Learnings from Operating an AI Running Coach for 11 Months: When Product Development Becomes Scientific Discovery
Figure 1: From left: Marone, Fujiwara, Takahashi. This candid shot captures our practice-first philosophy: prioritizing real conversations with athletes over formal photography.
This article is based on a presentation delivered at the AAII2025 (Asia AI Institute) International Symposium held in Thailand on November 13, 2025. We would like to express our sincere gratitude to all those who provided this valuable research presentation opportunity, as well as to all attendees who participated in the presentation.
TL;DR
11 Months in Practice: Operating "EKIDEN.AI," one of the world's first AI conversational running coaches, in collaboration with Olympic athletes and professional coaches
3 Major Pivots: From human-coach-centric → AI-dedicated; from product → research; from short-term goals → long-term vision (2027 professional team, 2032 Olympics)
Technical Achievements: Direct data access through partnerships with GARMIN and others; integration of objective and subjective data; demonstrated effectiveness of conversation-based reflection
Elite Athlete Performance: Arata Fujiwara's 2:41:17 (one week after a full marathon); Akihiro Kaneko's remarkable 2:13:05 record
Core Discovery: Clarification of human-AI role division
AI strengths: Perfect memory and analysis of data; early warning sign detection
Unique human value: ① Determining direction (Arata Fujiwara) ② Non-verbal understanding and reading the atmosphere (Marone)
7 New Research Areas: Sport Data Science, Human vs AI vs Hybrid coaching, etc.—fields that don't exist in traditional academic domains
Organizational Pivot: From product company → doctoral research support organization; aiming for deep integration of practice and research
Does AI Coaching Really Work?
"Does AI coaching really work?"—To answer this fundamental question, we have been developing and operating "EKIDEN.AI," one of the world's first AI conversational running coaches, for 11 months. The insights gained through this process have evolved beyond mere technical achievements into scientific discoveries about human-AI collaboration.
The project's distinctive feature lies in close collaboration with two professional coaches: Arata Fujiwara, a London Olympic marathoner and current SUZUKI ATHLETE CLUB coach, and Marone Kota Aziz, formerly of the top running team TWO LAPS and a personal trainer who has supported many Japanese national team athletes across various sports. Without their practical insights, this research would not have been possible. Olympic-level athletes and top coaches recognizing their own limitations and seeking optimal collaboration with AI—this process has brought discoveries beyond our expectations.
And what made the speed of these discoveries possible was the involvement of Nguyễn Quang Trung, a hacker based in Germany. Without his participation, characterized by rapid product development and excellent architectural design, it would have been impossible to advance product development and customer validation at the same pace as our research. Technical excellence has supported the rapid iteration of practice and research. We will share more about these technical challenges in another opportunity.
Figure 2: Let me start with a fundamental question: Does AI coaching actually work? What makes our study unique is collaboration with two professional coaches who worked alongside our AI system, providing insights no purely technical development could reveal.
An 11-Month Journey: Three Critical Pivots
These 11 months have been a journey experiencing three major directional changes. Each pivot became a turning point that fundamentally transformed our understanding.
First Pivot: From Human-Coach-Centric to AI-Dedicated
Initially, we designed the platform to support human coaches. The goal was to reduce the burden on human coaches and provide high-quality instruction to more athletes. However, through actual dialogue with athletes, we gained a fundamental realization.
What athletes were seeking was not "training menus" but "understanding, learning, and confidence." While many existing apps focus on menu generation, athletes were grappling with fundamental questions like "Why is this training necessary?" "How should I understand my current state?" and "Is it okay to continue like this?"
This realization led us to drastically reduce complex features and shift to a simple design focused on AI dialogue. Rather than creating menus, we would support athletes to understand on their own and train with confidence through dialogue—this became EKIDEN.AI's new direction.
Second Pivot: From Product to Scientific Research
As we continued operations, we were discovering something beyond mere business. User behavior patterns, psychological changes through AI dialogue, the process of professional coaches recognizing their own limitations—these were phenomena not captured by traditional academic research.
What was particularly impressive was that two professional coaches independently arrived at the same conclusions: "Comprehensive data grasp is impossible for humans" and "But it must be humans who make the final directional decisions." Such insights could only emerge from practice.
We came to recognize that this was scientific research. We were finding answers through practice to fundamental questions about human-AI collaboration, the essence of coaching, and mechanisms of behavior change.
Third Pivot: From Short-Term Goals to Long-Term Vision
Our initial goals were typical startup short-term objectives: beta release and user acquisition. However, as we recognized the depth of our discoveries and the breadth of possibilities, our vision changed significantly.
Launching the world's first AI-coached professional team in 2027 and aiming for Olympic participation in 2032—these are not mere business goals. They represent an academic commitment to scientifically validate AI coaching at the highest competitive level. Not experiments in a lab, but validation on the world's highest stage—the Olympics. This ambitious goal demonstrates the seriousness of our research.
Technical Breakthroughs: Three Key Discoveries
Breakthrough 1: Strategic Partnerships with Hardware Companies
We achieved direct data access through partnerships with major fitness device companies including GARMIN, POLAR, SUUNTO, and WHOOP. This was not merely a technical improvement but a transformation that fundamentally changed the user experience.
