Write a one-page use-case brief with a baseline, boundaries, evidence requirements, and a clear next decision.
Start with a recurring problem
A useful use case names a specific job someone already performs. Begin with the delay, repeated effort, or quality problem that matters.
The current process is your starting point for comparison. Record who does the work, which inputs arrive, and what a finished result looks like.
A bounded task makes a pilot easier to understand. “Draft a reply from approved notes” is clearer than “transform customer service.”
Compare simpler alternatives
Fixed rules often fit stable decisions with explicit conditions. For example, a form can reject an empty required field without generating language.
AI assistance may fit tasks involving varied language or incomplete structure. However, uncertainty in the output creates additional review work.
A manual option remains a legitimate baseline. Anthropic recommends starting simply and adding agent complexity only when the task justifies it.
Check whether success is observable
A useful output must be checkable by someone with the necessary evidence. If nobody can judge correctness, an impressive demonstration provides weak assurance.
Input availability matters before model choice. A missing policy, unclear ownership, or unreliable record can defeat an otherwise capable system.
A safe boundary limits what the pilot can change. Prefer a reviewable draft when an external action is unnecessary for learning.
Compare three fictional proposals
A fictional small team wants to reduce time spent on event administration. It considers form validation, attendee-question drafts, and automatic refund decisions.
The first proposal has explicit rules and needs no language generation. The second can be tested using approved event notes and human review.
The third proposal changes money and depends on exceptions. It requires a separate authority design and stronger evidence before any operational trial.
| Task | Simpler starting design | Evidence needed | Next decision |
|---|---|---|---|
| Check required form fields | Deterministic validation | Defined field rules | Implement ordinary form checks |
| Draft answers to event questions | AI draft with source review | Current event notes and sample questions | Run a bounded draft-only pilot |
| Issue refunds automatically | Manual decision with recorded policy | Authority, exceptions, safeguards, and recovery | Defer autonomous action |
Define a pilot that teaches you something
A pilot question states the uncertainty you want to resolve. Ask whether reviewable drafts reduce total work while meeting the same quality standard.
An exclusion list keeps the trial focused. Specify which inputs, users, actions, and sensitive information are outside its scope.
A stopping condition prevents a pilot from drifting into routine use. Name the owner and the evidence required for the next decision.
Practice: write the use-case brief
Your exercise describes one recurring task in six fields: problem, user, inputs, output, baseline, and review boundary.
Your comparison must include a non-AI alternative. Then identify one missing fact that could change your preferred approach.
A strong brief connects an observable benefit to a bounded experiment. Avoid calling a task suitable merely because a model can attempt it.
Compare with the worked answer
Worked answer: For the event team, test draft answers against the existing manual process. Use approved notes, keep publication under human control, and measure review effort. Keep form validation deterministic. Defer automatic refunds until authority, exception handling, and recovery are defined.
Use the method on your own task
Editable worksheet: Download the Markdown template and open it in a text editor. Use fictional or approved information; no account or paid service is required.
Download the worksheet (.md) ↓What to carry forward
Write a one-page use-case brief with a baseline, boundaries, evidence requirements, and a clear next decision.
Show your reasoning with the exercise above. Keep the evidence, the decision, and the remaining uncertainty together.