[Event Report] Don't Let AI Do the Implementation, Design the Recognition! — The Frontline Where AI-Driven Development Enters the 'Business Penetration Phase'

DXHR (hereinafter 'DXHR') held an event for engineers and executives titled 'AI-Driven Development Study Session: Horizontal Connections and Changing Human Roles' on May 25, 2026, at its co-working space, 'NeuroHub'.

This event provided a high-density environment for exchanging 'first-hand information' that only practitioners could offer, covering everything from the philosophy of 'designing recognition' instead of letting AI do the implementation, to the realities of FDE (Forward Deployed Engineer) and the behind-the-scenes of contract development that completed 15 person-months of work in just one day.

Background The application of generative AI has clearly shifted from the phase of coding assistance for engineers to a phase of permeation throughout corporate activities. While 'AI-driven development' is already a past discussion from the perspective of Silicon Valley investors, with the new battleground being how to implement AI into business, many large companies in Japan have only just begun to learn the basics of prompts. This temperature gap is the most acute issue for engineers and executives thinking about their next steps.

At the start of the event, organizer Mr. Ito explained the event's policy, stating, 'In the AI era, the value of real human connections is increasing,' and launched the event with three objectives: - Expand horizontal connections: In an era where information gathering is homogenized by deep research, cultivate a network that shares field knowledge of AI implementation through first-hand information that can only be obtained through real dialogue. - Learn from experts: Directly absorb the latest trends and practical know-how from participants and panel guests active on the frontline, such as those from NTT. - Create business opportunities: Foster natural collaboration and order placement opportunities among executives, engineers, and non-engineers, without excessive sales pitches.

LT Session & Panel Discussion Speakers and guests active in designing and operating AI agents shared trial-and-error experiences and next steps through three perspectives: 'Recognition Design' to entrust implementation to AI, the fundamental shift in AI-driven development trends, and the realities and division of labor of FDE.

1. Don't let AI do the implementation, design the recognition (Speaker: Mr. Furuido, CEO of WellAI) Mr. Furuido, who has a unique career path from a radiologic technologist to a programmer via side-hustle LP production, and currently leads four products including AI call centers and AI avatars, stated, 'The difference in the AI era is determined by the precision in eliminating ambiguity.' He experienced a vicious cycle where throwing vague instructions at AI resulted in the AI complementing the design with something 'plausible,' leading to misalignment with every revision. Even with Claude's Plan Mode, premise recognition gaps remained. Therefore, he introduced 'Grill Me,' where the AI first asks the user questions. By eliminating premises, boundary conditions, and exceptional cases in advance, he was able to move to a structure where 'Implementation is done by AI, and judgment is done by humans,' with almost no code written. Internally, he operates the 'AI Factory' with a three-layer structure of philosophy, process, and daily guidelines, along with the 'Four Principles of AI-Driven Development' (Pass information / Set guardrails / Let AI run / Keep improving). In conclusion, he strongly emphasized that in the AI era, it is the 'implementer' who will be replaced, and the one who remains is the 'person who can design recognition'—stressing the importance of having the power to align recognition as a condition for being the one who 'places the order' to AI, rather than the one being ordered by AI.

2. Panel Discussion: The Frontline of AI-Driven Development and the Reality of FDE The guest was Mr. Kumai (Kumai Research Institute), who claims to be the first to promote 'requirements definition and design' for AI-driven development (from 2023). He pointed out that from the Silicon Valley perspective, 'AI-driven development' is already in the past, and implementing AI into business is the new battleground. He shared a success story of completing a contract project estimated at 15 person-months in just one day (10 hours) by incorporating an automatic loop into Claude Code. On the other hand, he also pointed out the gritty reality of the field, such as cases in 40-50 person-month bug response projects where 'the DB environment was not built in the first place.'