PMGM7023 Chap.4 Human Judgment and Agent Execution
Human Judgment and Agent Execution
Eight layers of work, and where they went
Performed manually, one analysis passes through eight layers: defining the analysis, setting up the project, importing and profiling, reconstructing context the export lost, cleaning and constructing variables, writing the pipeline, verifying each result, and building the deliverable.
Each rerun repeats several of those layers, so iteration is slow, which is the real reason analyses receive fewer checks than they deserve.
Delegating those layers does not remove responsibility; it concentrates it.
The lecture describes the arrangement as a control design with four positions: you specify, the system executes, you challenge, and you decide.
The gate people skip
Specifying feels like work and gets done. Deciding is unavoidable and gets done. Challenging is the position that feels redundant precisely when the output is fluent, complete and well formatted.
Challenging means asking what the unit of analysis was, how many rows entered and how many survived, which exclusions applied and under what rule, whether a second reasonable definition moves the headline, and whether any claim in the summary rests on a step that actually ran.
None of those questions requires technical skill, and all of them require having decided in advance what the answer should look like.
A specification that can be checked
A brief is judged by one property: could a second person tell whether the finished work met it?
Four parts make that possible. The objective fixes the outcome, the unit and the comparison. The boundary fixes what may be read and changed, which is what makes delegation safe at all. The checkpoint forces a pause before the expensive step, while a wrong assumption is still cheap.
The acceptance test states how you will confirm the result, and it is the part most often omitted, without which plausibility becomes the standard of proof.
What stays scarce when execution becomes cheap
The lecture ends on an economic reading of the same boundary. Work where the inputs and the test of success can be written down, repeated and verified becomes abundant once execution is cheap.
Work where objectives are contested, cases are ambiguous, or somebody must remain accountable does not, and it becomes more valuable rather than less.
The published exposure estimates the course cites make the same point at the level of occupations, and the distinction worth carrying is that exposure measures how much of a job a system could touch rather than how much will disappear.
A whole workflow leaves a person only when the inputs and outputs are fully digital, the procedure is standardised, quality can be tested cheaply and objectively, and little context, trust or accountability remains. When any one of those fails, a person stays inside the system.
What this chapter covers
- 01
Eight manual layers and what delegating them changes
- 02
Specify, execute, challenge and decide as four control positions
- 03
The four parts of a brief someone else could verify
- 04
Codifiable execution against ambiguous, accountable work
Turning a request into a checkable brief
- 2State the objective with outcome, unit and comparison.
- 2State the boundary on what may be read and written.
- 2State a checkpoint before the expensive step.
- 2State an acceptance test a second person could apply.
Key terms
- Acceptance Test
- An acceptance test is a stated check that decides whether finished work met its brief. It is written before the work so that plausibility does not become the standard of proof.
- Human in the Loop
- Keeping a person inside the loop is a control arrangement: somebody specifies the work, challenges the output and owns the decision. It describes active gates rather than passive observation.
- Row Count Audit
- A row count audit compares how many records entered a process with how many reached the result, and accounts for the difference. It is the cheapest check on any delegated analysis.
- Codifiable Work
- Codifiable work is work where the inputs and the test of success can be written down, repeated and verified. It is the part of an analysis that delegates cleanly.
Human Judgment and Agent Execution FAQ
What should a manager specify before delegating an analysis?
Four things. The objective, naming the outcome, the unit and the comparison. The boundary, naming which files may be read and changed and where output goes. A checkpoint before the expensive step, so a wrong assumption surfaces while it is still cheap. And an acceptance test describing how the finished result will be confirmed, which is what stops a plausible answer being treated as a verified one.
An analysis comes back clean and confident. What should I ask?
Ask how many records entered and how many reached the final table, with the difference accounted for. Ask which records were excluded and under which rule. Ask whether a second reasonable definition of the outcome changes the headline figure. A result that survives only one definition is not yet a finding, and none of those three questions requires technical skill to pose.
Exam move
Write a four-part brief for every analysis you run this term, including the small ones, until the structure becomes automatic. Keep a short list of the challenge questions and apply them to any finished output before accepting it. Practise stating what a result cannot establish in the same sentence that reports it.
Working through Human Judgment and Agent Execution in PMGM7023? Sia is AskSia’s AI Management tutor — ask any PMGM7023 Human Judgment and Agent Execution question and get a clear, step-by-step explanation grounded in how PMGM7023 is taught and assessed. Read this chapter free, then take your hardest questions to Sia.
