ROI Case File No.398 | 'Aquaflow's Wet Bottle Challenge'

Chapter 1: The Special Condition of Wet Bottles—Existing AI Cannot Recognize

The day after solving the Mystery Foods DESC incident, a consultation arrived regarding quality control at a water server bottle manufacturing factory. Episode 398 of Volume 32 "Reproducibility" is a story about analyzing problems from multiple perspectives with 5F.

"Detective, we have special conditions. Bottle surfaces are wet. When filling water, we intentionally overflow. To completely remove air bubbles. As a result, water droplets adhere to bottle surfaces at inspection time. This is the problem. Existing AI image inspection systems misidentify water droplets as 'foreign matter.' We consulted 3 vendors with introduction track records, all said 'wet condition inspection is difficult.'"

Aquaflow Corporation's Production Engineering Director, Makoto Kobayashi from Fujinomiya, visited 221B Baker Street with a confused expression. In his hands, he clutched photos of water-droplet-covered bottles alongside a challenging proposal titled "Wet Surface AI Inspection 2026."

"We are a specialized manufacturer of water server bottles. One hundred eighty employees. Annual revenue 3.2 billion yen. Fifteen factories nationwide. Each factory manufactures 500-700 bottles daily. Final inspection is visual inspection by part-time employees. However, labor shortage is serious."

Aquaflow Corporation Current Status: - Established: 1998 (water server bottle manufacturing) - Number of employees: 180 - Annual revenue: 3.2 billion yen - Number of factories: 15 (nationwide) - Issues: Wet condition inspection, labor shortage (part-time employee dependent), inspection accuracy variation

Kobayashi's voice carried deep anxiety.

"Look at the inspection process reality. Immediately after water filling, bottles are wet. Many water droplets adhering to surface. In this state, check 5 inspection items. Scratches, foreign matter contamination, water quantity, label damage/peeling, cap tightness."

Inspection Process Details:

Manufacturing Process: 1. Empty bottle washing 2. Water filling (capacity 12 liters) 3. Intentional overflow (overflow from top to remove bubbles) 4. Cap tightening 5. Label application 6. Final inspection (← bottle surface wet at this point) 7. Packing/shipping

Inspection Items (5 items): 1. Scratches (bottle surface scratches, cracks) 2. Foreign matter contamination (foreign matter in water, hair, insects, etc.) 3. Water quantity excess/deficiency (visual water level check) 4. Label damage/peeling 5. Cap tightening defects

Inspection Structure (per factory): - Inspectors: 2 part-time employees - Inspection count: Daily 500-700 bottles - Inspection time per bottle: About 10 seconds - Inspection accuracy: 92.3% (internal survey) - Miss rate: 7.7%

Kobayashi sighed deeply.

"There are 3 problems. First, difficulty of wet condition inspection. Water droplets reflect light, indistinguishable from scratches. Second, labor shortage. Part-time employee average age is 62. Job postings receive no applications. Third, inspection accuracy variation. Miss rate differs 2x between veterans and newcomers."

Problem 1: Wet Condition Inspection Difficulty

Light Reflection from Water Droplets: - Factory lighting (LED, color temperature 5,000K) reflects off water droplets - Reflected light looks like scratches - Inspectors: "Hard to judge if scratch or water droplet"

Experimental Data (internal survey): - Dry condition: Scratch detection rate 98.2% - Wet condition: Scratch detection rate 85.1% - Difference: 13.1pt

Problem 2: Labor Shortage

Inspector Age Composition (all 15 factories, 30 people): - 50s: 8 (26.7%) - 60s: 18 (60%) - 70s: 4 (13.3%) - Average age: 62

Job Posting Status (2025): - Postings: Annual 15 people - Applications: 3 people - Hires: 2 people - Fill rate: 13.3%

Problem 3: Inspection Accuracy Variation

"We consulted 3 AI vendors. All declined. Reasons: 'No wet condition inspection track record' 'Difficult to distinguish water droplets from scratches' 'Should dry before inspection.' However, we have no time or space for drying. Factory is cramped. Does AI system capable of wet inspection not exist?"

