The Fundamental Prerequisite for Business Improvement Before AI Adoption: Resolving 'Tacit Knowledge' and Reconstructing Processes

Introduction: Why Business Improvement Fails and the Barrier of 'Visualization'

In many organizations, the terms 'business improvement' and 'productivity enhancement' are used daily. However, while they are touted as slogans, there are frequent cases where the reality on the ground fails to yield the expected results. There is no shortage of examples where top-down directives are issued, only to leave the staff exhausted.

A particularly notable situation recently is the confusion surrounding the use of rapidly evolving artificial intelligence (AI) technology. Although policies are set stating that 'AI is revolutionary and should be incorporated into operations to achieve improvements,' there is often a lack of tactics at the operational level regarding 'what specifically should be entrusted to AI and how'. As a result, staff members are unable to take effective action, facing the challenge that the consideration of adoption itself stalls.

Why does this dysfunction occur? One of the biggest factors in the failure of business improvement lies in the fact that the 'visualization and verbalization' of business processes has not been achieved.

Within organizations, there are many instances of tasks that rely on the experience and intuition of specific skilled individuals, so-called 'tacit knowledge'. In a state where the overall picture of operations and standard procedures are not objectively grasped, it is impossible to identify bottlenecks or devise improvement measures.

This problem becomes fatal, especially when utilizing currently mainstream generative AI and Large Language Models (LLMs). These AI technologies have a mechanism that outputs probabilistically appropriate answers to input text information (prompts).

In other words, if the instructions are not verbalized, the AI will not function. As long as it remains 'tacit knowledge,' it is impossible to give appropriate instructions to the AI, and the resulting output will be meaningless.

Therefore, the first step in business improvement is to convert the 'tacit knowledge' dormant within the organization into 'explicit knowledge' that anyone can understand and execute; in other words, it lies in the thorough 'visualization' of business processes.

Chapter 1: Structural Understanding of 'Trunk' and 'Branches'

When proceeding with the 'visualization' of business, the first thing many workplaces undertake is the creation of 'business manuals.' However, despite preparing detailed manuals, phenomena frequently occur where this does not lead to essential productivity improvement, or conversely, operations become rigid.

This stems from confusing the structural layers (hierarchies) of the 'trunk' and 'branches' in business.

What is the 'trunk' of a business process?

The 'trunk' of a business refers to the 'process itself,' such as the overall purpose of the business, the major flow from start to finish, coordination between departments and personnel (handoffs), and the input/output of information. This is the skeleton of the business and the blueprint showing what value the organization is creating.

What are the 'branches' of a business process?

On the other hand, the 'branches' of a business refer to the specific operational procedures and know-how in individual work steps that make up the 'trunk' process. Generally, what is documented as a 'business manual' corresponds to these 'branches.'

The Pitfall of Manual Preparation

The essence of the problem lies in focusing on the preparation of manuals, which are the peripheral 'branches,' while the business process itself, the 'trunk,' contains inefficiencies and waste.

For example, suppose there is a task of entering certain data into a core system. When creating a manual for this task, everything from how to log in to the system to the operation of the input screen and confirmation procedures is described in detail. However, when looking at the entire business (the trunk), it is not uncommon for that data to have already been digitized by another department in a previous process, and the input work itself would be unnecessary if only coordination were performed.

In this case, no matter how precise an input manual is created, it is merely a procedure manual for efficiently performing unnecessary work. Moreover, by being manualized, the 'existence of that task' is justified, creating the risk that inefficient processes will become fixed. In business improvement, it is essential to first review the 'trunk' and then optimize the 'branches' that are subsequently needed.

Chapter 2: Speak Instead of Write — The 'Monologue' Technique for Generating Business Flows

Even if we say 'we should visualize the core of our business,' it is difficult to find the time in a busy workplace to create business flowcharts or procedure manuals from scratch. Furthermore, the judgments and tips (tacit knowledge) that experts perform unconsciously are often left out when one tries to sit down and write them out.

Therefore, the recommended approach is not to write text, but to speak your thoughts while performing the actual work and have AI analyze them as an approach.

The 'Monologue' Recording Process

The specific steps are as follows.

1. Start Recording: Turn on the recording function on your smartphone or PC.

2. Practice Live Commentary (Monologue): While actually performing the work, voice all your actions as a live commentary.

   • First, I open Company A's invoice file.

   • Next, I visually check the date and total amount. Sometimes the consumption tax fraction doesn't match, so I check it with a calculator.

   • Once confirmed, I open screen B of the accounting software and enter the date. It is important to speak not only about the operations but also about the thought process, such as 'what I am checking' and 'why I am doing it.'

3. Transcription and Structuring by AI: Feed the recorded data into an AI (such as a generative AI with speech recognition capabilities), have it transcribed, and then provide instructions as follows.

   • Based on the audio just provided, please create a business flowchart.

   • Please create a draft of the procedure manual for this task.

   • Please list the points the operator is careful about and the branching conditions.

