---
id: "lib-2026-10-23-ota-ward-s-ai-phone-pilot"
type: "library"
title: "Ota Ward's AI phone pilot"
title_original: "大田区、年1000時間の問い合わせを「AI電話」で削減できるか？ 実証実験で見えた課題"
lang: "JA"
org: "キーマンズネット (ITmedia)"
source_type: "Practitioner post"
date: "2026-10-23"
week: "2026-W43"
url: "https://nosov.ai/library/lib-2026-10-23-ota-ward-s-ai-phone-pilot/"
original_url: "https://kn.itmedia.co.jp/kn/article/2610/23/2000002529/"
tier: "B"
tier_name: "Useful, limited"
tier_reason: "ITmedia Keymans Net case report on Ota Ward's AI phone pilot, unusual for publishing a missed target and the design changes it caused; credible, but the figures are the ward's own and the second pilot has no results yet."
summary_basis: "full-text"
original_minutes: "6"
tags: ["public sector", "voice ai", "pilot failure", "target setting"]
---

# Ota Ward's AI phone pilot

Original title: 大田区、年1000時間の問い合わせを「AI電話」で削減できるか？ 実証実験で見えた課題

Tier B (Useful, limited): ITmedia Keymans Net case report on Ota Ward's AI phone pilot, unusual for publishing a missed target and the design changes it caused; credible, but the figures are the ward's own and the second pilot has no results yet.

Summary based on the full text.

## In short

An ITmedia Keymans Net article (23 October 2026) by Taeko Motohiro on Ota Ward in Tokyo, which is testing Softsu's "Mirai AI" automated phone answering through its HOIP startup-pilot programme. The first pilot, in the disaster-prevention division in FY2025, set a goal of cutting staff handling by 80% and reached about 40% effective AI handling. The ward published the gap, analysed why, and reset the second pilot, in the building review division, to a 30% target with answers limited to objective map data.

## Key numbers

- ~40%: calls the AI handled effectively in the FY2025 disaster-prevention pilot, against an 80% goal
- ~1,000 hours: estimated annual phone time in the building review division, target of the FY2026 pilot
- 300 hours / 30%: FY2026 target reduction in phone handling time

## What it reports

- FY2025: callers asked about eligibility for subsidised earthquake breakers and furniture anchors. Effective AI handling was about 40% against an 80% goal.
- Causes named by the ward: callers who cannot state what they need, AI asking repeated clarifying questions until callers hang up, quiet voices and several people speaking at once.
- The ward counted as gains a clear, log-based picture of why AI phone handling is hard and lower internal resistance to trying again.
- FY2026 (1 October 2026 to 28 February 2027): building review division, about 1,000 hours of calls a year, mostly about zoning, coverage ratios and road designations that are already on the ward's public map site.
- Target lowered to a 30% (300-hour) reduction because the knowledge scope is much wider.
- Design changes: answers limited to objective information from the map site; questions needing judgement are declined. Branching scripts replaced by a staff-written Q&A set with synonyms and trade jargon linked to each answer.
- Backend uses ChatGPT and Azure OpenAI Service; the vendor says call data is not stored externally or used for model training.
- Expected benefits beyond hours: fewer interruptions for specialists, 24-hour answering, and voice access for residents who struggle with web search.

## How they got it

Interviews with Ota Ward officials Takefumi Nakatani (innovation programme) and Fusho Nagata (industry and economy division), with call-log samples supplied by the ward. No independent measurement.

## Limits

- "Effective handling" is not defined in the article.
- FY2026 results are not yet available; the 30% is a target.
- Trade-press case story based on the ward's and vendor's statements.

## Why it matters for companies

Voice agents for customer or employee helplines meet the same obstacles: callers who do not know how to phrase the question, and audio that is hard to transcribe. The ward's response is a useful template: restrict the agent to facts with one correct answer, build the knowledge base from staff phrasing and synonyms, and set targets from the last pilot's measured rate rather than from vendor claims.

## Key passages in translation

> 当初目標としていた「職員対応の80％削減」に対し、AIによる有効対応率は約40％にとどまった。

Against the original goal of "cutting staff handling by 80%", the AI's effective handling rate stopped at about 40%.

> AIにとっては『質問文そのものが不明瞭』となり、聞き返しが連続して途中で離脱されてしまうこともありました。

For the AI, "the question itself is unclear", and there were cases where repeated requests for clarification led callers to drop off midway.

> 今回の実証実験は、全ての電話対応をAIに置き換えることが目的ではありません。

The aim of this pilot is not to replace all phone handling with AI.

Original: https://kn.itmedia.co.jp/kn/article/2610/23/2000002529/