import bank and credit card transactions, and many offer receipt OCR within the app. However, the receipt OCR in these accounting platforms is designed for processing individual receipts — one at a time, within the app's mobile interface. For a freelancer with 100+ receipts accumulated over the year, processing them one by one through a mobile app is faster than manual typing but still a linear, per-receipt workflow. A batch extraction tool that processes all receipts in one upload and outputs a single spreadsheet gives you the full dataset at once — which you can then reconcile against the bank transactions already imported into your accounting software.

Frequently Asked Questions

Can AI really read handwritten Japanese receipts?

Partially. Modern vision-language models can read printed Japanese with high accuracy, but handwritten kanji remains the hardest test case. A model like the open-source Japanese-Receipt-VL-3B — fine-tuned specifically on Japanese mobile receipt images — handles standard printed receipts well. Handwritten fields with clear, block-style writing are typically legible to AI. But cursive or highly stylized handwritten receipts — particularly from older receipts where the vendor used a personal stamp and brush writing — will have lower extraction accuracy and should be spot-checked. If you frequently receive handwritten receipts, test a sample batch before committing to a workflow.

Does the tool handle Japanese era dates (令和, 平成, 昭和) automatically?

Most AI extraction tools extract the date as it appears on the document. If the receipt says 令和7年6月15日, that's what appears in your spreadsheet — which isn't directly usable in most accounting software. You can handle this with a computed column: define a rule like "Convert 令和 year to Western by adding 2018, 平成 by adding 1988, 昭和 by adding 1925" and the AI outputs the Western date. If your tool doesn't support computed columns, you can run a one-time Excel formula (=DATE(era_year+2018, month, day) for Reiwa dates) across the date column after extraction.

How accurately does it split 8% and 10% tax on receipts?

When the receipt clearly labels the 8% and 10% subtotals separately — as most Japanese convenience store and supermarket receipts do — AI extraction handles the split reliably. When a receipt shows only a single total without tax rate breakdown (common on older handwritten receipts or small vendor receipts), the AI can't split what isn't there. In that case, you'll need to calculate the tax breakdown manually based on the contents: food items at 8%, other items at 10%. Receipts issued under the Qualified Invoice System (インボイス制度) — mandatory since October 2023 for B2B transactions — are required to show the tax breakdown by rate, making extraction more reliable for business receipts received since then.

What about faded thermal paper receipts? Can AI read those?

It depends on the degree of fading. Thermal paper receipts typically begin to degrade within 6–12 months — the text fades from black to gray to invisible as the chemical coating breaks down. If the receipt is already mostly blank, no extraction tool can recover data that the human eye can't see. If the text is faint but still legible, AI vision models often handle it better than traditional OCR because they process the image holistically — inferring text from surrounding context — rather than relying solely on pixel-level contrast. The practical takeaway: digitize your receipts as soon as you receive them. Once a thermal receipt has faded, the data is gone. The Ministry of Economy, Trade and Industry (METI) recommends scanning receipts within 3 months of issue for this reason.

Can I import the output into freee or Yayoi instead of Excel?

freee, Yayoi (弥生), and MoneyForward all support CSV import for transaction data. Once your receipts are extracted into a spreadsheet, export as CSV and import into your accounting platform — matching the column order to the import format required by your software. freee's import template expects columns for date, amount, account item (勘定科目), description (摘要), and counterparty (取引先). The output from a well-structured extraction run maps directly to these fields. The import step replaces manual transaction-by-transaction entry in the accounting app — you upload one CSV file instead of entering 150 individual receipts.

Does an extracted spreadsheet satisfy NTA audit requirements?

The NTA does not require you to keep paper originals of receipts that were received electronically — the Electronic Bookkeeping Act mandates electronic storage for documents received electronically. For paper receipts that you digitize by scanning, the scanned copies serve as supporting evidence, but the NTA may request to see the originals during an audit. In practice, an extraction spreadsheet with source-file references (linking each row to the original receipt image) combined with a well-organized digital filing system satisfies the requirement that records be searchable by date, amount, and counterparty. The key is traceability: every data point in your tax return must be traceable back to a specific source document that can be retrieved on demand.

Making Kakutei Shinkoku Something You Survive, Not Dread

Here's what changes when you stop typing receipts by hand. A freelancer or small business owner with 150 annual receipts goes from spending a full workday on data entry — 6 to 8 hours, accounting for format-switching costs and error correction — to spending roughly 15 minutes on extraction plus 30 minutes on validation. The extraction itself takes a few minutes. The validation — sorting, scanning outliers, reclassifying unknowns — takes half an hour if you've been diligent about naming files and capturing clear images through the year. Total: under an hour versus most of a day.

This isn't about saving an afternoon. It's about removing the single largest anxiety source from the final tax return process — the fear that you've lost a receipt, misclassified an expense, or missed a deduction because the shoebox was too overwhelming to sort through properly. When every receipt lives in a spreadsheet with its date, vendor, amount, tax breakdown, category, and a link back to the original image, the tax return becomes a reconciliation exercise rather than a reconstruction project. You're verifying data, not creating it from scratch.

The Blue Return special deduction (青色申告特別控除) — worth up to ¥650,000 in taxable income reduction — requires clean records. A spreadsheet exports directly into your Blue Return Financial Statement's expense section, with every line backed by a traceable source file. That's the difference between filing with confidence and filing with crossed fingers.

Test the workflow on your own receipts. See if a stack of receipts that would take an afternoon to type becomes a spreadsheet you can validate over coffee.