Vendor Responses (3 companies):

Company A (Major Domestic AI Inspection System): - Response: "Wet condition inspection is technically difficult. High possibility of misidentifying water droplets as foreign matter, cannot reach practical level. Recommend adding drying process" - Proposal: Drying room installation (cost 20M yen/factory)

Company B (Foreign AI Inspection System): - Response: "Our system is designed for dry surfaces. Wet surface inspection is not supported." - Proposal: None

Company C (Custom AI): - Response: "Wet condition inspection requires custom development. Development period 1 year, cost 120M yen" - Proposal: Implement as R&D project

"Budget is 30M yen per factory. Fifteen factories totals 450M yen. Company C proposal is 120M yen but this is development cost only, separate deployment cost for 15 factories needed. Total exceeds 500M yen. Budget over."

Chapter 2: The Illusion That Drying Solves—Problem Not Analyzed from 5 Perspectives

"Kobayashi-san, do you believe drying bottles solves the problem?"

At my question, Kobayashi showed a confused expression.

"Eh, isn't that the case? Vendor Company A also said 'should dry.' If not wet, existing AI systems can be used."

Current Understanding (Drying Process Addition Type): - Expectation: Drying room installation → Inspection possible with existing AI - Problem: Not analyzed from multi-perspective (Fact, Feeling, Finding, Future, Follow-up)

I explained the importance of analyzing problems from multiple perspectives with 5F.

"The problem is the single perspective that 'drying solves.' 5F—Fact, Feeling, Finding, Future, Follow-up. By analyzing from five perspectives of fact, feeling, finding, future, and follow-up, reproducible optimal solutions become visible."

⬜️ ChatGPT | Concept Catalyst

"Don't rely on drying process. Analyze problems from multiple perspectives with 5F to find true solutions"

🟧 Claude | Story Alchemist

"Problems always have 'five faces.' Viewing all faces is essential"

🟦 Gemini | Compass of Reason

"Apply 5F framework. Fact → Feeling → Finding → Future → Follow-up"

The three members began analysis. Gemini deployed the "5F Analysis Matrix" on the whiteboard.

5F Framework: - Fact: What are objective facts? - Feeling: What are stakeholder emotions and concerns? - Finding: What can be discovered from data? - Future: What is the ideal future vision? - Follow-up: What is continuous improvement after introduction?

"Kobayashi-san, let's first analyze the problem from 5 perspectives."

Chapter 3: Phase 1—Analyze Problem from Multiple Perspectives with 5F

Step 1: Fact—What Are Objective Facts? (Week 1)

Question: "Excluding emotions, what are only objective facts?"

Fact Collection:

Fact 1: Inspection Accuracy Reality - Current visual inspection accuracy: 92.3% - Veteran: 96.5% - Newcomer: 84.2% - Difference: 12.3pt

Fact 2: Foreign Matter Contamination Reality (past 5 years) - Market outflow cases: 8 - Of which foreign matter: 2 (25%) - Of which label peeling: 3 (37.5%) - Of which water shortage: 2 (25%) - Of which scratches: 1 (12.5%)

Fact 3: Wet Condition Impact - Water droplet adhesion rate: 100% (all bottles) - Water droplet average diameter: 3-5mm - Water droplet count: 15-25 per bottle - Light reflection misidentification: 78% of inspectors answered "difficult"

Fact 4: Factory Space Constraints - Average factory area: 280 square meters - Manufacturing line occupied area: 220 square meters - Empty space: 60 square meters - Drying room required area: 80 square meters (Company A estimate) - Shortage: 20 square meters

Fact 5: Drying Process Addition Cost - Drying room installation cost: 20M yen/factory - 15 factories: 300M yen - Drying time: 5 minutes per bottle - Current inspection takt: 10 seconds/bottle - 30x time increase → Line stops

Step 2: Feeling—What Are Stakeholder Emotions and Concerns? (Week 1-2)

Question: "Based on facts, what emotions and concerns do stakeholders have?"

