The value SCAN identifies is only realized through execution. ADOPT is the framework that turns a chosen AI initiative into a specific, feasible, and measurable execution plan.
In the previous article, we used SCAN to weigh the potential value of AI initiatives along two axes: their capturable value and their strategic fit. What comes next is execution: putting the chosen initiatives into practice. And here the hard part is that success or failure is decided not by any single element, but by how all of them combine.
For example, even with an excellent AI model, it won’t run without the data to feed it. Even with a clearly defined target outcome, nothing changes if front line staff don’t actually use the system. And even with a sound plan, the initiative collapses if the transition path is infeasible. In short, value leaks away at the seams between the elements.
This article therefore proposes an execution-planning framework called ADOPT. ADOPT builds an execution plan through five interlinked elements: (1) the AI that acts (Agent), (2) the data that powers it (Data), (3) the outcome it is meant to deliver (Outcome), (4) the people and organization that make it move (People), and (5) the shift from existing operations that carries it into daily work (Transition). It turns a chosen AI initiative into a single execution plan in which the five fit together.
Overview of ADOPT
- Agent: What does the AI take on, and how does it collaborate with people? Define its role, autonomy level, and guardrails, and design the AI solution.
- Data: What data is needed, and is it in a usable state? Clarify data quality, obtainability, and governance, and set up new data mechanisms such as feedback loops.
- Outcome: What counts as success, and how is it measured? Carry over the capturable value SCAN estimated and make it the target; then set the KPIs, baseline, and measurement method.
- People: Whose work changes inside the company, and what makes them able and willing? Assess capability gaps, motivation, and adoption, and design the change management.
- Transition: How do you move from existing operations to a new way of working? What are the gaps you need to cross? Plan the pilot, rollback, sequence, and timeline.
Two principles are worth holding onto here. First, it is coherence, not the parts, that makes a plan work. Second, the Outcome is the center of this execution plan, defined by working backward from the capturable value SCAN estimated. Agent and Data are the technology wing; People and Transition are the human wing — it is the balance and synergy of these two wings that makes AI’s value real.
1. Agent (the AI actor) — Designing the Solution
First, what does the AI take on? A simple stand-in for a person, a co-pilot that supports human judgment, or something that acts autonomously? Decide this autonomy level, then design the points where humans stay in the loop, the guardrails that prevent deviation, and the escalation for when things fail. What you produce here is the AI solution design.
The risk points to check lie in technical feasibility, model maturity, and guardrail reliability.
Be careful not to treat agent as a foregone conclusion. The right answer may well be a conventional machine-learning model, a fixed workflow, or a data-processing pipeline. Choose the simplest means that reaches the outcome.
One decision deserves early attention: where the Agent’s “brain” — the LLM (Large Language Model) — lives. Do you use a cloud LLM via API (Cloud AI), or run an LLM on your own GPUs (Local LLM)? This choice greatly changes data sovereignty, confidentiality, response speed, cost structure, and operational burden. For general-purpose processing where you want to launch the latest capability quickly and cheaply, choose Cloud AI; where sensitive data cannot leave the premises, regulation is strict, or heavy processing runs continuously, choose Local LLM. Weigh this not on functionality alone, but comprehensively across performance, security, and cost. The details are collected in the appendix.
Ichimaru once ran a digital-transformation project. At that time it built a demand-forecasting model and rolled it out internally. But its business impact was limited, for two reasons.
1. The model it built back then did not factor in weather forecasts or nearby-event information, so under those conditions its predictions were often badly off, making it hard to run as-is.
2. The model output no reasoning behind its forecasts, so when area managers doubted a number, there was no way to justify it, and they never came to trust the model.
This time, it wants to design the AI solution around these two problems. Specifically, it builds AI-driven analysis and support functions: calibrating the forecast for weather and event information, and advising on error factors in past-accuracy assessment, on how to use the forecast, and on the ordering work itself.
And there was one more debate — over the LLM choice, Cloud AI versus Local LLM. Here the president said:
“Lately the number of inbound tourists, especially visitors from Europe, has been rising. At some stores they account for nearly half of sales. We are also considering introducing a loyalty AI down the road, so for anything touching personal data I would like to go with a Local LLM. This project does not handle personal data, but with the future in mind, I want to consider a Local LLM.”
Note: The president’s concern about European customers’ personal data for future loyalty AI usage stems from the EU’s GDPR (General Data Protection Regulation). To process the personal data of EU residents on a Cloud AI, one must sign legal documents with the Cloud AI provider — a Data Processing Agreement (DPA), Standard Contractual Clauses (SCC), and the like — to ensure the lawfulness of cross-border transfer. This is often a heavy, impractical burden for an ordinary Japanese company. A Local LLM, which keeps data in-house, avoids the problem.
