Skip to content
NTPU AI Learn

GoAI 101 · Module 1

Government AI: From Generative AI to Agentic AI

What can a government agency actually use AI for, and what is not a good fit?

Start Section 1
  • Duration60 min
  • LevelCore
  • PrerequisiteNo hard prerequisite; prior hands-on experience with generative AI helps
0 / 5 completed
Sync my progressJoin a session to keep progress across devices and share it with your instructor

Info

After you join, your completed lesson numbers are sent to your instructor together with your learner code, so they can see your progress. Use the codes your instructor gave you, and do not enter your name, company secrets or other personal data. Progress is kept with the rest of that session's data.

Learning objectives

  • Distinguish Generative AI, Chat AI, Agentic AI, and AI Agent.
  • Explain the role of Model, Context, Knowledge, Memory, Tools, and Workflow.
  • Classify AI's role in government work as Assist, Recommend, or Execute.
  • Identify the Human–AI–Human division of responsibility.
  • Judge whether a government work scenario is a suitable AI use case.

Judge: foundational concepts

12 min

From Generative AI to Agentic AI

Before a government agency adopts AI, the first question is not "is this tool easy to use" but "is this work actually suited to AI." This module builds shared vocabulary for that judgment: Generative AI (produces content), Chat AI (conversational interaction), Agentic AI (capable of autonomous multi-step planning and execution), and AI Agent (an execution entity that can use tools, connect to systems, and adapt its actions to context).

  • Generative AI: produces text, summaries, or drafts from input; it does not act autonomously on its own.
  • Chat AI: provides Q&A and assistance through a conversational interface, still driven turn by turn by a human.
  • Agentic AI: can break a goal into multiple steps, plan an execution order, and adjust the next step when needed.
  • AI Agent: an execution entity built on Agentic AI that actually calls tools and queries or writes to systems.
  • AI vs. traditional information systems: a traditional system produces a deterministic result from fixed logic; AI's output carries uncertainty and needs human verification and governance, it cannot be treated as deterministic program logic.

Info

These four terms form a capability spectrum, not four mutually exclusive product categories: the same system may both generate content and hold a conversation. What matters is how far it is actually allowed to go in a given scenario.
Check yourself

An agency wants to let AI directly approve a subsidy application that requires statutory eligibility review. Which judgment is most correct?

An agency wants to let AI directly approve a subsidy application that requires statutory eligibility review. Which judgment is most correct?

Key takeaway: Judging "what role should AI play in this work" matters more than "is this tool easy to use"; the four terms describe a capability spectrum, not mutually exclusive categories.

Core components and role tiers

12 min

The six components of an AI system

Any AI system, whether a simple Chat AI or a full Agentic AI, is built from six components. Understanding them lets you discuss concretely "what does this system actually know and what can it do," instead of treating AI as one opaque black box.

The six components of an AI system
ComponentDescriptionGovernment example
ModelThe underlying model that understands input and produces outputA caseworker uses an agency-approved model to summarize documents
ContextThe background information AI should draw on for this taskThis complaint case's content and known handling status
KnowledgeKnowledge sources AI can query, e.g. regulations, handbooksRelevant regulation text, internal operating procedures
MemoryInformation retained across conversations or tasksThe case number, and facts already confirmed earlier
ToolsExternal capabilities AI can call, e.g. a lookup system or calculatorQuerying a document-tracking system's status (read-only)
WorkflowHow multiple steps connect, and where human intervention is neededThe full process from intake to formal reply for a complaint case

Info

All six matter: a system with a Model but no Context cannot answer a real question specific to the agency; a system with Tools but no Workflow design is prone to acting where it should not.
Check yourself

A caseworker wants AI, when replying to a complaint, to recall that "a preliminary explanation was already sent for this case last week." Which of the six components does this mainly involve?

A caseworker wants AI, when replying to a complaint, to recall that "a preliminary explanation was already sent for this case last week." Which of the six components does this mainly involve?

Key takeaway: Model, Context, Knowledge, Memory, Tools, and Workflow together determine what an AI system knows and can do; none of the six is optional.