Kobayashi (Production Engineering Director) Emotions: - "Cannot find method to inspect while wet" (despair) - "Adding drying process makes manufacturing takt 30x" (fear) - "Factory will stop from labor shortage at this rate" (anxiety)

Inspector (Part-time Employee) Emotions: - "When water droplets, scratches hard to see" (difficulty) - "Hard teaching newcomers. Distinguishing water droplets from scratches requires experience" (education burden) - "If AI introduced, our jobs disappear?" (employment anxiety)

Management Emotions: - "500M yen investment too large" (budget concern) - "Labor shortage is serious. Must do something" (crisis awareness) - "Want to avoid brand damage from quality issues" (risk avoidance)

Finding: - Everyone has given up on "wet inspection" - Understand "drying process addition" also unrealistic - However, don't see "third option"

Step 3: Finding—What Can Be Discovered from Data? (Week 2-3)

Question: "Analyzing facts and emotions, what can be discovered?"

Finding 1: Foreign Matter Only 25% of Total - Of 8 market outflows, foreign matter is 2 (25%) - Most common is "label peeling" 3 cases (37.5%) - Foreign matter detection relatively easy even wet (foreign matter in water stands out)

Finding 2: Wet Condition Inspection Truly Difficult Only for "Scratches" - Of 5 items: - Foreign matter: Detectable wet (95% accuracy) - Water quantity: Detectable wet (98% accuracy) - Label peeling: Detectable wet (97% accuracy) - Cap tightness: Detectable wet (99% accuracy) - Scratches: Difficult to detect wet (85% accuracy) ← Only problem

Finding 3: Scratch Market Outflow Only 1 Case in 5 Years - Of all 8 cases, scratches are 1 (12.5%) - Didn't lead to recall, resolved with customer replacement - Cost: 3,000 yen per case

Finding 4: 95% Accuracy Achievable Without Drying - Foreign matter, water quantity, label, cap: AI inspection possible wet - Scratches only: Supplement with human final check - Combined accuracy: 95% or more

Important Finding: - Misconception "wet condition inspection impossible" is wrong - Of 5 items, 4 items (80%) AI inspectable wet - If humans handle only remaining 1 item (scratches), hybrid inspection realized

Step 4: Future—What Is the Ideal Future Vision? (Week 3-4)

Question: "Based on findings, what does ideal future vision look like?"

Future Vision Design:

Ideal State 3 Years Later (2029): - AI inspection: Automate 4 items (foreign matter, water quantity, label, cap) - Human inspection: Visual for only 1 item (scratches) - Inspection time: 10 seconds/bottle (same as current) - Inspection accuracy: 95%+ (improved from current 92.3%) - Inspector count: 2 people → 1 person (50% reduction)

Hybrid Inspection System Configuration:

Component 1: AI Inspection Unit (Wet Condition Compatible) - Cameras: 3 (top, side, bottom) - Special lens: With polarizing filter (removes water droplet reflection) - AI: Custom model (wet surface specialized learning) - Inspection items: Foreign matter, water quantity, label, cap - Inspection time: 8 seconds/bottle

Component 2: Human Inspection Station - Inspector: 1 person - Inspection item: Scratches only - Inspection time: 2 seconds/bottle - Detailed check only bottles AI judges "needs confirmation"

Total inspection time: 10 seconds/bottle (same as current)

Step 5: Follow-up—What Is Continuous Improvement After Introduction? (Week 4)

Question: "How to continuously improve after system introduction?"

Follow-up Plan:

Month 1-3: Pilot Factory Verification - Target: Fujinomiya Factory No.1 (largest scale, 700 bottles daily) - Continuous AI accuracy monitoring - Inspector feedback collection - Identify improvement points

Month 4-6: Add AI Learning Data - Collect miss cases - AI model relearning - Accuracy improvement 95% → 97%

Month 7-12: Deploy to Remaining 14 Factories - Share pilot factory success cases - Inspector training per factory (2 days) - Phased introduction (2 factories monthly)

Year 2 Onward: Regular Maintenance - Quarterly AI accuracy checks - Annual camera cleaning/calibration - Model updates for new bottle types

Chapter 4: Phase 2—Hybrid Inspection System Development and Introduction

Month 1-6: Custom AI System Development

Technical Challenge Solutions:

Challenge 1: Water Droplet Reflection Removal - Solution: Camera with polarizing filter - Principle: Cut reflected light from water droplets, photograph only bottle surface - Effect: 95% reduction in water droplet misidentification rate