Uncovering 'Tacit Knowledge' Using AI

The greatest advantage of this method is that unconscious actions are verbalized. Subtle judgment criteria that are usually omitted in manual creation, such as 'why I paused here' or 'what I saw that made me judge it as an error,' are recorded by performing the live commentary.

AI excels at extracting and organizing logical structures from rambling 'monologues.' Even without humans spending time creating documents, utilizing AI makes it possible to visualize the reality of business (the core and the details) in a short time and with high resolution.

Chapter 3: The ECRS Principle and the Importance of the 'Decision to Discard'

Once the overall picture of the business (the core) is visualized, the next step is to optimize the process. Here, we will explain the improvement process based on the 'ECRS Principle,' which is widely used in the fields of industrial engineering and production management. In particular, we will delve deeply into the traps that many DX and IT initiatives fall into, as well as the role of humans.

The Four Perspectives of ECRS and the Correct Order

It is an absolute rule to consider ECRS in the following order. This is because they are arranged in order of lowest cost and highest effectiveness.

• E: Eliminate

  • Can you eliminate the task itself? (Most important)

• C: Combine

  • Can you combine tasks?

• R: Rearrange

  • Can you change the order or the person responsible?

• S: Simplify

  • Can you simplify it or automate it by introducing tools?

The Typical Failure of IT/DX: Escaping to 'S' without 'E'

A common failure pattern seen in many places is skipping the consideration of 'E (Eliminate)' and 'C (Combine)' in this ECRS framework and jumping straight to 'R (Rearrange)' or 'S (Simplify/Systematize)' patterns.

The decision to 'eliminate a task' requires organizational coordination and risk assessment, making it psychologically and politically challenging. In contrast, introducing a new system or AI tool (S) provides a sense of accomplishment that 'something new is being done' and is often easier to get approved.

However, introducing an expensive system while leaving unnecessary business processes as they are is nothing more than 'digitizing waste and running it at high speed.' As a result, tasks become more complex than before due to system configuration and maintenance, leading to a counterproductive situation—so-called 'DX defeat'—where you end up paying higher system usage and maintenance fees than the labor costs you managed to save.

The Human Responsibility to Judge 'Why It Can Be Eliminated'

So, why is 'E (Eliminate)' not possible in many workplaces? It is because there are fewer people who understand the 'essential meaning of the task' and can take responsibility for deciding to 'stop' it.

Meaningless check tasks and document creation continue to be preserved for reasons like 'it was handed down from my predecessor' or 'I heard there was trouble in the past.' 'Eliminating a task' carries the risk of something going wrong. To weigh those risks and costs and assert that 'it is unnecessary in the current environment,' a deep understanding of not just the superficial procedures, but the underlying laws, regulations, and the business mechanism itself (the essence) is indispensable.

The 'Final Responsibility' That AI Cannot Take

Here, the role and limitations of AI become clear. AI can look at a visualized flow and propose that 'this process is redundant' or 'it can be eliminated from the perspective of ECRS.' However, it cannot bear responsibility.

AI provides optimal solutions based on 'past data' or 'general theory,' but it cannot take responsibility or 'fall on its sword' for any legal risks or damage to customer trust that might occur in the future as a result of eliminating that task.

No matter how much AI evolves and becomes capable of replacing all knowledge work, the work of 'understanding the essence, accepting risks, and making final decisions' can only be done by humans. Precisely because AI is being utilized for business improvement, humans need to further hone their essential judgment ability to question the 'Why' of business operations.

Chapter 4: Criteria for AI Substitutability

Through the ECRS process, tasks that should be considered for AI substitution (Simplify/Rearrange) are extracted. So, what kind of tasks are specifically suitable for AI? The areas where current AI technology, especially generative AI and recognition technology, excels can be broadly classified into the following three categories.
*While media generation is also a strong field, it is omitted here as these creative elements are rarely important in common knowledge work.

1. Recognition

A function that reads unstructured data such as images, audio, and handwritten characters and recognizes/converts them into digital data.

• OCR (Optical Character Recognition): Extracting text information from paper documents or PDFs.

• Speech Recognition: Converting meeting recordings, etc., into text.

• Image Recognition: Identifying objects or situations within images.

2. Generation & Summarization

A function that creates new text or summarizes content based on large amounts of text data. This is the area where LLMs excel the most.
Recently popular slide generation also falls into this category.

• Drafting: Creating drafts for emails, reports, daily logs, etc.

• Summarization: Extracting key points from long documents or meeting minutes.

• Translation/Proofreading: Performing translations between languages and checking for typos, omissions, or writing styles.

3. Anomaly Detection

A function that finds patterns that differ from the norm or specific regularities from within large amounts of data.

• Fraud Detection: Detecting data inconsistencies or unusual transactions.

• Classification: Categorizing data based on predetermined rules.

Chapter 5: Concrete Use Cases in Specialized Operations

Here, based on the AI strength areas (recognition, text generation, anomaly detection) organized in the previous chapter, we will explain specific utilization scenarios envisioned for specialized tasks such as accounting and tax, as well as administrative departments.

