日本語A&M GROUP Official SiteProfessional AI Multi-AgentSound Engine DOMEarEffect

Support — Operation

We design a workflow in which
a person checks and finalizes drafts created by AI.

To use AI in operations, you need rules on who verifies what, and where, before a result is finalized. We design everything from how AI output is presented to the approval flow and record keeping.

Disclosure
Disclosure of AI-Drafted Content
Review
Review and approval flow
Evidence
Records of grounds and operations

When

In situations like these

Concerned about AI errors

You cannot tell whether AI output can be used as is. If AI output is clearly labeled as an "AI proposal" and it is decided who checks it, where, and how it is confirmed, errors can be stopped before confirmation.

  • We are unsure whether we can notice errors
  • Where responsibility lies is unclear

Requires auditing or explanation

You may later need to explain why a decision was made. If the system is designed to keep a record of the materials behind each decision and of who did what and when, you can answer such questions afterward.

  • Evidence must be shown
  • We want a record of who verified it

It is not used on site

A system exists, but how to use it has not been decided. Without rules and a guide for use, the system will not be used. We create rules that fit on-site procedures and see them through until they take hold.

  • Usage differs from person to person
  • The verification steps get skipped

What we do

What We Provide

Disclosure of AI-Drafted Content

We show separately the parts written by AI and the parts checked by people. The AI-written text can be told apart from the text people have checked, so that a document not yet finalized is never treated as final.

  • Display rules
  • Replacement history
  • Version control

Review and approval flow

Decide who confirms, where, and after checking what. Define the stages of review, return, and approval, and who is responsible for each. The greater the impact, the more stages of approval we add.

  • Review procedure
  • Approval stage
  • Handling of returns for revision

Records of grounds and operations

We record the grounds for each decision and who did what. We keep links to the supporting documents and a record of operations, so the course of a decision can be traced afterward.

  • How grounds are kept on record
  • Record of operations
  • Permission design

In depth

We decide in advance what is left to AI
and which steps a person checks and confirms.

Whether AI can keep being used in daily operations depends less on performance than on operating rules. Where, and by whom, is AI output checked, and what counts as final? Without these rules, the people on the front line cannot fully trust AI results, and in the end a person has to review everything again.

We design everything around AI output: how AI-written content is clearly labeled as an "AI draft," the review and approval flow, how rejections are handled, and the supporting materials and operation logs. We also build into the operation a feedback channel that improves the instructions given to the AI based on input from the people on the front line.

Our in-house systems run every day in a way that places human approval at steps requiring judgment, so that only approved actions are executed. High-impact operations, such as stopping a process or writing data, are designed not to run when the data needed for the decision cannot be read.

Case study

Case Study

We present the flow and the finished form in the form of reports. Each one states clearly whether it is based on our actual operations or is an example assumed from how such work is normally done.

AI drafts are labeled as “AI proposals,” and a professional reviews and finalizes them.
Example based on our own operating experience

AI operations design

AI drafts are labeled as "AI proposals,"
and professionals review and confirm them.

This is an example from our product development. When AI drafts professional documents, an organization cannot use them if it is unclear how much is the AI's writing and who verified and finalized it.

  1. Establish a rule to mark the AI's entries with an "AI proposal" label
  2. Define the flow of review, return for revision, and finalization
  3. Keep the supporting documents and a record of operations
  4. We set up an improvement desk that receives feedback from the field
  5. Reflect only approved corrections in the instructions to the AI

ReportReport for this case

Flow
Make this the operating rule: AI drafting → review → return for revision / confirmation → record → improvement contact point → improvement of instructions.
Deliverables
  • How AI proposals are presented, and the review and approval flow
  • Records of grounds and operations
  • A contact point for improvements, and management of the change history of instructions given to the AI
Points to confirm
  • Whether any document is being treated as final before it has been finalized
  • Whether it is possible to trace afterward who corrected what and when
  • Whether proposed corrections are applied after approval
Only what a person has approved is executed from AI's proposals.
Example based on our own operating experience

Operations driven by approval

Only the AI proposals
that a person has approved are executed.

This is an example from our own company. The multi-agent system that oversees operations across the whole business drafts improvement proposals, but if it were to execute operations with real impact without human approval, it could harm the business.

  1. Show the AI's proposals separately from the targets of execution
  2. On the approval screen, choose to approve or pass on each proposal
  3. Execute only what has been approved and record the results
  4. Before high-impact operations such as stopping or writing, check that the data used for the decision has no gaps

ReportReport for this case

Flow
Make this the operating rule: AI proposal → approval screen → execution → recording of results → next-day check.
Deliverables
  • The approval screen and a record for each proposal
  • Recording the results of execution
  • Rules for confirmation before operations such as stopping
Points to confirm
  • Whether any unapproved operation is running
  • Whether it is possible to trace who approved it and when
  • Is a decision such as stopping being made while data is missing?
Reports written by AI are checked and finalized by the person in charge and the person responsible.
Example based on typical ways of working

The AI drafting and review workflow

A report written by AI is
checked and finalized by the person in charge and the responsible manager.

