AI Agent Accounting: Tax Accountants × Generative AI

From an “Input-Based” Tax Accounting Industry to an Era of Design and Approval

1. What Kind of Era Is the Tax Accounting Industry Entering?

In the tax accounting industry going forward, the key issue will no longer be the sheer volume of work itself.

What will matter is how each workflow is designed, which parts should be reviewed by humans, and how much can safely be entrusted to AI.

Traditionally, accounting firms have centered their work around collecting documents, entering data into accounting software, checking the data, and preparing monthly reports.

However, once AI agents are able to connect with business tools and files, data entry, initial checking, and draft report preparation will increasingly become areas that AI can handle.

As a result, we believe the gap will widen between accounting firms that use AI merely as a chat-based consultation tool and those that embed AI into their actual workflows and turn it into a reproducible operating system.

The important point in this figure is that simply being able to “use” AI is not enough.

To create real value as an accounting firm, it is necessary to envision what AI can do and then design the business procedures, authority settings, review flows, and client-specific rules required to make AI work safely and consistently.

At JGA, we will first create an internal environment where staff can use AI to perform actual work. From there, we will build an organization that understands AI agents and is capable of redesigning business workflows around them.

2. Why Is AI Agent Accounting Necessary?

The foundation of AI Agent Accounting is the mechanism that connects AI with business tools.

Until now, accounting software, storage systems, chat tools, and project management tools often had to be connected individually. As the number of connection points increases, operations become more complex.

This is where MCP becomes important.

MCP is a common connection interface that links AI with external tools. In the context of an accounting firm, it can be understood as a secure “gateway” through which AI accesses accounting data, supporting documents, client-specific rules, and past judgments.

The work of an accounting firm is not simply about entering data into accounting software.

It involves reading supporting documents, checking past treatments, applying client-specific rules, and, when necessary, confirming matters with clients.

For AI to support this series of tasks, it needs a mechanism that allows it to access accounting data safely.

In other words, the question for accounting firms going forward will not be whether they introduce AI.

The real question will be whether they can design a business environment in which AI can work safely.

3. JGA’s Approach: Controlled Automation, Not 100% Automation

JGA is not aiming to automate accounting work 100% from the outset.

Tax accounting work involves tax judgment, client explanations, important policy decisions, and final responsibility. These are not areas that should be handed over entirely to AI.

On the other hand, the areas that can be entrusted to AI are steadily increasing.

AI is well suited to preparing draft journal entries, checking journal entries, drafting monthly comments, extracting items that need to be confirmed with clients, and preparing first drafts of meeting minutes and proposals.

The essence of this system is that AI safely accesses accounting data and, in accordance with internal firm rules, prepares drafts, performs checks, and creates explanatory materials.

The source-of-truth data consists of MF Cloud Accounting, supporting documents, client responses, tax laws, and accounting standards.

AI outputs are not the official source of truth. They are work products for humans to review.

At JGA, we aim to build an organization where staff do not spend their time primarily on data entry.

Instead, they should spend their time on exception handling, client explanations, business improvement, and proposal activities.

AI is not positioned as a replacement for people. It is positioned as an execution partner that enables people to focus on higher-value work.

JGA’s basic policy is as follows:

AI handles drafting, checking, and document preparation.

Registration, submission, tax judgment, and important client-specific policy decisions are reviewed and approved by humans.

4. Overall Structure: Claude Code × MF Cloud Accounting × MCP Server

The overall structure of AI Agent Accounting can be summarized simply as follows.

MF Cloud Accounting is the accounting platform that holds source-of-truth data, such as journal entries, trial balances, master data, and transaction details.

The MCP Server is the connection interface that safely links AI with accounting data.

Claude Code is the AI agent that receives natural-language instructions and executes business procedures.

What matters in this structure is that AI does not freely rewrite accounting data on its own.

Instead, AI first retrieves data, prepares drafts, and then proceeds to the next process only after human review.

In actual accounting firm practice, various errors can occur, such as selecting the wrong entity, misunderstanding the accounting period, applying the wrong tax category, overlooking missing supporting documents, or missing client-specific rules.

For that reason, the more AI is used, the more important rule design and authority design become.

At JGA, we view AI Agent Accounting through three layers:

• Teach: Provide AI with internal rules, client context, and output formats.
• Connect: Enable AI to access necessary data, such as MF Cloud Accounting.
• Restrict: Prevent dangerous operations, mistaken external transmissions, excessive authority, and lack of audit trails.

When these three layers are in place, AI becomes more than just a convenient chat tool.

