從宏觀視野到日常辦公桌

前兩篇貼文與大家分享了2026年INAHTA國際大會的宏觀視野與政策洞見，今天我想切換視角，把鏡頭拉回「凡間」，談談作為一名身處第一線的醫療科技評估（HTA）研究員，面對AI狂潮真實的體會、掙扎與期盼。在大會上，各國代表熱烈討論著AI治理與合成數據的挑戰，但在我們的日常辦公桌前，AI的距離似乎既近又遙遠。

我們深知AI的重要性，但在缺乏由下而上（bottom-up）的實作規劃下，基層研究員往往感到無所適從。這並非AI本身的問題，而是介於理想政策與實際工作之間的巨大落差。

AI狂潮下的現實困境

時間壓力與手工勞動的悖論

在臺灣，我們經常面臨極度壓縮的專案時程，例如在短短42天內就必須產出一份具備高度科學嚴謹性的HTA報告（mini HTA或rapid review）。在這分秒必爭的壓力下，我們卻必須將大量珍貴的時間，耗費在手動搜尋疾病背景、指引，甚至要逐字輸入關鍵字，只為了爬梳國內相近臨床治療地位的產品。

基礎數位化的缺失

在連國內藥物許可證資料庫都缺乏完善且整合的搜尋系統的情況下，我們卻要跳躍式地去大談AI導入。這不禁讓人想問：「連基礎的『數位化』都還沒有做好，究竟要怎麼去考慮『AI如何輔助』？」

這種缺乏系統性介面輔助的純手工勞動，在本次INAHTA大會中，被國際專家直白地稱為「愚蠢工作（Stupid works）」。身為受過訓練的研究員，我們最大的挫折感往往來自於無法將精力集中在最具智力挑戰的核心分析上，而是被困在這些無可避免的底層勞動中。

由上而下的政策困局

安全防堵與實作需求的失衡

當我們期盼AI能成為解放雙手的輔助工具時，機構由上而下（top-down）的傳達卻往往只停留在資安防堵與單向的風險警告。長官們憂心機密外洩與暴露於未知風險，這絕對合理且必要。

禁止政策帶來的隱形風險

但如果只有一味的禁止，卻沒有去探究員工完成工作所需的真實需求，也沒有提供受控的內部安全環境與實務指引，基層人員不但無從將AI納入工作流程，更可能迫使有需求的研究員轉向依賴個人私有AI工具，反而讓機構面臨更大的隱形風險。

這種只有防堵卻缺乏建設性導入設計的現況，正是全球許多HTA機構共同面臨的治理盲區。

AI真正落地的願景

工作流程的解構與任務邊界

AI真正要落地，需要的不是雲端上的空泛討論，而是接地氣的「工作流程解構」。誠如國際頂尖機構所分享的成功經驗，我們應該將既定的HTA報告架構拆解，找出邊界明確的任務（Bounded tasks）交給AI處理前置作業。

AI應該扮演的角色

我們並不奢望，也絕對不贊成讓AI來做最終的系統性文獻回顧（SLR）與核心價值判斷。相反地，我們殷切期盼AI能幫忙完成從零到一的背景資料梳理。

驗證工作與人類判斷的價值

有了AI的輔助，驗證（validation）的工作絕對不會減少，甚至不一定能立即省下多少時間，但這絕對值得，因為這能釋放我們「不可思議的大腦（incredible brains）」，讓我們將所有的心力與專注，投注在最需要人類批判性思考的科學實證評估上。

