Press Release｜Government is becoming AI-powered! Kaohsiung National Taxation Bureau's "AI Little Helper": How can a manual search that takes a day be reduced to 5 minutes?

This article was originally reported and written by Matthew Chen, Senior Editor at Business Next (數位時代), and first published on the Business Next website. Headquarter.ai is authorized to share this translated version for reference.

Photo credit: Business Next

For more than 30 years, Peng-You Weng, Director of the Kaohsiung National Taxation Bureau (left in header image), has handled business tax matters. But recently, he noticed something unprecedented: tax applications and inquiry letters submitted by citizens had reached a level of professional sophistication his bureau had never encountered. The cases became significantly more difficult for staff to process.

He was puzzled. After further investigation, he learned that people had long been using AI tools such as ChatGPT to assist with the application process. Weng Peiyou called this situation an "unequal war," saying, "In this situation, if our colleagues do not have a legal, compliant, controllable, and verifiable AI tool, it is equivalent to fighting a war without weapons."

Equipping colleagues with AI tools is not just about fighting an "AI vs. AI" battle, but also carries a sense of survival for public agencies. Tax cases are complex, and with the emergence of various new business models, more and more discussions are needed. Furthermore, according to a report by *The Reporter*, civil servants have long faced a dilemma of labor shortages and overwork. "The number of cases has surged, but manpower is chronically insufficient. Sometimes, the public uses AI and receives incorrect information, requiring more time to correct, which puts enormous pressure on colleagues," said Weng Pei-yu. Adopting AI has become a necessary solution.

A forced transformation compressed a day's work into 5 minutes.

"This transformation is already 'inevitable'." Chang Wen-hsi (right in the first picture), director of the Financial Information Center, the information coordination unit of the Ministry of Finance, said: "In addition to the shortage of staff, the more long-term problem (in government agencies) is the knowledge transfer. If the professional knowledge and practical experience of senior colleagues in the tax field cannot be effectively transferred, and newcomers have to face the sharp questions raised by the public using AI, it will be a great risk to the system."

Thus, the Ministry of Finance's AI solution, "Business Tax Assistant," was born. The staff only needs to input the business tax questions into the system, and the AI can automatically compare multiple tax laws and data sources, compressing the traditional manual inquiry process, which may take up to a day, into a few minutes.

For example, you can input it in a very natural language way:

“Ms. A was reported for livestreaming sales at XX Night Market, attracting long queues. She also organizes group purchases via LINE and conducts irregular meet-ups for product delivery. Her monthly sales exceed NT$80,000. Should she register for business tax?”

Senior staff can readily recall the necessary reference to tax laws and even similar past cases. However, new staff need to spend a lot of time thinking of keywords, searching on different search engines, reading the results one by one, and taking notes as they go, which is very time-consuming.

The Ministry of Finance’s AI solution, the “Business Tax Assistant,” provides reference materials alongside each answer, enabling officers to make the final determination.

The business tax assistant will first analyze the problem, generate keywords based on the problem, then search for information, and finally provide a reference answer. Most importantly, they will attach the cited data so that the person in charge can make the final judgment.

"The most direct benefit is that new staff no longer need to spend half a day organizing data. AI can provide preliminary answers in just 5 to 10 minutes, allowing everyone to finally focus their energy on the areas that truly require judgment," said Weng Pei-yu. "The feedback from colleagues has also confirmed that although this system has just gone online, the direction is correct. However, the assistant is definitely only for assistance and reference; colleagues still need to develop basic professional skills."

The AI Behind the System: Headquarter.ai, Invested by Cherubic Ventures and Chris Lin

The system was built through a collaboration between the Financial Data Center and Chunghwa Telecom’s cloud team, with the AI application layer developed by Taiwan-based startup Headquarter.ai (通徹智慧)—a company funded by Cherubic Ventures (心元資本) and KKBOX co-founder Chris Lin (林冠羣). Its clients include the Taiwan External Trade Development Council (TAITRA) and well-known job banks.

Phote credit: Business Next - CEO of Headquarter.ai Chien Chang Huang

Headquarter.ai specializes in enterprise AI solutions and focuses on a technique known as Agentic Workflow, which decomposes real-world tasks and trains large language models to execute structured reasoning under human-defined constraints.

“We break down one complex query into about a hundred smaller steps,” said CEO Chien-Chang Huang. “Human experts guide the model on what constitutes a high-quality result. Even subtle differences between terms like ‘expert’ and ‘master’ can produce different outcomes, so rigorous tuning is essential.”

If someone directly asks an LLM a tax question, they might get the correct answer—but often receive vague or inaccurate responses.

“That’s why we conducted more than one hundred rounds of user interviews,” Huang said. “We had to understand the real workflow, what happens after a case arrives, what the final deliverable looks like, and—critically—what information can or cannot be surfaced due to confidentiality. Many precedents contain trade secrets or personal data, so a careful data-masking process is required.”

Could AI become a "killer application" in government?

For Director Chang, the biggest surprise after launch was not the efficiency improvements, but the extremely low adoption barrier.

“In the past, when rolling out new systems, staff often resisted due to new rules or commands they had to learn,” he said. “AI is different—officers can interact in natural language. The learning curve almost disappears.”

Will AI eventually replace government workers? Weng is candid:

“As long as a human signature is required, people cannot be replaced. AI cannot assume legal responsibility.”However, he emphasized AI’s immense value. “Modern public servants face far more complex workloads than before. Those who do not learn to use AI will see their productivity challenged. This is not only a public-sector issue—it is a universal trend across all industries.”

And this trend carries a warning for the private sector: if even government agencies—traditionally cautious, non-profit-oriented institutions—now view AI adoption as a “must-do transformation,” for efficiency-driven enterprises, AI is no longer an optional upgrade. It is a strategic decision that could determine long-term survival.