Eximbay's AX Standardization Journey Based on AWS Kiro

While individual members' AI usage increased, the fact that this usage didn't translate into a common organizational workflow emerged as the core challenge. Eximbay viewed this as the starting point for AX (AI Transformation) and decided to build not just simple tool proliferation, but a repeatable work structure and a system for accumulating execution assets.

The central axis was AWS Kiro. Eximbay chose Kiro because it could define and repeatedly execute work units, help anyone achieve a certain level of results, and preserve execution methods, once created, as organizational assets.

To spread this across the entire organization, Eximbay designed a 3-layer operational structure.

• Execution Layer: Executing tasks as agent units with Kiro
• Generation Layer: Automatically generating sub-agent configuration files (.md) based on natural language with Kiro Desk
• Accumulation Layer: Managing prompts, agent definitions, steering, and outputs as reusable assets through a version-control-based collaboration structure

In particular, Kiro Desk played a role in lowering the entry barrier for creating sub-agents. By describing roles and expected outcomes in natural language, it automatically generated configuration files, enabling members who didn't know the configuration structure to create sub-agents suited to their own work.

The version-control-based collaboration structure was also important. Eximbay operated by separating team branches from individual branches to manage team-specific prompts, agent definitions, templates, and steering rules, and merged valid individual experimentation results into team assets. As a result, AI usage began accumulating as organizational assets rather than remaining as individual know-how.

Changes appeared first in development organizations. Previously, humans handled everything from requirement understanding to implementation, testing, and documentation, but after adopting Kiro, this shifted to a model where developers define goals, structure, constraints, and quality criteria, while AI generates implementation drafts, test code, and documentation drafts in parallel. For the WeChat MCC automatic classification engine, work that used to take 1.5-2 weeks was reduced to within 1 week, and direct human intervention time dropped to about 3 hours. Batch reprocessing and data correction was also shortened from about 8 hours to about 2 hours.

In non-development organizations, user experience rather than features determined adoption. The existing project management tool's dashboard had sufficient features but was complex and difficult to use, so Eximbay redesigned a new dashboard with Kiro that allows key metrics to be viewed intuitively. The fact that it continues to be used in actual work shows that success or failure of AX in non-development organizations depends less on technical completeness and more on usable experience.

In summary, the core of organizational AX doesn't end with adopting a good tool. Execution structures must be designed with low entry barriers, results must be preserved as assets, and an experience must be created that anyone can naturally start using. Eximbay presented a direction where generation, execution, and accumulation are combined into a single operational structure, allowing AI to establish itself in the organization's way of working.