1. Efficiency in input tasks through 'recognition' technology

The processing of supporting documents accounts for a large portion of many administrative tasks. We will replace this with 'recognition' technology.

• Automatic reading of receipts and invoices:
  AI-equipped OCR reads the contents of receipts and invoices (dates, amounts, business partners) that were previously visually checked and manually entered by humans. This 'simplifies' (S) the input task itself, allowing humans to focus solely on verifying the results read by the AI.

• Digitization of bankbooks and handwritten memos:
  With improvements in image recognition technology, it has become possible to convert handwritten cashbooks and copies of bankbooks into text data with high accuracy.

2. Efficiency in communication and reporting through 'text generation'

Even in advisory tasks that require expert knowledge, AI can be utilized in the preliminary stages and peripheral tasks.

• Drafting emails for clients:
  For communication tasks that include routine elements, such as 'informing customers about the overview of the invoice system,' you can instruct the AI on the key points and have it create a draft of the email. The person in charge only needs to revise and confirm it, reducing the time spent thinking about the text from scratch.

• Generating comments for monthly reports:
  By inputting trial balance and management data into the AI in text format, it is possible to have it create a draft of report comments regarding 'factors for the year-on-year increase in sales' or 'fluctuations in expenses.' This allows for faster initial action in analysis tasks.

3. Advancement of audit and check tasks through 'anomaly detection'

Checking large volumes of data is an area where AI excels over humans, contributing to error reduction and quality improvement.

• Anomaly detection in journal entry data:
  By having the AI learn from vast amounts of past journal entry data, it can detect as alerts any newly entered journal entries where 'the combination of account titles differs from past patterns' or 'the number of digits in the amount deviates from the norm.'
  If vast input is not cost-effective, you can create a rulebook for processing and have it judge whether there are any abnormal values based on that.

Chapter 6: AI is not 'perfect' — Facing anxiety about accuracy

When considering the introduction of AI, we often hear the concern, 'I am anxious about entrusting tasks to AI because it is not 100% accurate.' This anxiety is a very correct and healthy feeling, as you are responsible for quality as an expert. AI, especially generative AI, sometimes confidently outputs content that differs from the facts.

However, there are two perspectives to stop and consider here.

1. Humans are not perfect either

AI is imperfect, but human work is also never 100%. Since human error always occurs, processes such as double-checks and supervisor approvals have always existed. The correct way to view this is not that 'checks become unnecessary because it is AI,' but that 'AI deliverables also require checks, just like human work.'

• Reduction of work personnel: Even if the same checking effort is required, the time to create deliverables from scratch is dramatically shortened by AI. Humans are freed from 'creation' and can focus on 'confirmation and judgment.'

• Uniformity of quality: While human work quality varies depending on physical condition and mood, AI can continue to output uniform quality that always follows a set rule by adjusting the prompt (instructions).

2. If you need the 'same result every time,' leave it to a program

If you require strict, standardized processing where the exact same result must be output 100% of the time and not even a millimeter of deviation is permitted, you should choose automation via rule-based 'programs (code)' rather than generative AI (LLM), which operates on probability.

You might think that 'writing programs is outside my expertise,' but nowadays you can have the AI write the program code itself. If you instruct the AI, 'Write a macro that compares columns A and B of this Excel data and turns them red if there is a mismatch,' the AI will generate the code. Of course, you still need to check (test) whether the code works correctly, but even here, the AI handles the labor of 'creation,' and humans can shift to the role of 'supervisor.'

For example, in the recently released WorkspaceFlows, there are two types of steps: steps where the AI makes decisions and steps that simply perform predetermined actions; you need to use these two appropriately depending on the situation.

Flows with a simple UI is also useful as a mere no-code workflow tool without AI, so there are times when you don't necessarily have to use AI.

Conclusion: Roadmap for Improvement

To succeed in business improvement using AI, you must start not with the introduction of technology, but with an inventory of your business operations. The roadmap explained in this article is restated below.

1. Visualization of business (turning tacit knowledge into explicit knowledge): Use 'talking to yourself' and AI voice recognition to visualize the reality of business operations (5W1H) without burden.

2. Process optimization via ECRS and human judgment: While having the AI analyze the visualized flow, ultimately, humans must take responsibility for making the decision to 'E (Eliminate).' Do not escape into easy system implementation.

3. Judgment of the right tool for the right job:

   • Tasks requiring flexibility: Utilize 'AI (recognition, generation, detection).'

   • Tasks requiring strictness: Utilize 'programs' (use AI for code creation).

4. Redefining the human role: Let AI and programs be the 'creators,' and humans focus on their work as 'responsible decision-makers and approvers.'

AI is not a magic wand, but if applied through appropriate processes and check systems, it becomes a powerful partner that can dramatically improve business efficiency. To break out of the state of 'not knowing what to do,' first start by giving a play-by-play of your current work through 'talking to yourself' and re-examining it objectively. Then, while borrowing the power of AI, exercise the 'ability to discern the essence' that only humans possess.