寫在最後：建立安全試錯空間

科技的浪潮已經拍打在我們的辦公桌上。期盼未來的HTA機構與決策者們，能真正走入凡間傾聽實作者的聲音，建立容許內部交流與試錯的安全空間（safe space）。

唯有正視這些繁雜且重複的底層勞動，導入受控的工作流程與數位輔助，我們才能在緊縮的時間與資源下，堅守HTA的專業品質，並真正找回身為研究員的科學尊嚴。

大家在各自的專業領域中，又是如何面對機構的AI政策與實際工作需求的落差呢？歡迎在底下留言與我交流！

English version below

From Macro Vision to Your Daily Desk

My previous two posts shared the macro perspectives and policy insights from the 2026 INAHTA International Conference. Today, I want to switch angles and focus on the ground level—sharing my honest account as a frontline Health Technology Assessment (HTA) researcher experiencing the AI wave’s real challenges, struggles, and hopes. During the conference, delegates enthusiastically discussed AI governance and synthetic data challenges. Yet at our desks, AI feels both near and far away.

We understand AI’s importance, but without bottom-up implementation planning, frontline researchers often feel lost. This isn’t a problem with AI itself—it’s the vast gap between ideal policies and actual work.

The Reality Beneath the AI Hype

The Paradox of Time Pressure and Manual Labor

In Taiwan, we often face extremely compressed project timelines. For instance, we must produce a highly scientifically rigorous HTA report—whether mini HTA or rapid review—within just 42 days. Under this time crunch, we spend enormous amounts of precious time on manual searches for disease backgrounds, guidelines, and even typing keywords one by one to map out domestically equivalent clinical treatment options.

The Missing Foundation of Digitalization

When even our domestic drug licensing database lacks a comprehensive, integrated search system, how can we jump straight to discussing AI integration? This raises an obvious question: “How can we consider ‘how AI can help’ when we haven’t even properly digitalized the basics?”

This pure manual labor without systematic interface support was bluntly termed “stupid works” by international experts at the INAHTA conference. As trained researchers, our greatest frustration comes from being unable to focus our energy on the most intellectually challenging core analysis. Instead, we’re trapped in unavoidable low-level labor.

The Top-Down Policy Dilemma

The Imbalance Between Security and Implementation Needs

When we hope AI becomes a tool to free our hands, institutional top-down communications focus only on cybersecurity safeguards and one-way risk warnings. Senior management’s concerns about data breaches and unknown risks are absolutely reasonable and necessary.

The Hidden Risks of Pure Prohibition

However, if there’s only blanket prohibition without exploring employees’ actual work needs, without providing a controlled internal safe environment and practical guidelines, frontline staff cannot integrate AI into workflows. Worse, it may force researchers with genuine needs to rely on personal private AI tools, exposing the institution to even greater hidden risks.

This governance blind spot—prevention-only without constructive implementation design—is shared by many HTA institutions globally.

The Vision for Genuine AI Implementation

Deconstructing Workflows and Defining Task Boundaries

Real AI implementation doesn’t require cloud-level abstract discussions. It needs grounded “workflow deconstruction.” Following the successful experiences shared by leading international institutions, we should deconstruct the established HTA report structure and identify bounded tasks that AI can handle for preliminary work.

The Right Role for AI

We neither hope for nor endorse having AI conduct final systematic literature reviews (SLR) and core value judgments. Instead, we earnestly hope AI can help complete the foundational data mapping work from scratch.

The Irreplaceable Value of Validation and Human Judgment

With AI support, validation work won’t decrease and might not immediately save time. But it’s absolutely worth it—because it liberates our “incredible brains” so we can devote all our energy and focus to the scientific evidence assessment where human critical thinking is most needed.

A Final Thought: Creating Safe Spaces for Trial and Error

The wave of technology has already reached our desks. I hope future HTA institutions and decision-makers will truly listen to those implementing the work on the ground, creating safe spaces for internal dialogue and experimentation.

Only by facing up to these complex and repetitive low-level labors, introducing controlled workflows and digital assistance, can we maintain HTA’s professional quality under tight time and resource constraints. Only then can we truly reclaim our dignity as researchers.

How are you navigating the gap between your institution’s AI policies and your actual work needs in your own field? I’d love to hear your thoughts in the comments below!