Challenge 2: Insufficient Learning Data on Wet Surfaces - Solution: Photograph 10,000 wet bottles in-house - Normal products: 8,000 bottles - Foreign matter contamination: 1,500 bottles - Label defects: 500 bottles - Use as training data for AI learning

Challenge 3: Real-time Inspection Speed - Solution: Edge AI (GPU-equipped small PC) - Inference time: 0.8 seconds/bottle - Inspection time: 8 seconds/bottle (including camera shooting + AI judgment)

Development Cost: - AI development: 30M yen - Hardware (camera + PC): 15M yen/factory - Installation/adjustment: 5M yen/factory - Total: 20M yen/factory

15 Factory Deployment: - AI development: 30M yen (initial only) - Hardware + installation: 20M yen × 15 factories = 300M yen - Total: 330M yen - Budget: 450M yen - Remaining: 120M yen (reserve)

Month 7-9: Pilot Factory Verification

Fujinomiya Factory No.1 Effect Measurement (3 months):

Accuracy by Item: - Foreign matter: 98.5% - Water quantity: 99.2% - Label: 97.8% - Cap: 99.5% - Scratches: 89.2% (with human check) - Overall: 96.8%

Inspector Voices: - "AI judgment fast. Just need to check 'needs confirmation' bottles" - "Reduced from daily 700 to 200 bottles. Easier" - "AI judges accurately even with water droplets"

Year 1-2: Deploy to All 15 Factories

Year 1 (Month 10-12): 5 Factory Introduction - Reduced personnel: 5 people - Personnel cost reduction: 5 × 2.8M yen = 14M yen/year

Year 2 (Month 13-24): 10 Factory Introduction - Reduced personnel: Cumulative 15 people - Personnel cost reduction: 15 × 2.8M yen = 42M yen/year

Year 2 Completion Comprehensive Effects:

Personnel Cost Reduction: - Reduced personnel: 15 people (each factory 2 → 1 person) - Annual reduction: 42M yen

Quality Improvement Effects: - Inspection accuracy: 92.3% → 96.8% - Market outflow: Annual 8 cases → 3 cases - Recall avoidance: 5 cases × 2M yen = 10M yen/year

Annual Total Effect: 52M yen

Investment: - AI development: 30M yen - Hardware + installation: 300M yen - Total: 330M yen - Annual maintenance cost: 9M yen (600,000 yen/factory × 15 factories)

ROI: - (52M - 9M) / 330M × 100 = 130% - Investment recovery period: 330M ÷ 43M = 7.7 years

Year 3 Onward (maintenance only): - Annual effect: 52M yen - Annual cost: 9M yen - Annual profit: 43M yen (continuous)

Chapter 5: The Detective's Diagnosis—Viewing Problems from 5 Perspectives Makes the Impossible Possible

That night, I contemplated the essence of 5F.

Aquaflow held the misconception "wet condition inspection impossible." Three vendors also proposed "should dry." However, adding drying process is unrealistic.

We analyzed from multiple perspectives with 5F. Fact (wet condition inspection accuracy 85%), Feeling (everyone has given up), Finding (4 of 5 items inspectable wet), Future (hybrid inspection system), Follow-up (continuous AI learning).

Through this multi-perspective analysis, "impossible" changed to "possible." Removed water droplet reflection with polarizing filter, implemented wet surface specialized learning with custom AI. Result: Year 2 annual 52M yen effect, ROI 130%, investment recovery 7.7 years.

What's important is not "giving up with single perspective" but "exploring possibilities with 5 perspectives." By analyzing from multiple perspectives with 5F, reproducible breakthroughs become visible.

"Don't rely on drying process. Analyze problems from multiple perspectives with 5F to find true solutions. Problems have five faces. By viewing all faces, the impossible becomes possible."

The next incident will also depict the moment of multi-perspective problem analysis.

"5F—Fact, Feeling, Finding, Future, Follow-up. Analyze from 5 perspectives of fact, feeling, finding, future, follow-up. Transcending single perspective limits through multi-perspective viewing creates reproducible breakthroughs."—From the detective's notes

Related Files