Match generative AI to business work — decision exercise
Exercise: evaluate a park maintenance assistant
Original fictional case. No paid services or cloud account required. Difficulty: beginner · Estimated duration: 15 minutes
A park authority receives volunteer voice reports about trail hazards. It wants a daily, route-specific maintenance brief with traceable hazard details. Reviewed incident records and permitted crew assignment preferences are available. Supervisors verify hazards before allocating work. A separate request needs an exact count of open incidents from a reliable table.
| Need | Required result |
|---|---|
| Volunteer recordings | Written hazard summaries linked to the correct route |
| Incident records | Findings about recurring hazard locations |
| Crew preferences | Relevant draft work briefs for assigned routes |
| Open incidents | Exact count from the incident table |
Your tasks
- Separate summarization, analysis and personalization. State each input and output; identify where audio support or transcription is needed.
- Propose a small experiment that measures hazard completeness and route-association accuracy. Preserve supervisor verification.
- Decide whether the exact-count request needs generation and explain the alternative.
Reference solution
Audio reports require audio-to-text capability or transcription before text summarization. Analysis turns incident records into findings about recurring locations. Personalization combines reviewed findings with permitted crew/route preferences to draft relevant text; it does not grant access to unrelated records. Evaluate whether important hazards survive the summaries and whether every detail remains linked to the correct route. Supervisors verify hazards before allocating work. A conventional query can compute the exact open-incident count; fluent generated prose is unnecessary for that task. Check the selected model’s documented directions and evaluate it on representative cases rather than assuming one model performs every step.
Rubric (6 house points)
Two points each for: distinct inputs and outputs across all three jobs; completeness/route accuracy with required verification; a justified deterministic exact-count approach.