Our research revealed that in a world without EKIDEN.AI, athletes often used tools like ChatGPT by taking screenshots of training data and manually uploading them. This "mere few seconds of effort" created not only a barrier to comprehensive data integration but also a significant psychological obstacle. By removing this barrier through direct data access, user engagement frequency and satisfaction improved dramatically.
Coach Arata Fujiwara experienced this change firsthand: "Since eliminating screenshots, athletes clearly dialogue with AI more frequently. I realized that a small difference in effort significantly impacts continuity."
Breakthrough 2: Integration of Objective and Subjective Data
We discovered that objective data like heart rate, pace, and sleep alone was insufficient. Subjective data such as RPE (Rate of Perceived Exertion), daily impressions, mental state, and subtle sensations like "something feels off"—these subjective data proved essential for both AI and human coaches.
However, inputting subjective data is tedious. Therefore, we focused on excellent UX design. One-tap RPE input, voice-recorded impressions (coming soon), emoji-based mood input—these allowed us to build a system for collecting detailed subjective data without psychological burden.
As a result, the judgment accuracy of both AI and human coaches improved. While objective data shows "what happened," subjective data shows "how the athlete felt." Only with both can comprehensive and accurate judgment be made.
Breakthrough 3: The Power of Conversation-Based Reflection
We observed that reflection through dialogue with AI, rather than merely displaying data, leads to deeper understanding of training and confidence building.
Traditional apps displayed graphs and numbers and stopped there. But EKIDEN.AI asks questions: "How was today's training?" "Your heart rate is somewhat high—how did you feel?" This dialogue process encourages athletes' introspection and enhances their self-analysis abilities.
Quantitatively, the frequency of reflections increased and session duration extended. Qualitatively, athletes' confidence that they "understand their training" clearly improved.
Elite Athletes Speak: The Reality of AI Coaching
Beyond theory and analysis, actual top athletes are achieving remarkable results.
Arata Fujiwara: Two Full Marathons in Two Weeks
Arata Fujiwara achieved a remarkable record in November 2024. He ran a full marathon in 3:15:37 on October 26, and just one week later, on November 2, he recorded 2:41:17 (average pace 3:48/km).
Behind this feat of "two full marathons in two weeks, with the second one over 30 minutes faster," was detailed dialogue with EKIDEN.AI. The AI analyzed his data, positioned the first marathon as a "tune-up race for the main event," provided stimulation with a 3-day LT training session (3:42/km, faster than race pace), and achieved perfect peaking.
Fujiwara says: "After years of working on marathons, I gained new insights through dialogue with AI. Most importantly, training reflections increased, and that process itself boosted motivation. Rather than just looking at data, the dialogue with AI about 'why' deepened my understanding."
Akihiro Kaneko: Dialogue with the World
Akihiro Kaneko (Commodiida) recorded 2:13:05 (9th overall) at the Gold Coast Marathon in July 2024. His high school 1500m best was 4:37, and he started from a circle at university. Despite struggling with injuries, after running Hakone Ekiden and New Year Ekiden, he now competes with the world.
For him, EKIDEN.AI is a "daily consultation partner." "Should I push today, or rest? EKIDEN.AI serves as 'another perspective,' combining weather, fatigue, subjective sensations, and data to think together. Not about speed, but how to make meaningful progress—being able to think about that with AI is now enjoyable."
Kaneko's words capture the essence of EKIDEN.AI. The goal is not to create the fastest athletes, but to support everyone who wants to make today a meaningful step. That's the coaching we aim for.
Core Discovery: Human-AI Role Division
The most important discovery was the clarification of human-AI role division that the two professional coaches reached through practice.
Human Coach Limitations—Honest Self-Recognition
Both coaches commonly recognized AI's comparative advantage in comprehensive data grasp, exhaustive memory, advanced analysis, and early warning sign detection.
Coach Arata Fujiwara says: "Honestly speaking, when watching multiple athletes, completely grasping data is impossible. Training content from three months ago, subtle pace changes, heart rate trends—humans simply cannot remember all of this."
Coach Marone agrees: "24-hour monitoring is physically impossible, and our ability to simultaneously analyze multivariate data has limits. When tired, judgment also fluctuates. Acknowledging this was the first step toward AI collaboration."
This honesty of "Olympic-level coaches acknowledging their own limitations" was the starting point of our research.
AI's Comparative Advantage—Perfect Memory and Analysis
AI's strengths are clear:
Perfect memory: Instant access to all data for all athletes; zero information loss
Advanced analysis: Complex pattern extraction; immediate calculation of multivariate correlations
Early warning sign detection: Detection of overtraining and injury risk precursors
Consistency: Objective judgment not swayed by emotions or fatigue
24/7 availability: Continuous monitoring and real-time feedback
Coach Arata Fujiwara particularly notes AI's "warning sign detection" capability: "Small changes that humans would miss—subtle increases in heart rate, declines in sleep quality, discrepancies between RPE and objective data—being able to detect these early is a major strength of AI."
