---
id: "lib-2026-10-28-ai-design-doc-review-made-reviewers-lives-easier"
type: "library"
title: "AI design-doc review made reviewers' lives easier. And the team?"
title_original: "AIエージェントに設計書レビューを任せたらレビュワーは楽になった。でも現場は？？"
lang: "JA"
org: "NTT DATA (Zenn)"
source_type: "Practitioner post"
date: "2026-10-28"
week: "2026-W44"
url: "https://nosov.ai/library/lib-2026-10-28-ai-design-doc-review-made-reviewers-lives-easier/"
original_url: "https://zenn.dev/nttdata_tech/articles/6a06f484ebfab7"
tier: "B"
tier_name: "Useful, limited"
tier_reason: "NTT DATA engineer's account of AI design-document review on a roughly 20-person commercial project; the analysis of where judgement work moves is sharp, but the author says the figures are deliberately blurred and based on feel."
summary_basis: "full-text"
original_minutes: "12"
tags: ["ai technology", "future of work", "ai review", "design documents", "workload shift"]
frames: ["verification-capacity"]
---

# AI design-doc review made reviewers' lives easier. And the team?

Original title: AIエージェントに設計書レビューを任せたらレビュワーは楽になった。でも現場は？？

Tier B (Useful, limited): NTT DATA engineer's account of AI design-document review on a roughly 20-person commercial project; the analysis of where judgement work moves is sharp, but the author says the figures are deliberately blurred and based on feel.

Summary based on the full text.

## In short

A Zenn post from NTT DATA (28 October 2026) by an engineer whose team uses GitHub Copilot custom agents to review design documents before they go to human reviewers. The team turned reviewers' implicit criteria into Markdown rules, including which checks the AI may decide and which it must hand back. Reviewer check time fell from about an hour to minutes in some cases. The project as a whole did not get cheaper: the sorting and judgement reviewers used to do moved to the engineers who wrote the documents. The author's point is that raising a correct comment and translating it for the recipient are different jobs, and the AI only does the first.

## Key numbers

- about 1 hour to minutes or tens of minutes: reviewer time for consistency and omission checks per document, as felt, in some cases
- about 80%: share of AI review comments judged valid
- about 10%: comments humans first rejected that turned out to be correct
- 40+: AI comments on a single design document
- around 20: people on the waterfall cloud project where it was trialled

## What it reports

- Setting: full-scratch, cloud-native commercial system, waterfall, about 20 people, designs written in Markdown in GitHub. AI review started around late 2025, first with typo, terminology and template checks.
- Fully delegating design from requirements failed; even the newest models could not find compromises across several customers' and stakeholders' interests.
- The team chose to use top-tier models and not invest heavily in harness work, betting that today's top model capability will be cheap in a few years. Models are pinned per agent and switches recorded.
- Rules capture why a reviewer comments, when to comment, when not to, and who decides. Each review point is split into "is it written" (AI decides) and "is it right" (human decides for items such as responsibility boundaries with other systems). For human-decided items the AI returns a checklist, not a verdict.
- Rules are tuned against human reviews: false positives get exception conditions (for example, no auth comment when a shared platform owns auth), and misses are added.
- In prototype use by most of the engineers: about 80% of comments valid, about 10% were right after humans first rejected them.
- One document drew over 40 comments. Because engineers run the AI before review, they now do the sorting reviewers did. Right after introduction, total project effort felt flat or higher.
- Failure modes: correct comments lost because the recipient lacks meeting context; wrong comments accepted because they read as authoritative (for example, flagging an omission the designer made on purpose).
- Fix in progress: each comment carries "who does what next", and design-level judgements route back to the PM and reviewer team.

## How they got it

The author's team built and ran the agents on a live project and compared AI and human reviews. The post states that the setup is generalised and that the figures are blurred; most numbers are reviewers' impressions.

## Limits

- Figures are deliberately blurred and felt, not measured.
- Still a prototype; total effort including rework has not been measured, and effect on build quality will only be known at release.
- One project with high novelty and customer-specific requirements.

## Why it matters for companies

Most AI review pilots count the reviewer's time saved. This one shows the cost can move to the author instead, especially junior staff who receive abstract comments without context. Teams adding AI review should decide who receives which comments, require each comment to say what to do next, and route design judgements to the people whose job is to judge. Measuring only reviewer time will overstate the benefit.

## Key passages in translation

> レビュワーの工数が減った分、設計書を書いたエンジニアの工数が増えるになります。

The effort reviewers saved became added effort for the engineers who wrote the design documents.

> メタな指摘を「出す」ことと、それを相手に合わせて「翻訳する」ことは、別の仕事でした。AIが肩代わりできたのは、前者だけです。

Raising a meta-level comment and translating it for the person receiving it were separate jobs. The AI could take over only the first.

> 判断のコストは、判断が役割の人が払うべきです。

The cost of judgement should be paid by the people whose role is to judge.

Original: https://zenn.dev/nttdata_tech/articles/6a06f484ebfab7