Business automation using AI | How to separate processes to be delegated and processes to be checked by humans
Business automation using AI is a method of using AI in processes that require certain judgments, such as reading and classifying texts, searching for information and creating drafts. Rather than entrusting the entire task to the AI from the beginning, we will separate the ideas created by the AI and the operations to be finalized and executed by humans. Here, we will use the example of how to sort inquiries to organize what needs to be checked in order to put it into practice.
What is the difference between AI automation and normal business automation?
If the inputs and conditions are clear, they can be processed using spreadsheet functions or existing system rules. On the other hand, the process of reading and categorizing the meaning of inquiries, or compiling proposed answers from multiple materials, is a candidate for testing AI. Not leaving processes that do not require AI to AI clarifies the scope of verification. If you want to start by designing routine processes, please see How to start business automation.
| Process | How to try it first | What people should check |
|---|---|---|
| Records of reception times and customer IDs | Existing system settings and rules | Input omissions and duplicates |
| Classification of inquiry text | Creating a classification proposal using AI | Validity and oversight of classification |
| Create a reply | Create a draft by referring to approved materials | Rationale, terms and conditions, personal information |
| Sending to customer | Initially carried out by the person in charge | Addressee and final message |
If you plan to design everything from email reception to reply proposal as one task, you can check the rule judgment, hold, and confirmation of the person in charge in Email sorting and reply draft procedure.
How to try classifying inquiries?
First, check the environments your company is authorized to use and the inquiry data available. If you don't need to use actual customer information, start with hypothetical text. We will define categories such as Price,'' Obstacle,'' and Other,'' and create a rule that if the item cannot be categorized, it will be Waiting for confirmation.''
Next, a person selects examples from past inquiries that have permission to use, and records the correct classification and rationale. In addition to ordinary text, include text that includes two topics, text that lacks information, and text that is highly urgent. Compare the AI's proposed classification with the human judgment and record what type of mix-up occurred. You cannot judge the quality in production just by matching a few examples.
| Items to record | Entry example (fictional) |
|---|---|
| Input | "I contacted you to confirm the billing amount and to be unable to log in" |
| Classification added by someone | Waiting for confirmation because there are multiple topics |
| AI classification proposal | Disability |
| Confirmation result | I overlooked the billing amount question |
| Next action | The person in charge will handle the two topics separately |
This table does not indicate the percentage of correct answers for classification. In our own evaluation, we also record the number of cases, errors in each classification, and the correction time of the person in charge.
Download the sheet to check AI classification results (CSV) and compare AI proposals and human judgments one by one. Inquiries about sample entries are fictitious.
Can I leave the execution and external sending to you?
Drafting and sending externally are separate processes. In the first attempt, the AI will create a classification proposal and even a draft, and a representative will open and review the original inquiry and reference materials. If you want to include sending or updating customer information, design the operation authority, approver, execution log, and stop method before adding it to the target. There are confirmation items in AI agent approval and authority management.
IPA's Document on the safe use of generative AI and AI agents also cites unauthorized use and inaccurate output as challenges. Please decide what information can be entered and who will be in charge of checking it, based on your company's rules and usage environment.
Document on the safe use of generative AI and AI agentshttps://www.ipa.go.jp/jinzai/ics/core_human_resource/final_project/2026/ai-security.html
Should I try it in training or consult about installing Grokbot?
If you have not decided on the target task yet, you can choose the process to be delegated from AI training for presidents, where the president himself tries out AI using actual work as a subject. If you have decided on the target work and the materials to be used, you can organize the collaboration partners and people's approval using the Confirmation table when proceeding from training to Grokbot introduction and Grok Bot company introduction procedure. We also decide whether it is necessary to use or build Grokbot depending on the business.
When consulting, please let us know the task you would like to delegate, the type of materials you will be using, whether you will send them externally or update them, and who will approve them. Items that have not yet been decided may remain undecided.