Sealed AI dev setup costs 3〜7% of labor, Zenn guide says

A common impasse in contract development: teams want to use AI but cannot explain the cost, so proposals stall. Its answer is to stop quoting AI as a blanket per-person, full-period line and instead split the estimate by phase, since headcount, duration and tool needs differ at each stage. The worked example is a large mission-critical payment system, chosen because its strict security requirements make the resulting pattern reusable elsewhere.

The architecture is built around a single entry point. With one gateway in front of Amazon Bedrock, personal-information masking, logging and authentication can all be handled in one place, and Claude Code is run through that same path via an environment variable. The body frames AI output as a draft in every phase, with a human approval gate — and draws a firm line at core processing such as amount calculation and balances, where people write the code and two-person approval applies.

The estimate then pairs published unit prices with assumed Bedrock usage per person per month, which is heaviest in implementation at $108, derived from 1.5 million input and 60,000 output tokens a day over 20 days. The smallest phases are mostly initial setup, and ordering phases consecutively lets the same environment be reused, a gap of roughly 1.5 million yen that can be used as proposal material.