12 min

AI role tiers and common government application patterns

AI role tiers: Assist / Recommend / Execute
RoleDescriptionHuman responsibility
AssistAI provides a draft, summary, or options, not a conclusionA human decides whether and how to adopt it
RecommendAI proposes a ranked suggestion or judgment based on dataA human reviews the basis and makes the final decision
ExecuteAI actually performs an action within a clearly authorized boundaryA human sets the boundary in advance and the action remains traceable and reversible
  • Common public-sector application patterns: document summarization and comparison, case classification, draft writing, meeting-note organization, assisted information lookup.
  • Most of these patterns sit at Assist or Recommend; very few move into Execute, and only after rigorous assessment.
  • The same application pattern may map to a different role tier depending on the agency and the risk level involved.

Info

A role tier is not a one-time label: the same use case may launch at Assist, be observed for a while, and only later be considered for Recommend once enough evidence accumulates. Module 5 covers the detailed control levels.
Check yourself

A system automatically generates a classification suggestion and priority ranking for complaints based on their content, but the caseworker still confirms the final classification. What role is this?

A system automatically generates a classification suggestion and priority ranking for complaints based on their content, but the caseworker still confirms the final classification. What role is this?

Key takeaway: Most government AI application patterns should start at Assist or Recommend; moving further to Execute needs a case-by-case assessment, not a one-time decision.

Government scenarios and responsibility boundaries

14 min

Three government scenarios, and one that does not fit

Government scenarios: three cases

(1) Summarizing and comparing official documents or policy papers: AI helps summarize long documents and flag differences; the caseworker still confirms nothing important was missed. (2) Classifying citizen complaints and drafting replies: AI classifies and drafts a reply; the formal reply must be reviewed by the caseworker before it is sent. (3) Meeting minutes and action-item tracking: AI turns a transcript into action items; the owner and deadline are still confirmed by the responsible staff member.

Warning

A high-risk example that is not a good fit: using AI to automatically approve an administrative decision that requires statutory eligibility review or discretionary judgment. Such decisions carry legal responsibility and due-process requirements, AI may only play an Assist role by providing reference material; it cannot replace review and sign-off by a legally authorized officer.

Info

Human–AI–Human: a human defines the task and the judgment criteria, AI helps produce a draft or recommendation, and a human verifies the result and remains accountable. Every exercise in this course uses simulated scenarios only, never real agency or personal data.
  • Three questions to judge use-case fit: Does this work have a clear standard of correctness or quality? How severe is a wrong output, and can it be caught in time? Is there already someone responsible for final review?
  • Only when you can answer all three is it worth designing a full workflow; if you cannot, stay at a plain Assist level and observe first.
Check yourself

Which of the following best fits the condition for being "worth developing further into an AI use case"?

Which of the following best fits the condition for being "worth developing further into an AI use case"?

Key takeaway: The three common scenarios show how AI can help government work, but each one keeps a human review step; the high-risk scenario is a reminder: role tiering must never be skipped for the sake of "efficiency."

Practice and application

10 min

Practice: Should AI Do This?

Activity 1 asks you to make a full suitability judgment for a government work scenario, and to start filling in the Government AI Use Case Card that this course builds on throughout. This card is referenced again and again: Module 2 turns it into a workflow, Module 3 decides what context it needs, and so on.

  1. Pick or receive a government work scenario
    It may be a real type of work from your own unit, but use simulated data in the exercise.
  2. Record what AI may and may not do
    List the allowed actions and the prohibited actions separately.
  3. Note the human owner and the initial risk
    Who reviews it? What happens if the AI output is wrong?
  4. Judge whether it is worth designing further
    Write one sentence on why this case is (or is not) worth developing into a full workflow.
  • Government AI Use Case Card minimum fields: Use Case Name, Business Goal, Current Process, Primary User, AI Role, Input, Expected Output, Human Owner, AI Allowed Actions, AI Prohibited Actions, Risk Level, Reason to Use AI, Success Indicator.
  • Never enter a real national ID number, real citizen personal data, or sensitive agency information in this exercise, use simulated or de-identified data only.
Check yourself

When filling in the Government AI Use Case Card, what is the main purpose of the "AI Prohibited Actions" field?

When filling in the Government AI Use Case Card, what is the main purpose of the "AI Prohibited Actions" field?

Key takeaway: This module's output is not just a concept, it's a Use Case Card that follows you all the way to Module 8; the earlier you make this judgment clear, the less likely the later workflow, context, and governance design will go off track.

Save as PDF

What next?