Skip to content

GoAI 101 · 6h · Module 2

Government AI Workflows and Human Gates

How does AI enter a workflow, and when must work go back to a person?

Start Section 1
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 a Task from a Workflow and break work into steps with clear inputs and outputs.
  • Mark who decides at each step (AI, human, or rule) and recognize Sequential, Parallel, and Routing patterns.
  • Add human approval, escalation, hand-back, completion conditions, and a failure path to a workflow.

From Task to Workflow

10 min

From Task to Workflow: Break Down the Steps Before Adding AI

Module 1 asked how far AI should go on a piece of work. This module goes one level down: which steps make up that work, where AI joins, and where a person must take over. Government work is rarely "ask once, answer once". It is a multi-step workflow: intake, classification, legal lookup, drafting, review, reply.

Task (a single job)

  • Finished in one interaction, such as "summarize this document".
  • One input, one output.
  • Errors are contained and easy to check.

Workflow (a process)

  • Several Tasks chained by order or condition.
  • One step's output is the next step's input.
  • Errors travel downstream, so you must know who owns and checks each step.
  • Input: what must be in hand before the step starts, such as the full complaint text or the case assignment.
  • Output: what the step must hand over when it ends, such as a classification result or a draft reply.
  • Completion condition: what counts as done, such as "the classification was confirmed by the case officer".
  • Test for a good split: can this step be checked on its own? If not, it usually needs splitting further.

Info

Write down the steps, inputs, and outputs first, then decide which step AI takes. Adding AI to an unclear process only automates the confusion.
Check yourself

Which statement best shows that a step has been broken down clearly enough?

Which statement best shows that a step has been broken down clearly enough?

Key takeaway: Split the work into steps with inputs, outputs, completion conditions, and independent checks before deciding where AI joins.

Workflow patterns and decision roles

10 min

Workflow Patterns and Decision Roles: Who Decides the Next Step

Three basic workflow patterns
PatternMeaningGovernment example
SequentialEach step depends on the previous resultClassify, summarize, then draft, one after another
ParallelIndependent steps run at the same time, then mergeLook up relevant regulations and the likely responsible agency together
RoutingConditions send the case down different pathsSend a complaint to a different unit depending on its type

Every point in the flow that decides "which way do we go next" is a Decision Point. A decision point has one of three deciders, and the flow chart must show which:

Decision points and deciders
DeciderSuited forWatch out for
Rule (fixed condition)Clear conditions written in advance, such as "merge duplicate case numbers"Most predictable and easiest to audit
AI (suggestion or judgment)Understanding text, summarizing, or first-pass classificationAI only suggests; a person confirms anything that affects rights
Human (case officer or supervisor)Statutory authority, discretion, or formal external effectName a role; "relevant staff" is not a role

Warning

If AI decides the routing, a wrong call sends the case to the wrong unit. Give such decision points an exit: "low confidence or unclear type goes to a person".
Check yourself

A complaint requires looking up "relevant regulations" and "the likely responsible agency", and neither depends on the other. Which pattern fits best?

A complaint requires looking up "relevant regulations" and "the likely responsible agency", and neither depends on the other. Which pattern fits best?

Key takeaway: Use Sequential for dependent steps, Parallel for independent ones, and Routing for conditional branches; mark every decision point as decided by a rule, AI, or a person.

Case: an unfinished complaint workflow

10 min

Case: An Unfinished Complaint-Handling Workflow

This course uses a simulated teaching scenario: a fictional city government handles citizen complaints and wants AI to help with classification, summaries, and draft replies. Below is the flow the business unit has drawn so far. Find the problems before redrawing anything.

Unfinished complaint workflow (simulated)
  1. Complaint received
  2. AI classifies
  3. AI summarizes
  4. AI looks up laws and responsible agency
  5. AI drafts a reply
  6. Reply sent
  • Gap 1: no person confirms the AI classification, so a wrong category sends the case to the wrong unit.
  • Gap 2: the draft reply is sent directly with no human approval gate, yet an official external reply is a high-impact action.
  • Gap 3: there is no exception path for threatening language, unclear content, or a failed legal lookup.
  • Gap 4: there is no completion condition or failure path, so when AI errs or a lookup fails the flow stalls or silently continues.

Info

Three questions to test completeness: who owns each step, where must a person approve, and what happens when the flow cannot continue?
Check yourself

In this complaint workflow, which gap carries the highest risk and should be fixed first?

In this complaint workflow, which gap carries the highest risk and should be fixed first?

Key takeaway: When reviewing a flow, first check for human approval before high-impact actions, then check that exception paths and failure paths exist.

Human gates and exception paths

10 min

Human Approval, Escalation, and Hand-back

Four ways a person steps into the flow
MechanismWhen to use itComplaint example
ApprovalBefore a high-impact action takes effect, a named role must explicitly agreeThe case officer reviews and approves the draft reply before it is sent
EscalationThe case exceeds the current role or AI scope, so a higher authority decidesThe responsible agency is unclear, so a senior officer decides
Hand-backAI stops handling the whole case and a person takes over completelyThe complaint contains a threat to someone's personal safety
Failure pathA pre-planned response when a step cannot finishLegal lookup fails: mark it incomplete and notify the case officer; never fill it with AI guesses
How to write a completion condition

"The case officer has approved the draft reply, the case status is updated to replied, and an approval record exists" is done. "AI has produced a draft" is only an intermediate output, not completion.

  • A human gate must say who reviews, what they check, and what they can do (approve, edit, return, hand back).
  • "Have someone take a look" is not a gate; with no named role, no checklist, and no record, it does not count.
  • Silently continuing after a failure is the most dangerous design; the flow must state stop, notify, and person takes over.

Warning

Human–AI–Human: a person defines the task, AI assists or executes, and a person reviews and stays accountable. AI cannot replace statutory authority or approve its own output.
Check yourself

A complaint contains language threatening a case officer's personal safety. What is the most appropriate handling in the workflow?

A complaint contains language threatening a case officer's personal safety. What is the most appropriate handling in the workflow?

Key takeaway: A human gate needs a named role, a defined check, and a record; exceptions and failures need a next step written in advance, not improvised.

Practice and application

10 min

Practice: Complete the Flow Chart with Human Gates

Continue with the unfinished complaint workflow from the previous slide. You do not draw from a blank page and you do not build any automation platform. Your job is to check the flow and add human gates and exception paths so it is complete and responsibility is clear. The result goes into the "workflow steps" field of the integrated Canvas.

  1. Mark the division of work
    Label each step as decided by AI, a person, or a rule, and write its input and output.
  2. Add human approval
    Insert an approval gate before each high-impact action; state who reviews, what they check, and what they can do.
  3. Add exception paths
    Write at least one escalation condition and one hand-back condition.
  4. Write the completion condition and failure path
    State what counts as done and what happens next when a step fails.
  • Output: a "flow chart with human gates", entered into the "workflow steps" field of the integrated Canvas.
  • What is assessed: whether the flow is complete and responsibility is clear, not how complex the drawing is.
  • Use only simulated or de-identified data. Do not enter secrets, real personal data, or unauthorized sensitive information.
Check yourself

What is the "failure path" in a flow chart mainly used to record?

What is the "failure path" in a flow chart mainly used to record?

Key takeaway: A flow is complete when a person approves before every high-impact action and every exception or failure has a clear next step.

What next?