Profile / Case study
AI Voice Agent for Ho Chi Minh City Tax
Ho Chi Minh City Tax launched an AI call centre on 8 August 2026 that handles about 90% of routine taxpayer enquiries. I worked on the voice agent as Senior Business Analyst and Consultant Lead at FPT Smart Cloud.
At a glance
- Client
- Ho Chi Minh City Tax
- Scale
- 350,000+ enterprises, 650,000+ household businesses, 8 million+ individual taxpayers
- Solution
- AI call centre: website chatbot and phone voice agent, 24 hours a day
- Technology
- FPT.AI platform, under the FPT and Ho Chi Minh City Tax cooperation agreement of November 2025
- My role
- Senior Business Analyst and Consultant Lead, voice agent
- Launch
- 8 August 2026
The situation
Ho Chi Minh City Tax serves the largest taxpayer base in Vietnam: more than 350,000 enterprises, more than 650,000 household businesses and more than 8 million individual taxpayers. Three problems had built up at that scale.
- Overloaded phone lines. Routine questions (where is my tax office, who handles my file, how much do I owe, am I under an exit ban) competed with complex cases for the same officers.
- Notices that never arrived. Taxpayers who had changed address or contact details missed debt and exit-ban notices.
- Impersonation fraud. Scammers posing as tax officers by phone made taxpayers distrust genuine calls, which made legitimate outreach harder.
A policy change raised the pressure. From 1 January 2026, household businesses moved from presumptive tax to self-declaration, so several hundred thousand small taxpayers in the city needed guidance on obligations that were new to them. In November 2025, FPT and Ho Chi Minh City Tax signed a cooperation agreement to apply AI agents on the FPT.AI platform to taxpayer support.
The solution
The AI call centre pairs a chatbot on the tax authority's website with a voice agent on the phone, and both run 24 hours a day. The voice agent works in two directions: it calls taxpayers with reminders and notices, and it answers the enquiries they call in with.
A typical outbound call runs like this:
- A caller ID the taxpayer can checkThe call arrives under the registered caller ID "Thue TP.HCM", so the taxpayer can see who is calling before answering.
- Details only the tax authority holdsThe agent states the taxpayer's name, tax ID and the amount owed. A scammer would not have these details, which is what lets the taxpayer trust the call.
- The obligation, explainedThe agent explains the obligation and answers follow-up questions about debt, exit-ban status, office addresses and the officer or department responsible.
- Payment only in eTax MobileThe agent never asks for a transfer or a card number.
- Handover and recordAnything outside the routine set is routed to a tax officer. The call is recorded and transcribed for review.
Notices also reach taxpayers through apps they already use: Momo, ZaloPay, VNPay, Vietcombank and ACB.
The work is phased so that the highest-volume, most rule-bound calls went first.
| Phase | Scope |
|---|---|
| 1 | Tax debt reminders and exit-ban notices |
| 2 | Reminders for declaration, filing and annual settlement |
| 3 | Inbound enquiries and more complex cases |
The head of Ho Chi Minh City Tax described the AI as a support tool for officers: officers keep the cases that need judgment.
My role and contributions
My job was to turn the tax authority's organisation and rules into something an AI agent could follow, and to prove it worked before taxpayers heard it.
Call routing design
Ho Chi Minh City Tax runs 29 branches and 16 departments, and a caller rarely knows which one owns their question. I designed the inbound routing model that maps a caller's need to the right unit, with a two-level intent taxonomy (broad topic, then specific request) and fallback logic for when the agent is unsure. I documented it as routing diagrams, user journeys and call scripts.
Outbound scenarios
I scoped the reminder scenarios the agent runs: tax debt with follow-up questions, declaration obligations and filing deadlines, including when the agent hands over to an officer.
Spoken Vietnamese
Tax content is full of amounts, dates, tax IDs and abbreviations that a text-to-speech engine reads badly by default. I wrote the normalisation script that converts them into natural spoken Vietnamese, so a debt figure is read as an amount and a tax ID as digits.
Configuration control
The agent's configuration went through four versions as scope grew. I audited and merged those versions so that one agreed configuration went to test.
Requirements and acceptance
I wrote the business requirements document and built the user acceptance test workbook of 137 test cases covering routing, scripts, fallbacks and handover.
Leadership reporting
During the live pilot I prepared the leadership report: what the pilot showed, the open risks and the decisions needed to proceed.
Results
- A verifiable official channel. A registered caller ID, identity details stated at the start of the call and payment only through eTax Mobile give taxpayers a way to tell a genuine call from a scam.
- Recognised inside FPT. The Voice Agents for HCMC Tax team received a commendation from FPT's AI Product Division on 10 August 2026, and I received an individual commendation on 26 August 2026.
The launch figures come from the tax authority's public statements, not from my own measurement.
Sources
Public reporting, in Vietnamese. The launch coverage describes the system; the November 2025 article covers the cooperation agreement with FPT.
- Ho Chi Minh City Tax launches AI call centre and digital payment platform Nhan Dan
- HCMC Tax launches AI call centre answering 90% of taxpayer questions Sai Gon Giai Phong
- HCMC Tax's AI virtual assistant phones taxpayers with debt reminders VOV
- AI automatically calls with tax reminders and exit-ban notices Thanh Nien
- Ho Chi Minh City Tax pioneers AI to support taxpayers Thoi bao Tai chinh Viet Nam
- FPT partners with HCMC Tax to apply AI in taxpayer support Bao Dau tu