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
Learning objectives
Distinguish Generative AI, Agentic AI, and AI Agent, and explain what each means in government work.
Judge whether a government task can use AI, based on business purpose, error consequence, and suitability.
Classify the role of AI as Assist, Recommend, or Execute, and choose an appropriate level.
State permitted actions, prohibited actions, and the accountable person to complete the AI Use Judgment Card.
Judgment basics: concepts and suitability
9 min
Three terms: Generative AI, Agentic AI, AI Agent
Before a government agency adopts AI, the first question is not "is this tool any good?" but "how far should AI go on this task?" To answer it, we start with three terms that give everyone a shared language. They describe a spectrum of capability, not three mutually exclusive product categories.
The three terms and a government example of each
Term
What it can do
Government example
Generative AI
Produces text, summaries, or drafts from input; does not act on its own
Condenses a long interpretation letter into a one-page summary, which the caseworker verifies
Agentic AI
Breaks a goal into steps, plans the order, and adjusts the next step based on results
On receiving a petition, plans its own steps: classify, look up rules, draft a reply
AI Agent
An executor that actually calls tools and reads from or writes to systems
Connects to the case system to check progress, or even updates case status directly
The further along the spectrum, the closer AI gets to actually acting, and the greater the impact when it is wrong.
AI output is uncertain; it is not the deterministic result of a traditional information system running fixed logic, so it needs human verification and governance.
What matters is not the product name but how far AI is actually permitted to go in this scenario.
Info
One system may be able to generate content, plan steps, and call tools. Being capable does not mean everything should be switched on; deciding where to stop is the agency's judgment.
Check yourself
A system can connect to the case system, look up petition progress, and directly change a case status to "closed". Which term fits best, and what is the key risk?
Key takeaway: The three terms form a spectrum of capability: the closer AI gets to actually acting, the more it needs defined permission boundaries and human control.
9 min
Business purpose, error consequence, and suitability
To judge whether a government task suits AI, start with three questions. What is the business purpose, and is there a quality standard you can check? How serious is the consequence if AI is wrong, and can it be caught early? Is someone already responsible for the final review? If you cannot answer all three, do not design further.
How far AI can go, by error consequence (simulated scenarios)
Task
Business purpose
Error consequence
Suitability judgment
Turn meeting minutes into an action list
Save organizing time
Minor: omissions can be caught during human confirmation
Suitable: Assist is enough
Classify petitions and draft replies
Shorten triage and first-draft time
Moderate: misclassification delays handling; an improper reply damages trust
Suitable, but every case needs caseworker review
Decide eligibility for a subsidy
Determine by law whether the applicant qualifies
Severe: affects citizens' rights and involves statutory authority and due process
AI may only organize material; the decision belongs to the officer with statutory authority
Running case for this course (simulated)
A city government's Civil Affairs Bureau (a simulated unit) receives many citizen petitions every month. Caseworkers must classify each one, look up the relevant rules, and draft a reply. The bureau is considering AI assistance. Later modules reuse this case to build every card and table.
Warning
AI cannot replace statutory authority. Decisions involving legal eligibility, discretion, or citizens' rights must be reviewed and signed off by the officer with that authority; "efficiency" is never a reason to skip it.
Check yourself
Which of the following best meets the conditions for being worth designing further as an AI use?
Key takeaway: Check business purpose, error consequence, and whether someone reviews before deciding AI can be used; the more serious the consequence of error, the fewer steps AI should take.
Role levels and responsibility boundaries
9 min
Three roles: Assist, Recommend, Execute
Once a task is judged suitable for AI, the next decision is what role AI plays in it. A role is not a property of the tool; it is an authorization decision the agency makes for this application.
AI role levels and human responsibility
Role
What AI does
Human responsibility
Petition example
Assist
Provides drafts, summaries, or options; not a conclusion
Decides whether and how to use it
Summarizes a petition into three key lines
Recommend
Proposes a classification or ranking based on data, with its basis
Reviews the basis and makes the final decision
Suggests a category and handling priority; the caseworker confirms
Execute
Carries out an action inside a boundary authorized in advance
Sets the boundary beforehand; keeps actions traceable and reversible
Automatically routes the case to a named section and sends an acknowledgement
Most government AI applications sit at Assist or Recommend; only low-risk, clearly bounded, rigorously evaluated steps are considered for Execute.
One application can use different roles at different steps: summarizing is Assist, classifying is Recommend, and only sending the acknowledgement is Execute.
A role is not fixed forever: start at Assist, gather evidence, then evaluate whether to raise it.
Check yourself
A system generates a classification suggestion from a petition and ranks handling priority, explaining its basis, but the caseworker confirms the official classification. Which role is this?
Key takeaway: A role is the agency's authorization decision: most applications start at Assist or Recommend, and moving up to Execute requires case-by-case evaluation.
9 min
Permitted actions, prohibited actions, and the accountable person
After choosing a role, write it as a concrete action boundary. Saying "AI only assists" is not enough: list what AI may do, what it may not do, and who is responsible when it goes wrong.
Petition AI application: example action boundary (simulated)
Item
Content
AI role
Recommend: suggest a classification and draft a reply
Permitted actions
Summarize the petition, suggest a category, search public rules, draft a reply
Prohibited actions
Send a reply without caseworker review, change a case to closed, decide statutory eligibility, handle personal data without authorization
Main risks
Misclassification causes delay; citing outdated rules; improper reply tone or content; personal data leakage
Accountable person
The Civil Affairs Bureau petition caseworker (reviews and signs off every reply); the section chief owns the authorized scope of the application
Info
Human–AI–Human: a person defines the task and the standard, AI helps produce a draft or recommendation, and a person reviews the result and remains responsible. "A human looked at it" is not enough; say who reviews, what they check, and that they can reject.
A prohibited action is usually not something technology cannot do; it is something the business or governance does not allow.
The accountable person must be as specific as a job title or role, never "relevant staff" or "the system administrator".
These four fields (role, permitted, prohibited, accountable person) are the starting point for every later module.
Check yourself
In an action boundary table, what is the main purpose of the "Prohibited actions" field?
Key takeaway: Responsibility boundaries are only real once the role is written as permitted actions, prohibited actions, and a specific accountable person.
Practice and application
9 min
Practice: complete the AI Use Judgment Card
For the simulated petition case provided in the course, complete the first card: the AI Use Judgment Card. The case background is supplied, so you do not choose a topic; spend your time on judgment and reasoning. The card feeds the first four fields of the integrated Government AI Governance & Acceptance Canvas, and later modules build on it.
1Read the case background Confirm the business purpose, the current handling method, and who the users are.
2Choose the AI role Pick Assist, Recommend, or Execute and write your reasoning, including why not one level higher.
3Write permitted and prohibited actions List at least two of each; prohibited actions must be specific, not just "do not break the law".
4Write the main risks and the accountable person List the most important risks and their consequences, and name a specific person and their review responsibility.
Canvas fields covered: use case name (useCaseName), business goal (businessGoal), AI role (aiRole), accountable person (humanOwner).
Do not enter real national ID numbers, citizens' personal data, sensitive agency information, or any account or password; use simulated or de-identified data only.
When filling in "Accountable person" on the AI Use Judgment Card, which entry is most appropriate?
Key takeaway: The judgment card is the starting point of the integrated Canvas: the sooner role, permitted and prohibited actions, and the accountable person are clear, the less the later workflow, permissions, and acceptance work drifts.
