AI Agents for Accounting: Receipts, Journal Entries and E-Records Law

The tool features and regulations in this article were checked against official pages and National Tax Agency (NTA) materials on October 2, 2026. For tax decisions, check with your tax accountant or the NTA.

When you automate accounting work with an AI agent, do not hand over everything. Give it reading and drafting. People keep finalizing journal entries and deciding how to store records in line with the law. With this split in mind, let's look at where to start.

Accounting Tasks That Are Easy and Hard to Automate

Tasks that are easy to automate have a fixed format and produce results people can compare against the source. Examples are reading the date, amount, and vendor from receipts and invoices, suggesting account titles, and saving files under set names.

Tasks that are hard to automate need context to decide. For transactions where it is unclear whether an expense relates to the business or which tax category applies, the AI can suggest options, but people decide. People also finalize year-end closing and tax filing.

From Reading Receipts to Journal Entries

The flow breaks into five stages: receive, read, draft, human approval, and register in the accounting software. At every stage, keep the original file; never delete it.

Building It with n8n

To build an accounting flow in n8n, you can use the Gmail Trigger or Google Drive Trigger to pick up files that arrive by email or in a folder. The Extract From File node pulls text out of PDFs. To pick dates and amounts out of that text, use the Information Extractor node. This node turns text into structured data, with the fields you want defined by attribute descriptions, a JSON example, or a JSON schema.

When you handle receipts in n8n, record the extracted results and the original file in the same row. If a spreadsheet lists the file location, date, amount, vendor, and suggested account title side by side, people can open the original file right away when approving.

Building It with Dify

To build an accounting flow in Dify, you can use the LLM node. Dify's LLM node accepts image and document input and supports structured output with a defined JSON shape. To read receipt images, choose a model that can handle images. For human approval, use the Human Input node, which pauses the workflow and sends a form. The form can also go out by email, and the recipient can choose to approve or send it back.

Connecting Accounting Software via API

Pass the extracted results to your accounting software through its API or through an import file. For three major Japanese accounting products, these are the methods confirmed in official materials.

When you connect an AI agent to freee, start with uploading receipts and drafting transactions so people can review before anything is finalized. You can check the features of the freee accounting API in the API schema freee publishes. The Money Forward Cloud developer site describes exchanging an API key for a JWT to make calls, and presents this as intended for AI agents and automation scripts. To call it from n8n, use the HTTP Request node.

The Money Forward developer site posted a notice on September 24, 2026 that the Cloud Accounting API and MCP server now support API key authentication. There is also a notice from August 3, 2026 that Cloud Accounting supports Anthropic's official Claude connector. Check this developer site for the latest on connecting AI agents to Money Forward.

Yayoi Accounting Next imports files with set fields in a set order. Dates outside the accounting period being processed cannot be imported. If you build the CSV with n8n's Convert to File node and have a person review it before importing, you add a check before import. We explain API connections in How to Connect Services in n8n.

What to Check Under Japan's E-Records Law and Invoice System

According to the NTA, transaction information from electronic transactions must be stored as electronic data under the Electronic Books Maintenance Act (Japan). Electronic transactions include receiving transaction information by email (including attachments) and receiving it through websites. What must be stored is not limited to invoices; it also includes purchase orders, contracts, receipts, and quotes.

• Take measures to prevent tampering: a timestamp, a system that keeps a history of corrections and deletions, or internal administrative rules
• Have a display and printer available for viewing the stored data
• Make the data searchable by three keys: date, amount, and counterparty

As ways to make data searchable, the NTA gives two examples: building an index in a spreadsheet, or naming files in the order of date, amount, and counterparty and keeping them in a specific folder. The date, amount, and counterparty the AI agent extracts can be used directly for file names and the index. If you scan and store paper receipts, check the scanner storage requirements in NTA materials.

In some cases, such as when sales in the base period are ¥50 million or less, you do not need to meet the search requirement as long as you can respond to requests to download the data during a tax audit. There is also a grace measure: if the head of the competent tax office recognizes that there was a reasonable cause for not meeting the standard rules, and you can respond to requests to download the data and present printed copies, simply storing the data is enough. Check NTA materials to see which applies to you.

Under Japan's Qualified Invoice System, you generally must keep qualified invoices to claim input tax credits for consumption tax. An invoice shows the seller's registration number, the transaction date, the total amount by tax rate and the applicable rate, the consumption tax amount, and more. Registration numbers can be checked on the NTA's Qualified Invoice Issuer Publication Site, which also offers a Web API. Having the AI agent read the registration number and match it against the publication site is easy to automate. Note that for purchases without an invoice, a transitional measure allows a set percentage of the input tax amount to be credited until the end of September 2031. Do not leave how to handle this transitional measure to the AI; people decide.

Deciding Where People Review

If people review every item, automation loses much of its value. Set conditions for which transactions go to a person. For transactions that meet none of the conditions, pull a few at month-end and compare them against the original files. When you find a reading error, add that case to the conditions.

Also decide which operations in the accounting software the AI agent may perform. Start by allowing only receipt uploads and draft creation, not finalizing or deleting. We explain how to think about permissions in AI Agent Permission Management, and how AI agents differ from RPA, which repeats fixed operations, in AI Agents vs. RPA.

The Back Office Category on Employee Store

Employee Store is a marketplace where companies can adopt AI agents built by developers. One of its categories is back office. Sellers can list in any format, including n8n or Dify workflows and their own custom-built agents.

Pricing is one-time purchase, monthly, or setup fee plus monthly, and payment is through Stripe. After purchase, you message the seller in the deal room, check the deliverables, and accept them. Delivery is, in principle, within 3 business days, and you are refunded if it runs late. Before you adopt an agent, check on the listing page which accounting software it supports and which operations it performs automatically.