From 60 Minutes to 10: How We Deployed AI Agents to Japanese Government Offices and Railway Stations

A Filipino ski instructor walks into Myoko City Hall in Niigata Prefecture. She needs to register as a foreign resident — tax enrollment, pension, health insurance, the works. The forms are in Japanese. She does not read kanji. The staff speak some English, but not enough to explain how the Japanese pension system works or why she needs to separate her garbage into eight categories.

Before January 2026, this interaction took 30 to 60 minutes. Staff would pass a handheld translation device back and forth, typing sentences one at a time. During peak ski season, the line would stretch out the door.

We built an AI agent to handle this. It now takes under 10 minutes.

!RIN agent deployment — Myoko City administrative AI for foreign residents

---

The Problem Goes Beyond Translation

Any language model can translate "you need to enroll in nenkin" into Tagalog. The translation alone is useless. Nenkin is the Japanese national pension system, a concept that does not map directly to the social security systems in the Philippines, Vietnam, or Australia.

What the resident actually needs to hear is: "Japanese law requires you to pay into a pension fund every month. This is separate from health insurance. If you leave Japan before retirement age, you can apply for a partial refund called dattai ichijikin. Here is how much you will pay based on your income bracket."

That kind of explanation is domain-specific concept mediation: taking a Japanese administrative system and making it comprehensible to someone whose reference frame is a completely different country's bureaucracy.

The core technical challenge behind everything we build at EFFOMA is making that mediation work reliably at scale, which is the reason we chose Claude as our reasoning engine.

What We Built

EFFOMA operates from Ho Chi Minh City, Sydney, and Tokyo (Minato). Our platform, Agentfront.ai, powers specialized AI agents deployed on physical hardware, including digital signage, interactive touchscreens, and mobile devices, in government offices, railway stations, and healthcare facilities across Japan.

Claude by Anthropic is the reasoning layer at the center of this platform.

!ARISA AI agent — production deployment at exhibition

RIN — Myoko City Hall

Client: Myoko City Government, Niigata Prefecture Deployment partner: Flyout Planning Corporation (Tokyo) Live since: January 2026

RIN is deployed at Myoko City Hall and the Myoko High-Plateau Branch Office on digital signage stations with integrated hardware. The agent handles the complete registration workflow: greeting the resident in their language, reading their residence permit data, and guiding them through every required form and procedure with culturally adapted explanations.

Automating the workflow itself was relatively straightforward. The deeper challenge was making Japanese administrative concepts genuinely understandable to someone whose entire bureaucratic reference frame is a different country. Claude's cross-cultural reasoning makes this possible at a level that simple translation never could.

!Myoko ski resort region — where foreign residents register during peak season

Measured results: - Registration time: 30-60 minutes down to under 10 - Staff throughput during peak season: 3-4x improvement - Form errors: near-zero (data comes from the IC card, not manual entry in an unfamiliar script)

ARISA and ALICE — Seibu Railway Stations

Client: Seibu Railway Co., Ltd. and Yokoze Town Deployment partner: Flyout Planning Corporation Live since: February 2026 (Ashigakubo Station; expanding to Yokoze and Seibu-Chichibu stations)

The Chichibu region west of Tokyo draws international tourists for ice pillar light-up events, mountain hiking, and historic temples, including the rare Soro-kaicho pilgrimage that occurs once every 12 years. But rural railway stations here had zero multilingual support. A tourist arriving at Ashigakubo Station had no way to find out where to eat, which trail to take, or when the last train leaves.

ARISA and ALICE are sister AI agents deployed as 3D avatars on station digital signage:

- ARISA handles detailed tourism guidance: restaurants, hiking trails, temple hours, event schedules, and cultural context - ALICE has a warmer, more approachable style suited for families and casual visitors

When a visitor scans a QR code generated by the agent, the conversation transfers to their personal smartphone. The agent becomes a companion guide that stays with them throughout their trip, answering questions long after they have left the station.

!Seibu Railway partnership — ARISA and ALICE AI tourism agents

The agents draw from curated knowledge bases: Yokoze Town tourism data, Chichibu Regional Hospitality Tourism Association content, Seibu Railway schedules, and local event calendars. Each agent works from verified, locally maintained information rather than searching the open internet.

!Chichibu region — natural scenery surrounding Seibu Railway deployment area

SOFIA — Healthcare

SOFIA assists patients with appointment scheduling, check-in, hospital navigation, and insurance process explanation. SOFIA has one hard rule: she never provides medical advice. Medical questions are routed to healthcare professionals immediately, with a conversation summary attached so the patient does not have to repeat themselves.

