Generative AI and Language Model Foundations · Lesson 4

Selecting Tasks and Building Human-Verified Workflows

8 min read

Lesson objectives

Evaluate task suitability using consequence, verifiability, reversibility, data sensitivity, and grounding.
Design a human review point that can materially change the outcome before action.
Build a simple workflow from task definition through verification and approval.

Choose the task before optimizing the prompt

Before asking “How do I write a better prompt?” ask “Is this an appropriate task for this kind of system?” A fast screen looks at five dimensions: consequence of error, verifiability, reversibility, data sensitivity, and availability of a source or success criterion. A low-consequence task that can be reviewed and reversed, such as an internal draft, tolerates more experimentation. A high-consequence task that is hard to verify or will act automatically on people, money, or permissions needs stronger controls and may not be appropriate for generative automation at all.

Screening dimension	Practical question	If the answer is uncomfortable
Consequence of error	What harm follows from a wrong output?	Reduce system authority or add qualified review
Verifiability	Is there a fact, rule, or test to compare against?	Do not make the output the final decision
Reversibility	Can the result be undone easily before anyone is affected?	Stop automatic execution before approval
Data sensitivity	Does the input contain confidential or personal information?	Minimize data and use an approved channel or do not send it
Grounding	Does the system have the sources needed for this case?	Retrieve the right source or state the gap

A human-verified workflow

A general workflow to adapt: 1) define the task and success criterion; 2) prepare the minimum necessary data and context; 3) request an output in a checkable form; 4) verify claims and structure against the source or test; 5) label errors and uncertainty rather than hiding them; 6) have the authorized person make the decision or approve the action when the result has consequences; 7) record the prompt version, source, and verification result when repeatability or auditability matters.

“Human in the loop” does not mean merely placing an approval button on the screen. Review is useful when the person has the time, information, and authority to reject or change the output. If a system presents one hundred decisions in a minute and someone is expected to click “approve all,” human presence is nominal. Put the review point where it can still change the outcome: before external sending, before updating a sensitive record, before publishing public information, or before a decision that depends on an evidence-sensitive claim.

Exercise

Exercise 1 — 12 minutes: rate four fictional tasks on the table dimensions from 1 (low) to 3 (high): draft an internal heading; summarize a policy supplied in full; propose a reply to a complaint that will be reviewed before sending; and make a final automatic decision on an application with financial consequences. Do not only total numbers; write the decisive constraint that changes each workflow design.

Check your answer

Guidance: the internal heading is low consequence and reversible; policy summarization needs source checking; the complaint reply needs tone and fact review before sending; the final financial decision is high consequence and may be unsuitable for delegation to a model, especially if criteria, appeal rights, or data quality are unclear. A score helps comparison but does not replace judgment about the decisive constraint.

Lesson summary

Task selection comes before prompt optimization. Increase controls as error consequences rise, verification becomes harder, or data becomes sensitive. Put human review where a person can reject or change the output before it becomes an action.