Unique and Essential Human Value—Two Insights
However, what only humans can do also became clear. And interestingly, the two coaches identified different but complementary human values.
Coach Arata Fujiwara: "Determining Direction"
What Coach Fujiwara emphasized was that final strategic decision-making must be done by humans.
"Even if AI presents options and data, it's humans who ultimately decide 'which direction to go.' Risk-taking or safe strategy, short-term results or long-term growth, performance or health—these trade-offs involve value judgments. And it's humans who take responsibility for those decisions."
He uses a striking metaphor: "AI provides a perfect map. But it's humans who choose the destination. AI shows us where we can go, but where we should go is decided by the athlete and coach."
Coach Marone: "Non-Verbal Understanding and Reading the Atmosphere"
What Coach Marone identified was uniquely human sensory capabilities that don't appear in data.
"There are moments when an athlete says 'I'm fine,' but you can tell from their expression, attitude, and tone of voice that 'they're really tired.' The atmosphere the moment you enter the training ground, the team's overall motivational state, subtle changes in relationships with others—these cannot be captured numerically."
The Japanese expression he uses, "reading the atmosphere" (空気を読む), actually represents a universal human capability that transcends culture. Understanding non-verbal communication, grasping context, discerning timing—these are essential capabilities for human coaches in any culture.
The Ideal Collaborative Model: Perfect Complementarity
Integrating these two insights reveals an ideal collaborative model:
AI's role: Comprehensively present options and data. Communicate "You have three options. The data shows this."
Human's role: Determine direction and perceive non-verbal information. Judge "Having seen the data, this is what this athlete needs now."
This is not replacement. It's perfect complementarity. By having AI handle the heavy lifting of data analysis, human coaches can focus on what only humans can do—strategic judgment and emotional support.
Discovery of Seven New Research Areas
From this 11-month practical experience, at least seven new research areas have become apparent:
Sport Data Science - New methods for integrating objective and subjective data; quantification of subjective indicators
Comparative Study of Human vs AI vs Hybrid Coaching - Respective effectiveness; optimization of application scenarios
Behavior Change Patterns in AI-Coached Athletes - Motivation, learning curves, development of autonomy
Quantification of Human Coaches' Cognitive Limitations - Memory capacity, processing ability, judgment biases
Structural Analysis of Direction-Determining Function - Strategic decision-making processes; mechanisms of value judgment
Non-Verbal Communication in Coaching - Information uncapturable by AI; cultural variation
Psychology of Human-AI Interaction - Trust building, dependency risks, optimal collaborative patterns
These are areas where traditional academic specialization and dominant theories and knowledge (at least to our knowledge) don't necessarily exist. We believe these areas, with data science at their core, hold great potential for new interdisciplinary research. We believe these areas cannot be discovered without deep integration of practice and research.
From Product Company to Doctoral Research Support Organization
With this recognition, we have decided on a fundamental organizational pivot: from product company to doctoral research support organization.
This does not mean abandoning the product. Rather, it means deeply integrating product and research to advance both practical application and scientific knowledge. The product called EKIDEN.AI simultaneously becomes a research platform. Users receive the best coaching while simultaneously contributing to scientific advancement.
And we will train researchers who can earn doctoral degrees in this new field. "PhD in AI × Running Coaching"—this currently does not exist in any university program. We will pioneer that path.
Our goals of launching the world's first AI-coached professional team in 2027 and Olympic participation in 2032 are expressions of this determination. These are not mere business goals but scientific commitments to validate these insights at the highest competitive level.
Conclusion: When Product Development Becomes Scientific Discovery
Our central insight is that "product development can become scientific discovery."
This happens when the following conditions align:
Collaborate deeply with domain experts - Not superficial advice, but deep involvement in daily practice
Experts themselves can clearly recognize their own limitations - Honesty where Olympic coaches can say "I can't do it"
Operate an actual system for a sufficient period and observe actual behavior patterns - In the real world, not in a lab
Flexibly pivot based on learning - Courage not to stick to the initial plan
Recognize that some insights can only be gained from practice - Truths that cannot be reached by desk theory alone
This represents a new paradigm: AI × expert collaboration. Practitioners and researchers working together can advance both practical application and scientific understanding.
We are not just building a better coaching tool. We are discovering fundamental truths about how humans and AI can collaborate. And those discoveries will become universal insights applicable beyond running coaching to all professional domains.
Acknowledgments
This article was written based on a presentation delivered at the AAII2025 (Asia AI Institute) International Symposium held in Thailand on November 13, 2025. We would like to express our sincere gratitude to all those who provided this valuable research presentation opportunity, as well as to all attendees who participated in the presentation.
For more information about EKIDEN.AI, please visit our official website (https://ekiden.ai) or join the Strava club "EKIDEN.AI". If you are interested in research collaboration or doctoral programs, please feel free to contact us.
At Musashino University's School of International Data Science, we welcome colleagues who will advance research with entrepreneurial thinking alongside us. Integrating practice and research to generate new insights—won't you join us in this challenge? We look forward to changing the world together starting April 2026.