---

Why Claude

I want to be specific about this, because "we evaluated multiple models and chose Claude" is a meaningless sentence.

Here are three concrete things that determined our choice:

It Handles Japanese Formality Registers Correctly

Japanese has multiple levels of formality (keigo). Government communication requires the highest register, sonkeigo and kenjougo. An agent that speaks casual Japanese in a municipal office destroys credibility with both staff and residents.

During our evaluation, we tested identical prompts across models for formality consistency in Japanese government contexts. Claude maintained correct keigo across multi-turn conversations without degradation. This was not true of every model we tested.

For our deployments, this matters more than benchmarks. The city hall staff watch these interactions. If the agent speaks inappropriately, they lose trust in the system, and we lose the deployment.

It Explains Concepts Instead of Translating Words

I gave an example earlier: the pension enrollment problem. Claude does not transliterate "nenkin" or provide a dictionary definition. It constructs an explanation calibrated to the resident's likely frame of reference.

When a Vietnamese resident asks about health insurance, Claude explains the Japanese system in terms of how it differs from Vietnam's social health insurance (BHXH), because that comparison is useful. When an Australian asks the same question, the explanation references Medicare, because that is the relevant reference point.

Claude's existing cross-cultural knowledge is deep enough that our domain-specific knowledge layer builds on it rather than replacing it. We did not prompt-engineer this from scratch, and that saved us months of work.

It Respects Boundaries Under Pressure

Our agents operate in environments where getting something wrong has real consequences. A resident who receives incorrect tax guidance faces penalties. A patient who gets medical misinformation could be harmed.

Claude consistently respects the scope boundaries we define. When a resident asks RIN a legal question that goes beyond standard registration guidance, RIN transfers to a human staff member with a conversation summary. It does not guess. It does not hedge. It escalates.

We have run adversarial testing, with attempts to make the agents break character, provide unauthorized information, or bypass escalation rules. Claude's safety training maps well to our production requirements. The model's instinct toward caution is an asset in government and healthcare contexts, not a limitation.

!ARISA — AI agent poster for Agentfront.ai platform

---

What Multilingual Actually Means in Practice

I want to address this directly, because listing a language count sounds like a feature checkbox. Any API can claim multilingual support.

The hard part is generating text that works, not just text in another language. Specifically:

Concept adaptation. Japanese administrative systems do not have equivalents in most countries. Our knowledge layer maps Japanese procedures to explanations that make sense in the resident's cultural and bureaucratic context. This is domain engineering, not prompt engineering.

Register matching. A government agent must be formal in Japanese but conversational in the resident's language, because the resident is stressed, confused, and standing in a government office in a foreign country. Matching the right tone in the right language at the right moment is a nuanced problem.

Code-switching. Tourists and residents frequently mix languages mid-sentence. A visitor at Ashigakubo Station might start in English, use a Japanese place name, then ask a follow-up in Mandarin. The conversation needs to handle this fluidly, not restart every time the language shifts.

Script and layout edge cases. Right-to-left languages, complex script rendering, and languages where formality markers restructure entire sentences all require dedicated testing. Going from five languages to twenty was straightforward. Going further surfaced problems we did not anticipate.

The value is not in the number. The value is in getting each interaction right for the person standing in front of the screen.

---

Where We Operate

Tokyo, Japan (Minato) — Client operations and partnerships for the Japanese market. Japan is our primary deployment region.

Sydney, Australia — International headquarters. We are building enterprise and government solutions for the Australian and New Zealand markets.

Ho Chi Minh City, Vietnam — Engineering headquarters. The Agentfront.ai platform and cloud infrastructure are built and maintained here. We are preparing Vietnamese government pilot programs.

Our Japan deployments are delivered through our partnership with Flyout Planning Corporation, a Tokyo-based company specializing in AI and event technology for government and tourism.

---

Impact Summary

---

What Is Next

2026: Expanding municipal AI deployments to additional Japanese cities. Launching Vietnamese government pilots. Adding multi-modal capabilities — agents that can read physical documents via camera and explain them in real time.

2027: Packaging Agentfront.ai for self-service enterprise deployment. Agent-to-agent coordination for complex multi-step workflows. Australian government deployments.

---

About EFFOMA

EFFOMA is a creative technology company building AI agents and extended reality solutions, with offices in Ho Chi Minh City, Sydney, and Tokyo (Minato). Our conversational AI platform, Agentfront.ai, powers production deployments across government, tourism, and healthcare in Asia-Pacific, with Claude by Anthropic as the core reasoning engine.

We are members of the NVIDIA Inception Program, Microsoft for Startups Founders Hub, and AWS Activate.

Website: effoma.com Technology: effoma.com/tech Contact: [email protected]