2. Data — Fuel for the AI Engine
Next comes the data that feeds the solution. What data is needed? Where do you get it? Do you have access rights? Can you keep its quality? Can you satisfy governance and privacy? And how do you collect the feedback the AI itself generates?
The risk points to check lie in the data’s volume, quality, availability, and governance.
Be aware that data is often the bottleneck: the outcome you can reach is decided not by the model you adopt but by the data you can actually obtain. Moreover, the more feedback data accumulates, the more value your data generates. Put another way: if you do not design feedback collection, each cycle starts again from zero.
For Ichimaru, two datasets are clearly needed from the start.
One is weather-forecast data. Today, area managers look at the model’s demand forecast side by side with the next day’s and next week’s weather forecast, and adjust the number based on their own experience and intuition. This is work an AI can take over. Using public information from the relevant government agencies has an added benefit: it avoids extra data-licensing costs.
Note that the idea is not to change the running core of the current model, but to add a correction to its primary forecast purely as post-processing. This captures the effect of factoring in the weather while keeping the system-modification effort to a minimum.
The other is nearby-event information. Here two things are a little tricky:
1. What the area managers currently rely on is nearby-event information gathered from the web. This information is mostly unstructured, and needs data-shaping before an algorithm can process it.
2. The scale of an event — above all the number of attendees — is often not stated, so the area manager judges the expected turnout based on experience and intuition, and adjusts expected demand accordingly.
On these two points Ichimaru had an AI vendor run a quick desk-check trial using a Local LLM. On data-shaping, the LLM handled it well and produced clean structured data. On the other hand, while the LLM could predict event turnout reasonably well when given hints, a gap from the area managers’ judgment remained.
Here, one dataset worth collecting comes into view: the area managers’ corrections to the LLM’s adjustments. Continuously accumulate the input conditions, the LLM’s primary correction, and the area manager’s final revised result, and feed that back into the AI. In this way the AI’s correction skill gradually improves. This points to a broader benefit: AI gives you a way to turn implicit in-house know-how into explicit data, and then reuse it.
Note: Structured data refers to data created, processed, and stored according to a clear, strict format such as rows, columns, and attributes (e.g., relational-database tables, CSV, spreadsheets). By contrast, things without a fixed format — like web articles or email bodies — are called unstructured data. The latter is hard for algorithms to handle as-is and needs shaping.
3. Outcome — Success Targets and the Yardstick
What counts as success for an AI initiative, and how do you measure its effect? Carry over the capturable value SCAN estimated and make it the target, then set the KPIs and KGI (Key Goal Indicator), the baseline, and the measurement mechanism. Measure it as net value, after subtracting build and run costs, and separate leading indicators from lagging ones.
The risk points to check lie in business validity, market readiness, and timing.
Be careful to measure realized value, not activity. Time saved is not cost captured (SCAN’s “0.3-of-a-person” problem). The outcome of an AI initiative is the capturable value SCAN estimated, made concrete — and it is the center point of ADOPT. Every other element is planned to serve it.
Ichimaru has two success targets. One is to lower the average waste-cost rate from the current 5% to 3%. The other is to cut the area managers’ ordering-plan work time by 50%.
First, let us look at Ichimaru’s cost structure. That reveals what these two targets mean for the business.
Ichimaru sells roughly 5 million bowls of ramen a year. After deducting coupons and perks, the average unit price is ¥1,200. Ingredients are 30% of sales, labor 30%, and gross margin 40% — a somewhat high margin for the industry.
But here, the average waste cost is 5% of ingredient cost, mostly unsold stock caused by ordering inaccuracy. In money terms, that is about ¥90 million a year. However, the real issue is the structure behind that 5%.
Break it down, and stores open less than three years (about 20% of the total) consistently account for 80% of the waste cost. The area managers spend the most time on the ordering plans for these very stores, and this is still the result. In other words, as things stand, just holding today’s waste rate stretches the area managers to their limit; and as the number of stores grows, the waste rate is likely to worsen, threatening Ichimaru’s performance.
So, to spell out the two targets and their strategic meaning in more detail:
Target 1: Waste rate falls from 5% to 3% — at the current scale, a cost reduction of about ¥36 million a year.
Strategic meaning: Because person-dependent ordering plans become semi- or fully automated, the waste rate does not rise even as the number of stores increases — removing a major operational-risk factor in the expansion strategy.
