8 Steps to AI Adoption — Writing the Business Case and Calculating ROI
This is Hamamoto from TIMEWELL.
"We want to adopt AI. But we don't know where to start." "What do we need to explain, and how, to get the business case approved in the executive meeting?" These are questions I hear from corporate planning and DX teams almost every week.
Nomura Research Institute data shows 57.7% of Japanese companies have already deployed generative AI. At the same time, MIT research paints a sobering picture: 95% of companies are failing to see a return on their AI investment. "Deploying it" alone doesn't deliver value — the process design that connects implementation to outcomes is what determines success.
If you're preparing a business case right now, feel free to use this as your starting framework.
The Full Picture of an AI Adoption Project
Eight steps, taken in sequence. Think of this as an iterative journey, not a single pass.
| Step | Phase | Key Activities | Estimated Duration |
|---|---|---|---|
| 1 | Challenge inventory | Identify and prioritize business challenges | 2–3 weeks |
| 2 | Goals and targets | Define KGIs and KPIs | 1–2 weeks |
| 3 | Process mapping | Document As-Is and To-Be states | 2–4 weeks |
| 4 | Solution selection | Compare tools, validate technology | 2–3 weeks |
| 5 | ROI estimation | Calculate returns, prepare business case | 1–2 weeks |
| 6 | Proof of Concept (PoC) | Small-scale test, verify impact | 4–8 weeks |
| 7 | Full deployment | Phased rollout | 8–12 weeks |
| 8 | Reinforcement and improvement | Monitoring and PDCA | Ongoing |
A typical timeline is two to three months to PoC start, and roughly six months to full deployment. "Full company rollout starting next month" is the single biggest cause of failure — don't rush.
Step 1: Challenge Inventory
The first thing to do in an AI adoption project is not tool selection. It's identifying "which business processes have which problems."
The method is simple: ask three questions of managers in each department.
- What takes the most time? (Example: preparing monthly reports takes three days every month)
- What's concentrated in specific individuals? (Example: only veteran employees can handle certain customer inquiries)
- Where do errors most often occur? (Example: a few mistakes per month in manual data entry)
Map the resulting challenges on two axes — "ease of AI application" and "magnitude of impact" — and the starting point becomes clear.
Prioritizing Challenges
| Evaluation Axis | High | Low |
|---|---|---|
| Ease of AI application | Text processing, routine tasks, work where data has already accumulated | Ambiguous judgment criteria, no data, high volume of exception cases |
| Magnitude of impact | Company-wide, high frequency, high cost | Only specific individuals, low frequency, low cost |
The area where both are high is the first target. If you find "a routine task consuming dozens of hours per week company-wide," that's your starting point.
Step 2: Goals and Targets
Once the challenge is identified, set the AI adoption objective and measurable targets. Leave this vague and you'll have no basis for impact measurement later — ending with "we don't really know how it went." In my experience, projects that skip this step have about an 80% failure rate.
Here's the difference between good and bad targets:
- Not OK: "Improve operational efficiency" (what, by how much?)
- OK: "Reduce monthly report preparation time from the current three days to one day"
- OK: "Raise first-response rate on customer inquiries from 60% to 85%"
- OK: "Reduce data entry error rate from five per month to one or fewer"
With numerical targets, you can calculate ROI retroactively and give the executive meeting a clear answer to "was it worth it?"
Looking for AI training and consulting?
Learn about WARP training programs and consulting services in our materials.
Step 3: Business Process Mapping
Before bringing in AI, clarify both the current workflow and the ideal state after AI adoption. The current state is called "As-Is," the ideal is "To-Be." Skip this and you end up with "we deployed AI, but everyone reverted to doing it manually."
Three things to document:
- Write out the current flow. Interview the people doing the work to make visible "what actually gets done in what order"
- Identify which steps AI can replace. Classify each as: full automation, AI assistance, or decision support
- Clarify which steps humans must retain. Leave final judgment, exception handling, and customer-facing interactions in human hands
As a side note: this step fairly often uncovers "actually, the workflow itself was inefficient, even before AI." The business process audit has value on its own merits — don't skip it.
Step 4: Solution Selection
This is where we finally talk about tools. Once the challenge and targets are clear, the selection criteria define themselves.
Solution Evaluation Criteria
| Evaluation Item | What to Check | Importance |
|---|---|---|
| Fit with the challenge | Does it solve the identified problem? | Most important |
| Security | Data storage location, encryption, access controls | Most important |
| Ease of deployment | Integration with existing systems, initial setup burden | High |
| Customizability | Can it be adjusted to fit your company's workflow? | High |
| Cost structure | Initial cost, monthly fee, usage-based billing | High |
| Support | Japanese language support, deployment assistance, training | Medium |
| Scalability | Can it accommodate expanded future use? | Medium |
Never choose based on "it's popular" or "a competitor is using it." The only criterion is: does it fit your challenge?
Step 5: ROI Estimation
When it comes to getting a business case approved, ROI estimation is unavoidable. You need to show "what we stand to gain from AI" in a form executive leadership can use to make a decision.
Identifying Cost Items
| Cost Category | Example Items | How to Estimate |
|---|---|---|
| Initial cost | Licensing, implementation consulting, development | Vendor quotes |
| Operating cost | Monthly subscription, maintenance | Monthly × 12 |
| Human cost | Effort of the deployment team, training costs | Monthly rate × months invested |
| Indirect cost | Impact on existing work, productivity drop during transition | Extrapolate from past IT deployments |
Identifying Benefit Items
| Benefit Category | Example Items | Calculation Method |
|---|---|---|
| Time savings | Reduction in working hours | Hours saved × hourly rate × number of people |
| Error reduction | Avoiding rework caused by mistakes | Cost per error × number of errors eliminated |
| Revenue improvement | Higher close rates from faster response | Improvement rate × average deal size |
| Risk mitigation | Reducing key-person dependency | Explain qualitatively |
ROI Formula
ROI (%) = (Annual Benefit - Annual Cost) ÷ Annual Cost × 100
For example: an AI tool costing ¥6M per year that generates ¥9M per year in operational efficiency translates to ROI of 50%. Projects projecting ROI of 30%+ in year one tend to be approved more readily.
