Government AI Copilot Case Study: Sentiment Reports from 4 Hours to 30 Minutes

Adoption Background and Requirements Inventory

The agency's starting point for adopting Government AI Copilot was not technical evaluation, but a requirements inventory checklist. Prior to procurement, staff defined four things: what to monitor, report layout, alert triggers/recipients, and permission access rules, establishing a solid basis for configuration.

• Monitoring Scope：Official agency full name and abbreviations, executive names and titles, current-year key policies, and colloquial terms used by the public and media. Official policy names often differ from public terminology; setting only official names will miss critical discussions.
• Report Format：Retains current approval format field orders, topic classifications, and reporting hierarchies. System outputs align directly with existing layouts, eliminating the need for staff to reformat and allowing executives to keep their reading habits.
• Alert Rules：Defines what constitutes an anomaly (rapid spike in negative volume, concentrated discussion on specific policies) and designates notification targets. Without predefined rules, alerts either miss critical events or become overwhelming spam.
• Permission Planning：Lists visible topic scopes and report subscription lists per department based on operational division. The value of cross-departmental sharing lies in a shared factual foundation, which does not mean opening all content to everyone.

Procurement began only after completing the four inventories. InfoMiner is included in the Joint Supply Contract, with data servers located within Taiwan and holding ISO 27001 certification, allowing security review documents to be fully prepared during procurement.

Before: Four Hours of Staff Time Every Day

Before adoption, staff spent every morning manually browsing dozens of news sites and social platforms, copy-pasting relevant coverage, judging each item positive or negative, formatting briefings, then routing them through approval — routine sentiment reports took about 4 hours. This meant executives always saw "the morning's sentiment," while afternoon events often didn't reach decision-makers until the next day.

A Day Before Adoption

• After arriving at work, opening dozens of news websites and social platforms one by one to browse daily content.
• Copying and pasting relevant reports into document files, tagging sentiment and responsible business units item by item.
• Reformatting content into the agency's customary briefing layout, adding summaries and recommended watchpoints.
• Submitting to supervisors for hierarchical sign-off, repeating the reorganization if revisions or supplemental materials are requested.
• Briefings reached executives only in the afternoon, excluding afternoon developments entirely.

A bigger issue was cross-departmental: departments produced separate clippings in varying formats, presenting conflicting views of the same event, making cross-comparison difficult and preventing cumulative, traceable context.

This workflow also carried three hidden costs: inconsistent assessment criteria made it hard for executives to weigh issue severity; reports scattered across past documents required manual file-by-file review for historical tracking; and heavy reliance on specific staff created operational gaps during absences.

From Joint Supply Contract Procurement to Launch: What Happened in Two Weeks

The key to launching within two weeks was not rushing, but the complete absence of hardware deployment. Delivered as a cloud service with data servers located within Taiwan, the agency needed no server purchases or server room setups; combined with direct ordering through the Joint Supply Contract without separate tenders, the two most time-consuming steps were eliminated.

Among the five phases, the agency only needed to invest manpower in requirements interviews and the trial comparison run, with the vendor handling the rest. Monitoring sources already cover over 500,000 channels without requiring agency integration or maintenance, allowing efforts to focus on defining clear judgment standards.

After: The AI Copilot's Daily Workflow

The agency adopted InfoMiner through the Joint Supply Contract, taking approximately two weeks from requirements confirmation to launch. The daily workflow corresponds to the perception, understanding, and tracking stages of the AI Copilot 5-stage model:

• Sense and understand：The system monitors news media, PTT, Dcard, Facebook, and other sources 24/7, automatically completing Chinese sentiment analysis and topic classification. Collection and analysis finish before dawn, so staff arrive at work to a fully classified sentiment overview.
• Daily intelligence briefing：Daily sentiment reports are generated automatically: key topics, volume and sentiment shifts, main arguments, and relevant agencies. Staff need only verify and add assessments, reducing routine reporting time to ~30 minutes. The briefing arrives in their inbox before work, ready for review and internal context addition before submission.
• Anomaly alerts：When negative sentiment spikes abnormally, the system immediately alerts relevant decision-makers within 24 hours, so critical issues are not delayed until the next day's report. Overnight incidents are also monitored, sending notifications directly to designated recipients.
• Cross-department sharing：All departments view standardized analysis results on a single platform with traceable event history, eliminating reliance on separate news clippings for cross-department communication. Policy tracking can directly access the historical record for any issue.

Notably, the AI did not replace staff interpretation — it replaced copy-pasting and formatting. Human review remains a required step before reports are submitted, which is the right way for government agencies to use generative analysis tools.

Quantified Results

Three replicable practices

What is replicable in this case is not the procurement decision itself, but three practices during adoption. None require additional budget, yet they determine whether the system is actively utilized or neglected after launch.

1. Keyword taxonomy co-designed by staff and vendor, reviewed quarterly

The keyword taxonomy determines whether the system captures key insights and is the most direct factor affecting results. Vendors understand monitoring and exclusion logic, but only agency staff know which policies are sensitive or what colloquial terms the public uses. The agency co-designed an initial taxonomy with the vendor and reviews it quarterly: adding seasonal priorities and new terms, and removing closed issues. Without reviews, the taxonomy stays frozen at launch.

2. Maintain a comparison period to align ranking rules with agency judgment

Priorities identified by the system will not always match agency perspectives. The agency established a trial comparison period prior to launch where automated daily reports ran alongside manual workflows. Staff reviewed topic rankings daily, highlighting over- or under-ranked items to tune algorithms. This comparison period also built trust: only after confirming no critical issues were missed did staff feel confident submitting automated reports for sign-off.

3. Clearly define the boundary between humans and AI

The agency clarified boundaries at the start: AI replaces copy-pasting and formatting, while evaluation and approval remain human. The system collects, classifies, and displays, while staff verify facts, add internal context, and remain accountable for content. This boundary solves two problems: executives still receive reports verified by humans, and staff shift focus to professional analysis. Embedding human-AI division of labor into operating procedures passes internal scrutiny far better than claiming full automation.

From Daily Sentiment Briefings to an AI Governance Copilot

Automated briefings and real-time alerts are the foundation of the government AI copilot's five stages — sense, understand, assess, recommend, track. On the same data foundation, agencies can further enable incident assessment reports, council interpellation preparation and task tracking, and combine the RAGi agency knowledge base so AI recommendations fit the agency's own policy context.

The role of InfoMiner Government AI Copilot is to integrate external real-time sentiment with internal agency knowledge, assisting executives and staff in mastering situations, assessing risks, formulating decisions, and tracking implementation.

See the full InfoMiner Government AI Copilot solution →