Target 2: The area managers cut their ordering-plan work time — currently 10 hours a week — by 50%, saving, across four people, about 960 hours a year in total.
Strategic meaning: The time freed across the area managers amounts to about 0.5 of a person-year; redirected mainly to store-opening prep and negotiations, it converts what was to be an outside consultant into in-house capability — curbing outsourcing costs and preventing business know-how from leaking out.
Their strategic meaning becomes clear.
4. People — Stakeholders and Role Design
Even when AI creates a surplus, in the end it is people who harvest the value. Unless the front line staff actually change their behavior, the outcome is never realized. So in the People dimension, you identify whose work changes, and how, and design how to make those people able and willing to act.
First, define three roles clearly: Governance, which watches progress, quality, and risk; Drive & Execute, which actually does the work; and Support, which backs up the operation. Don’t leave who does what ambiguous. When roles hang in the air, the AI moves but the organization does not.
On top of that, analyze the capability gap between the current and the desired state. What abilities does the new role require, and how much is missing today? Define the capability-building actions to close that gap (training, OJT (on-the-job training), hiring, redeployment) and weave them into the plan.
The risk points to check lie in the capability gap, motivation (incentives), and operational adoption.
For Ichimaru, the role design looks like this:
• Governance is the president, who watches progress against key metrics and prompts action when things fall behind.
• Drive & Execute falls to the area managers, who master the AI’s corrections and make the final ordering decisions.
• Support is the store managers and staff, who enter detailed information on nearby events and uphold the quality of the data feedback.
The key capability to build is the area managers’ mastery of the AI tool. What the AI can do, in what situations it tends to be off, and on what basis to then make the ordering decision — through a process of testing the AI and conversing with it, area managers come to understand these fine points deeply. Through this they will fundamentally change the ordering process they have used until now and improve its accuracy and efficiency.
And the president’s governance, above all, is critical. Rather than waiting for the lagging indicator — the waste-cost rate — the president regularly checks leading indicators (the gap between demand forecast and actuals, correction status, the area managers’ AI-adoption rate, the event-information input rate, and so on), sends messages personally, and prompts action. It is precisely this governance from the top that becomes the greatest driving force to make behavior change happen — an indispensable element in adopting AI.
5. Transition — The Goal and Path of Change
Transition is the dimension that binds the previous four into a single change. It gives each of Agent, Data, Outcome, and People the meaning of “what changes, and how, from today to tomorrow,” and draws the path from today’s state to the target state as one continuous journey. If the Outcome is the destination, Transition is the journey that carries the four dimensions to it.
The risk points to check lie in the size of the gap to cross and organizational rigidity.
What to ask here, at its core, is competitive advantage. What is today’s competitive advantage? How do you maintain and strengthen it tomorrow? And can this initiative build a new competitive advantage? For these three questions, provide clear answers. Under the strategic fit SCAN identified, chart the path to the goals through the AI implementation plan.
Draw that change concretely, layer by layer. How do the tools and data assets change? How does the workflow change? How do people’s capabilities change? How does the organization change? And ultimately, how does the business change? Line up today’s state (As-Is) and tomorrow’s state (To-Be) layer by layer, and the full picture of change — and the distance you must close — comes into view.
Ichimaru’s advantage today is its distinctive taste — house-made noodles and broth — and its brand. How to strengthen that tomorrow: here lies the goal of the transition.
First, while protecting the taste and the brand, use AI to raise ordering accuracy and eliminate the waste weakness, and redirect the freed time to store openings, accelerating store expansion. Second, gain a strength that compounds as stores multiply: the accumulating feedback data and the know-how of AI correction. In this way, using AI reliably supports Ichimaru’s strategy of store expansion and cash-flow improvement.
Now, drawing the change by layer:
• Tools & Data: a post-processing correction AI, structured event data, and correction feedback as a new data asset.
• Workflow: the ordering plan shifts from adjustment by intuition to AI correction + area-manager confirmation.
• Capability: the area managers gain the skills to use AI and to develop new stores.
• Organization: from dependence on outside consultants to in-house capability, under the president’s leading-KPI governance.
• Business: the waste rate stays stable regardless of store count, lowering the operational risk of the expansion strategy.
The rollout approach also needs design. Weighing risk and feasibility, first pilot in a single area, and once the forecast-accuracy threshold is met, expand region-wide. One system-level trick is a design that leaves the running core of the current demand-forecast model untouched and adds a correction as post-processing based on additional data. It is a reversible step — you can revert to manual planning at any time. Such a 2-Way (reversible) design is often an effective tactic in adopting AI.
