Redesigning Yahoo! Shopping's development process with AIDD
AIDDで開発プロセスを再設計する — AI駆動開発の試行錯誤から得た学び
B · Useful, limited Account by the Yahoo! Shopping team at LY Corporation, including its commerce-domain CTO, of fitting AWS's AI-DLC to a 27-year brownfield system; the sequence and design choices are specific, but the single outcome figure is self-reported and approximate.
Summary based on the full text·Original: about 7 minutes to read
In short
A LY Corporation tech blog post (late October 2026; date from the URL) by six people from Yahoo! Shopping, including the commerce-domain CTO. Preparing cross-system changes on a system with 100+ components and 3,000+ repositories depended on a few domain experts. The team adopted AWS's AI-DLC, where agents break business intent into requirements and designs and humans review and decide. It worked in a development-only trial but failed on design for the existing system, because domain knowledge was not organised for AI. They paused, built an AI-readable knowledge base exposed through MCP, then rebuilt the process around their two existing approval gates. Planning effort fell from about two months to about one.
Key numbers
- 100+ / 3,000+ / 27components, repositories and years of history in Yahoo! Shopping
- about 2 months to about 1 monthplanning effort per cross-system project, before and after
- 4in-person hands-on workshops of 3-4 hours each used to roll it out
- 2approval gates (overall design, detailed design) kept from the existing process, each with senior-engineer review
As reported by the source.
Japanese·Web page·opens techblog.lycorp.co.jp
What it reports
- Timeline: AI-DLC introduced at an AWS briefing in July 2025; development-only trial in August-September 2025; the attempt to extend it to planning and design on the brownfield system hit context volume and complexity limits.
- Root cause: documents existed per system but in different formats, places and levels of detail; no document covered relationships between systems.
- Fix: information organised by design phase (system structure for overall design, code-level specs for detailed design) and by domain (search, cart). They defined a ubiquitous language, drew C4 diagrams, recorded repository and system owners, and served all of it, with release schedules and API docs, through an MCP server. Missing documentation was written new.
- Process: Phase 1 produces overall-design approval material (requirements, user stories, workflow plan) and goes to senior review; only then Phase 2 narrows target repos and produces detailed design and unit breakdown for a second review.
- Review load: questions routed by owner (business to planners, system to engineers, UX to designers), and the AI must propose candidate answers from MCP data before asking.
- A customer-journey step, not in AI-DLC, was added from the designers' own practice.
- Prompts are packaged as Agent Skills that force MCP tool use per phase; a project manager can run design through review documents by command alone.
- Result: a planner ran AIDD from proposal to overall-design approval; planning effort went from about 2 months to about 1.
How they got it
The team's description of its own rollout, with sample (partly blurred) AI-generated design documents and workshop feedback. The time figure is given as an approximate before-and-after with several changes combined.
Limits
- One outcome number, approximate, with no count of projects or method.
- No data on design errors caught in review or rework after approval.
- The authors credit several overlapping changes for the time reduction.
- Vendor method (AWS) adapted by a large customer; the vendor link is stated openly.
Why it matters for companies
The useful lesson is the order of work: the method failed until the company's own system knowledge was structured, so the first investment was documentation, ownership data and a shared vocabulary, not prompts. Keeping existing approval gates and routing questions to named owners made the AI output reviewable. Companies with large legacy estates should expect AI-driven design to stall on missing context, and should budget for building that context first.
Key passages in translation
Japanese original
この過程で、手法そのものではなく、AIに渡す情報基盤に根本的な課題があることが見えてきました。
In this process we saw that the fundamental problem was in the information base we gave the AI, not in the method itself.
Japanese original
この課題を受けて、AI-DLCの導入を一度立ち止まり、まずAIが扱える形で情報基盤を整備し直すことにしました。
Given this, we paused the AI-DLC rollout and decided first to rebuild the information base in a form AI can handle.
Japanese original
定量面では、企画工数が導入前の約2ヶ月から1ヶ月ほどに短縮されました。
In quantitative terms, planning effort fell from about two months before adoption to about one month.
Summary based on the full text. Read the original for the full argument and data.Read the original on techblog.lycorp.co.jp