How a volunteer ride service in Nara lets AI agents answer the calls it cannot.

Every booking comes by phone.

In Tsurumai, a neighborhood of Nara City, the residents' council runs a volunteer ride service. It takes older residents to clinics and shops on Mondays, Wednesdays and Fridays. A one-way ride costs ¥100.

Neighbors started it themselves, led by the chair of the residents' council.

It began in March 2025 as a pilot. In a neighborhood survey, 52% of the 106 households that replied said they would use it.

Almost every booking arrives by phone. Some days bring ten or more calls. Web forms did not catch on with older riders.

We drive for the service and build its booking system. One person could not answer every call; answering alone had reached its limit.

Agents that listen. People who stay in charge.

First, a booking form. It ran on Google Apps Script, and bookings were collected in a Google Sheet. We began with AI-written code pasted in, then built it out with an early version of Claude Code.

Then, a booking system on Cloudflare. Bookings live in one database. Drivers see a live schedule that refreshes every minute, with map directions and a reminder before each pick-up. A public page shows how full each day is — dates, times and counts only, no names.

A phone desk agent. A person answers first; the agent takes the calls they cannot. It needs four things to book: name, day, time and destination. It also asks where to pick the caller up and what the trip is for. It checks the name against the list of registered riders, and stops talking when the caller speaks. At the end it reads the booking back; it is built to book once the caller agrees. No one approves it first. If a detail is missing, it books nothing, tells the caller someone will call back, and alerts the team. If the name does not match the rider list, it still books the ride and marks it for a person to check. It does not say it is an AI when it answers.

A booking checker, in development. A second tool turns a call record into a draft. It checks clashes, free time slots, names and addresses, and waits for a coordinator to approve.

Parts we can swap. We started on a hosted voice platform. In May 2026 we moved to our own parts on a Mac. Since October 2026, live calls run on OpenAI's realtime voice model, with Gemini Live built in as a backup we can switch to by hand. The model behind the conversation can be changed.

Where call data goes. The caller's voice goes to OpenAI's realtime voice model. After the call, the text goes to another OpenAI model, which pulls out the booking. The call record goes into the ride service's booking system. Each call is reported to the team in Telegram, and calls that need a call back also go to Mattermost. The text of each call, and a recording of our agent's side, stay on the reception Mac.

Where Claude fits. We have built with Claude Code since its early days. Today Claude Code is listed in our agent ledger as the resident developer of the phone engine. Claude also helps maintain the booking system. Other agents run on other models, and live calls run on OpenAI's realtime voice model. That is the point: no single vendor holds the keys.

Try the scripted demo of a booking call →

The numbers so far.

• 1,337Ride bookings, 3 Mar 2025 – 28 Aug 2026Booking database · cancellations excluded
• 28→143Bookings per month, first month → busiestMar 2025 → Jul 2026

These numbers show how much the ride service is used. They are the work of the residents' council, its volunteer drivers and its coordinator, not a measure of what AI changed.

• Our May 2026 phone AI setup cost about ¥1,550 a month: ¥550 for the phone number and about ¥1,000 of electricity. Every AI part was open source and ran on our own Mac. On the hosted platform it had been about ¥21,500 a month. Live calls now use a cloud voice model; that cost is not included here.
• On 12 September 2026, in a live test, the agent answered a real call and entered the booking automatically. It was a test booking, removed afterwards. The call lasted 127 seconds and cost about ¥16.
• In late September 2026 the desk was still in test mode: a person answered first, and the agent took the call only when no one could.
• Since 4 October 2026 the desk is live. A person still answers first, and the agent takes the calls no one can.
• Not published yet: how many live calls the agent books on its own. Not measured yet: how much staff time it saves. We will publish both once the numbers are checked.

Who makes it work.

• The residents' council runs the service. Volunteer drivers do the driving, and we drive too.
• Interpreters are a role, not a title: people who know what riders need and can also instruct the agents. They turn needs into instructions, and results back into help.
• A member who spent years at home after working as a front-end engineer builds the phone desk and booking tools with Claude.
• Two 19-year-old twins help on rides, supporting older riders.
• In July 2026 we reported the results to the mayor of Nara City, along with three hard problems: the phone load, having no legal entity, and costs. We received advice on the next steps.
• The ride service is also preparing to become a nonprofit, so it can sign contracts and receive grants in its own name.

Three rules we keep.

• Read everything back. Never guess. Book only what the caller confirms.
• Keep a person close. A person answers first and gets a report on every call the agent takes.
• Make every part replaceable. The service must outlive any one vendor.

Running something similar in your town? Write to us.

info@womnara.com