In Closing — A Good Plan Is Coherent
ADOPT turns a chosen AI initiative into a single execution plan in which five dimensions mesh together. Agent and Data underpin “can we build it?”; People and Transition carry “will it take root?”; and at the center always sits the Outcome, which inherits the capturable value SCAN estimated.
But nailing the five individually is not enough. Value leaks not within an element but at the seams between elements — even a superb AI will not run without the data to support it; even a clear outcome will not be realized if the front line does not move. So the last thing to ask is simply: do the five mesh with one another?
Whether the plan is worth executing can be checked at four checkpoints. Is it solid (does it withstand the main risks)? Is it specific (do the owners, data, and dates actually exist)? Is it feasible (are the cost and technology within reach)? Is it measurable (can you verify the realization of the capturable value SCAN estimated)? Only when it satisfies these four — and the coherence of the seams — does the plan begin to move.
The value SCAN identifies is only realized once the plan is fully executed. What makes that final step certain is ADOPT.
Appendix: Cloud AI vs. Local LLM
Definitions
- Cloud AI: Using a large cloud LLM (ChatGPT / Claude / Gemini, etc.) via API.
- Local LLM: Running an open model (Llama / Gemma / GPT-oss, etc.) yourself on GPUs in your own on-prem / private environment.
1) Functionality
For basic LLM functions — conversation, summarizing and writing text, reasoning, coding, translation, speech-to-text, and so on — there is structurally almost no difference. You rarely need to choose Cloud vs. Local by the kind of function (“what it can do”); the differences show up in the performance, security, and cost facets below.
2) Performance
| Aspect | Cloud AI | Local LLM |
|---|---|---|
| Accuracy | Can use frontier-class, largest models; highest. Able to use the newest models | Open-model-centric, a step behind, but the gap is narrowing. Purpose-specific fine-tuning can even reverse it on particular tasks |
| Response speed | Network-dependent. Subject to rate limits, congestion, and external outages | Stable and low-latency within your own network. Can run offline |
| Scalability | Effectively unlimited. Responds instantly to sudden load swings | Capped by GPU capacity. Scaling up needs possible architecture change and procurement lead time |
| Availability & fault tolerance | Depends on the AI vendor’s SLA. Risk of spec changes, discontinuation, and price hikes (lock-in) | You can design availability under your own control, but failure response is also your responsibility |
| Operational burden | Low. Infrastructure and maintenance are on the AI vendor’s side. Can be run by a small team | Medium to high. Requires an AI partner, or in-house talent for GPU operations, data processing, and MLOps (the operational framework for machine learning) |
3) Security & Data Compliance
| Aspect | Cloud AI | Local LLM |
|---|---|---|
| Data sovereignty & confidentiality | Data leaves your premises. Manageable via contracts and region selection, but cross-border and third-party-processing risks remain | Data stays within the organization, under your control. No external transmission, minimizing the exposure surface |
| Regulation & compliance | Depends on the vendor’s compliance scope. Needs conformity checks against industry and regional regulations (personal data, finance, medical, etc.) | Can be governed by your own policy; strong for confidential and regulated industries (medical, finance, government) |
| Audit & traceability | The visible range of logs and processing history is constrained by vendor specs | You can fully grasp and audit inputs, outputs, and processing paths in-house |
4) Cost
| Aspect | Cloud AI | Local LLM |
|---|---|---|
| Initial cost | Very low. You can start immediately; suited to PoC | Requires GPU procurement and environment setup — on the order of hundreds of thousands to a few million yen |
| Running cost | Pay-as-you-go. Cheaper at low-to-mid volume, but cost spikes with heavy, continuous processing | Mostly fixed cost. At high, steady volume the per-token cost flips to cheaper (favorable on Total Cost of Ownership) |
Summary: Cloud AI vs. Local LLM
- Function is not a differentiator. The axes of choice are three: “performance (accuracy, speed, availability),” “security (data sovereignty, regulatory fit),” and “cost (total cost of ownership relative to volume).”
- Cloud AI fits when: you want a fast launch, demand swings widely, you need the latest accuracy, in-house resources are thin, and heavy continuous processing is not expected.
- Local LLM fits when: sensitive data cannot leave the premises or regulation is strict, low latency / offline is required, and heavy continuous processing makes the total cost of ownership flip in its favor.
- The real answer is often not either/or but a hybrid (confidential = Local / general = Cloud) or routing by task difficulty.