Honestly, benefit estimates always come with uncertainty attached. That's exactly where the numerical targets from Step 2 matter. If you can build the logic "if this number is achieved, ROI will look like this," executive leadership has a much easier decision to make.
Step 6: Proof of Concept
PoC is the phase of testing small-scale and verifying impact before full deployment. The biggest risk here is "PoC death" — getting stuck in PoC cycles without ever progressing to production.
Three Rules for Preventing PoC Death
- Set a time limit. PoC maximum: eight weeks. One extension allowed, with written justification
- Define success criteria in advance. In numerical terms: "if X is achieved it's a Go; if Y is not reached it's a Stop"
- Don't exclude the frontline. Operational staff must be involved from the start. Running PoC through the IT department alone creates a wall when it comes time to deploy
PoC Design Template
| Item | What to Document |
|---|---|
| Validation theme | What are we testing? (Example: automation rate for inquiry handling) |
| Success criteria | Numerical target (Example: automated response rate ≥70%) |
| Validation period | Start to end date (Example: April 1 to May 31) |
| Scope | Department, work function, data range |
| Participants | Project members and roles |
| Go/Stop decision date | Mid-review and final decision dates |
| Next actions | Full deployment plan for Go; exit criteria for Stop |
Step 7: Full Deployment and Rollout
Once PoC confirms the impact, move to full deployment. The ironclad rule: phase the rollout rather than going company-wide all at once.
Rollout sequence:
- Start live operation in the pilot department (the one that ran the PoC)
- Expand to adjacent departments (those with similar work content)
- Company-wide deployment (all departments, full rule documentation)
At each stage, document "what went well" and "what was unexpected," and apply the lessons to the next wave. Collecting frontline feedback at each stage and continuously improving operating guidelines is what drives adoption.
Step 8: Reinforcement and Improvement
Deployment is not the finish line — reinforcement after deployment is where the real work begins.
CloudAce research shows that 80.2% of companies that set KPIs for AI utilization achieved their targets. Conversely, 81.8% of companies that missed KPI targets cited "unstable AI output quality" as the reason. Regular monitoring and improvement cycles are essential.
Reinforcement checklist:
- Are you measuring usage rates monthly?
- Are you regularly reviewing impact KPIs?
- Is there a mechanism for collecting improvement requests from the frontline?
- Are you communicating tool updates and new features internally?
- Are you sharing success stories internally?
The fifth point — sharing success stories — is consistently underestimated, but I consider it the critical factor in adoption. "The department next door achieved this result with AI" is more motivating than any top-down directive.
Elements to Include in the Business Case Document
To translate everything above into a document that clears the executive meeting, here's the format that tends to work:
| Section | Contents | Suggested Length |
|---|---|---|
| Executive Summary | Proposal overview and expected impact (one page, self-contained) | 1 page |
| Background and challenge | Current state problems and case for AI adoption | 1–2 pages |
| Objectives and targets | KGI and KPI definitions | 1 page |
| Solution overview | Selected tool and rationale | 1–2 pages |
| Return on investment | Costs, benefits, ROI projection | 1–2 pages |
| Execution plan | Timeline, team structure, milestones | 1 page |
| Risks and mitigations | Anticipated risks and responses | 1 page |
A total of eight to ten pages is appropriate. Thick documents don't get read.
Five Common Failure Patterns
Five failure patterns I've seen repeatedly in consulting engagements.
Failure 1: Tool-First
"ChatGPT is popular, let's deploy it." Skip challenge identification and start with a tool, and you end up with "we deployed it but don't know what to use it for." This is the most common pattern.
Failure 2: No Executive Involvement
The person driving the project is enthusiastic, but executive leadership is disengaged. Budget and resources stay insufficient; it ends up as a side project for the designated person. IPA data shows 85.1% of Japanese companies feel a shortage of DX talent. Without executives committed to allocating resources, nothing moves forward.
Failure 3: Permanent PoC Loop
Running PoCs repeatedly without moving to production. "Let's collect a bit more data first." "Let's run one more validation." The delays pile up and months pass. Prevent this by defining decision criteria in advance.
Failure 4: Full External Outsourcing
Delegating everything to the AI vendor, leaving no internal know-how. The moment the vendor contract ends, operations stop. Using external resources is a legitimate choice, but if internal talent development doesn't happen simultaneously, it won't sustain.
Failure 5: Company-Wide Launch Without a Pilot
Going directly to full company rollout without a pilot. The front line isn't ready, confusion and frustration spread, and "AI just doesn't work" becomes the prevailing sentiment. A phased approach is non-negotiable.
Summary
Based on everything above, decide on your next action.
If you haven't yet finished the challenge inventory — start by asking three questions to managers in each department. If the challenge is visible — set numerical targets and move to ROI estimation. If you're already in PoC — document the time limit and decision criteria to prevent PoC death.
AI adoption is a project that produces consistent results when you follow the right process. Skip steps and failure is certain. Check where you are in the eight steps right now, and take the next one.
TIMEWELL's WARP provides end-to-end consulting support — from AI adoption strategy development through PoC design, full deployment, and internal talent development. Former DX and data strategy specialists from major enterprises will work with you to design a project framework tailored to your organization's situation. Even just a business case document review is a fine place to start — reach out any time.