It becomes an execution partner that works within the accounting firm’s business workflows.

5. The Five Elements Required to Turn AI into an Accounting Firm Staff Member

When incorporating AI into accounting firm operations, it is important to clearly define what AI should know, what AI is allowed to do, and what AI must not do.

At JGA, we will develop the following five elements in order to make AI function as a member of the accounting firm staff.

CLAUDE.md is the AI’s self-introduction and the firm’s internal rulebook.

Skills are business manuals for tasks such as journal entry checks, tax category judgments, and monthly reviews.

MEMORY.md functions as a client chart that stores client-specific history and special rules.

settings.json defines authority settings.

Hooks are safety mechanisms that structurally stop dangerous operations.

Traditionally, the know-how of accounting firms has often accumulated inside the minds of experienced staff members.

However, in the age of AI, it is important to turn that tacit knowledge into explicit knowledge in the form of rules, checklists, and client-specific records.

AI is a powerful tool.

The more powerful it is, the more it needs reins.

Operations such as deletion commands, external transmissions, accounting data updates, and access to client information should not be controlled only by written instructions.

They must be stopped through authority settings and Hooks.

Simply telling people to “be careful” does not prevent accidents.

Dangerous operations must be made impossible at the structural level.

This is what JGA means by harness design.

6. The Business Workflows JGA Will Implement

The first areas JGA will address through AI Agent Accounting can be divided into three main workflows.

The first is creating draft journal entries from invoice PDFs and transaction details.

The second is checking existing journal entries.

The third is preparing monthly reports, proposals, meeting minutes, and client confirmation items.

Workflow 1: Creating Draft Journal Entries from Invoice PDFs

AI will extract transaction dates, amounts, counterparties, tax rates, and other relevant information from invoice PDFs, compare them with the accounting master data, and prepare draft journal entries.

However, in the initial phase, AI will only create draft journal entries.

Whether those entries should actually be registered will be reviewed by humans.

Areas such as tax categories, fixed assets, director or family-related transactions, entertainment expenses, donations, foreign currency transactions, and non-recurring transactions are not areas where AI should register entries too easily.

These items will be sent for client confirmation or staff review.

Workflow 2: Journal Entry Checks

The second workflow is journal entry checking.

AI will retrieve journal entries, master data, and trial balances from MF Cloud Accounting and, based on predefined checking rules, identify outliers, missing subaccounts, missing departments, unusual tax categories, duplicate journal entries, and other potential issues.

In journal entry checks, the goal is not to have AI correct everything automatically.

It is important to classify items into categories such as “skip,” “automatic correction proposal,” and “client confirmation required.”

The purpose is to clearly identify which items need review and to prepare them in a way that makes human judgment easier.

Once a checking procedure has been determined, it will be reflected in Skills and checklists.

This reduces differences in checking quality among staff members and improves the reproducibility of monthly accounting work.

Workflow 3: Monthly Reports, Proposals, and Meeting Minutes

The third workflow is the preparation of monthly reports, proposals, meeting minutes, and client confirmation items.

AI will use trial balances and trend reports to analyze fluctuations, identify outliers, and prepare draft comments.

Humans will then adjust the language so that it communicates effectively to business owners and refine the tax implications, cash flow issues, and scope of proposals.

The important point here is that AI-generated work products should not be treated as final deliverables.

AI prepares the draft.

Humans edit it and finalize it in a form that can be properly communicated to the client.

This division of roles makes it possible to shorten preparation time while improving the quality of client explanations.

7. Why Guardrail Design Will Become a Source of Competitiveness

The most important factor in AI Agent Accounting is not the intelligence of the AI itself.

It is the design of the business environment in which the AI operates.

No matter how advanced the AI is, it cannot be used in practice if it selects the wrong entity, registers journal entries incorrectly, or sends client information externally by mistake.

For that reason, JGA will design guardrails for each type of risk.

These include confirming the target entity, requiring approval before registration, confirming external transmission, restricting deletion commands, using only official tools, and preserving audit logs.

Simply instructing AI not to register entries incorrectly is not enough.

The execution environment itself must stop risky actions.

For example, authority for journal entry creation and updates should be set to Ask, deletion-related operations should be set to Deny, and before registration, the client name, accounting period, number of entries, and total amount should be displayed for confirmation.

Implementation will proceed in stages.

At first, operations will be Read only, limited to data retrieval and analysis.

Next, the process will move to Draft, where AI creates drafts and correction proposals.

After that, JGA will move to Approve & Write, where writing is permitted only after human approval.