In our research department, we have started having AI draft reports, but how far staff verify the content before releasing it differs from person to person.

  1. Mark the AI's entries with an "AI proposal" label
  2. Set up a flow in which the person in charge checks against the evidence and the responsible manager finalizes
  3. Record the reasons for sending work back
  4. Correct the instructions to the AI based on the reasons for sending work back

ReportReport for this case

Flow
Operate in this order: AI drafting → staff check → confirmation by the responsible person → record of returns for revision → improvement of instructions.
Deliverables
  • Rules for displaying AI proposals
  • Review flow and permissions for staff and managers
  • Record of returns for revision
Points to confirm
  • Whether any report has gone outside before being finalized
  • Is the reason for the return for revision recorded?
  • Are the same rejections decreasing?

Operation

Finalization is done by
the professional and the person in charge.

Our Professional AI is also operated so that the passages written by AI are clearly marked, and a person traces them back to the source, verifies them, and finalizes them.

Deliverables

What we deliver

  1. Operations design document

    Rules for display, review, approval, and records. We compile operating rules into a single volume, from how AI output is presented to review, approval, and recording.

  2. Design of permissions and records

    We decide who can do what and what is kept: who can view, correct, and finalize what, and the scope of the records that are retained.

  3. Field guide

    A guide that staff can use without hesitation. It covers screen operations and the criteria for making judgment calls, written in the words of the people who do the work.

Process

How to proceed

  1. Understanding the current situation

    We ask about your current review and approval flow. We ask about how checking and approval work today and who verifies what.

  2. Identifying risks

    We identify the impact when AI makes an error. We map how far the effects extend when the AI is wrong, and decide at which stage to stop it.

  3. Design

    We design the rules for display, review, approval, and records. We design them to follow the flow of the work.

  4. Trial

    We try it in actual work and revise the procedures. By trying it in actual work, we revise the procedures wherever confusion or rework arises.

  5. Establishment in daily operations

    We prepare guides and help the new way of working take root on site. We prepare the guide, hold briefing sessions, and set up a contact point for feedback from the people on site.

Estimate

What Determines the Cost

We prepare a quote after asking about the following points. With the AI quote, the AI asks about the same points in turn.

  1. Number of tasks

    The number of tasks for which operations are designed. The more tasks there are, the greater the volume of rules and guidance.

  2. Approval stage

    Number of review and approval stages and people involved. The more approval stages and people involved, the more design and coordination work is needed.

  3. Record Requirements

    The scope of the evidence and operation records we keep. The wider the scope of records kept, the more work goes into building the system.

  4. Guides and training

    Whether a guide is prepared and whether briefing sessions are held. The scope differs depending on whether you choose a guide only or include briefing sessions.

Why A&M GROUP

The engineer who receives your inquiry will go on to handle the development.

  1. There are no handovers between staff members.

    The engineer who did the interviewing designs and implements the system directly. Requirements do not drift through back-and-forth specification documents.

  2. We operate them every day in our own business

    We actually operate 8 AI systems that do the work of several thousand people in our own business, including the multi-agent system that oversees the operation of the entire business. Our advice comes from experience of running them, not from proposals on paper.

  3. We propose the suitable method, not only AI

    Where AI is not needed, we will tell you so. If an off-the-shelf tool is enough, we will propose that approach as well.

  4. We also handle operation and improvement after delivery.

    We run it in actual operation and take responsibility through to detecting faults and continuing to fix them. We operate our own 8 AI systems under this structure every day.

Operating scale — from system logs

AI that carries the work of thousands of people,
operated every day in our own business.

Every proposal is a method we have run and verified in-house. Our AI processes every day a volume of work that would require multiple departments if handled manually.

approx. 100,000cases/day

Checks bid notices published by government agencies and others 4 times every day, with AI judging deadlines, participation conditions, and fit with our company

1,108cases

Bidding projects whose deadlines and progress are managed by AI

4,084items

Orders, inventory, pricing, advertising, and finance on one platform

86functions

In-house system functions that can be run directly from AI assistants

8systems

AI systems we actually run in our own business

Values retrieved from each system's records (as of November 9, 2026). "Thousands of people" is a figure we calculated: we converted this workload into the effort it would take people to process it with the same tasks, and took the number of people needed to finish it within twenty-four hours. It is not a measurement by a third party.

View the workload breakdown

Contact

Please contact us about AI operation design.

Choose a contact point and describe your request. Together with our AI guide, we will organize the details and prepare a rough estimate on the spot.

Request a quote for this plan