Finally, the system will be rolled out more broadly based on client-specific rules and audit logs.

Not aiming for automatic registration from the beginning ultimately leads to safer and faster implementation.

Starting with review as the default and forcibly stopping dangerous operations through Hooks is the basic premise for practical implementation.

8. Implementation Roadmap and JGA’s Execution Policy

JGA will introduce AI Agent Accounting in stages.

First, we will create an environment where it can be tested safely and operate in Read only mode for test clients.

After that, we will standardize draft journal entries, lists of correction items, and client confirmation items, and then test approval-based registration starting with low-risk transactions.

Implementation requires several prerequisites.

Accounting data must be organized to a certain extent.

Internal rules for accounts, subaccounts, departments, and tax categories must be documented.

Registration-related operations must be subject to approval.

Client-specific folders and supporting document storage locations must be clearly defined.

Particularly important rules include one session per entity, double confirmation before registration, excluding high-risk transactions from automatic processing, documenting client confirmation items, turning internal rules into Skills, and reflecting exception handling in MEMORY.md.

At first glance, these may appear to be small operational rules.

However, when using AI in actual practice, these small rules are exactly what determine quality.

JGA’s goal for AI Agent Accounting is not to rely on individual staff members’ manual checking skills.

It is to create a system that can consistently produce the same level of quality.

Conclusion: What Accounting Firms Need in the AI Era

What accounting firms need in the AI era is not a superficial statement that they are “using AI.”

What is needed is the ability to design which parts of the workflow should be entrusted to AI, which parts should be approved by humans, how evidence should be preserved, and what kind of value should be returned to clients.

JGA will shift from an accounting firm centered on data entry to an accounting firm that uses AI to streamline drafting, checking, and document preparation, while humans focus on judgment, explanation, proposals, and business improvement.

To achieve this, we will use Claude Code, MF Cloud Accounting, and MCP Server, while developing CLAUDE.md, Skills, MEMORY.md, settings.json, and Hooks, and implementing controlled AI Agent Accounting step by step.

In the tax accounting industry going forward, it will become difficult to differentiate based only on speed.

What will matter is the ability to use AI safely, understand each client’s specific circumstances, and convert accounting and tax information into meaningful explanations and proposals for business owners.

JGA aims to become an organization that not only entrusts work to AI, but also designs a system in which AI can work safely, balancing both the quality and speed of tax accounting services.

Final Note: Separating Where GitHub Should and Should Not Be Used

When rolling out AI Agent Accounting across the firm, the use of GitHub will also be important.

However, the most important point is that client information and personal information must not be placed in GitHub.

What should be stored in GitHub is not client data.

What should be stored there are frameworks that make work reproducible, such as CLAUDE.md, Skills, checklists, prompt templates, output formats, and sample Hooks settings.

In other words, GitHub should be positioned not as a place to collect client data, but as a place to manage versions of business templates and workflows.

Client names, contact persons, email addresses, supporting documents, trial balances, journal entry data, consultation history, and similar information should never be placed in GitHub.

Instead, they should be separated and stored in the places where they properly belong, such as MF Cloud Accounting, supporting document management folders, and client-specific management areas.

This approach is also important from a security perspective.

In 2026, Money Forward announced the completion of a detailed investigation and additional security measures regarding unauthorized access to the GitHub environment used by the company.

According to the company’s announcement, no unauthorized access to the production database or leakage of information from the production environment was confirmed.

At the same time, the announcement explained that certain GitHub repositories contained personal data related to customers, such as names and email addresses.

The lesson to draw from this incident is not that GitHub itself should be avoided.

Rather, if GitHub is used, it is essential to decide from the outset what should be placed there and what must never be placed there.

In principle, GitHub should not contain any client information or personal information.

It should contain only frameworks such as business rules, templates, checking procedures, and AI execution controls.

In that sense, the fact that customer information was included in GitHub repositories raises operational concerns from an external perspective as well.

At JGA, we will use GitHub as a foundation for standardizing business operations.

However, what we manage there will not be client data.

It will be the design documents required to operate our workflows safely.

By setting clear premises — no client information, no personal information commits, no environment variables or authentication credentials, reviewed change histories, and approval processes for risky changes — we will build an environment that allows AI Agent Accounting to be rolled out safely across the firm.

Reference: Money Forward, “Completion of Detailed Investigation and Strengthening of Security Measures Regarding Unauthorized Access to GitHub (Fourth Report)”

This article has been reconstructed for external publication based on JGA Tax Corporation’s internal overview of AI Agent Accounting.