Finding Good Use Cases

Where AI fits well

Good candidates share traits: the task is frequent and repetitive, involves language or patterns in data, has a clear way to check quality, and mistakes are cheap to catch and fix (a human reviews the draft). Examples: drafting replies, summarising meetings, classifying tickets, extracting fields from documents, searching internal knowledge, code assistance. Poorer candidates are one-off tasks, decisions where an error is severe and hard to detect, and anything where no one can say what "correct" means.

A use-case canvas

Fill one in per idea. It forces the questions that decide whether a pilot is worth starting.

Problem        Support agents spend 3 min per ticket finding policy text
Users          40 support agents, 1,200 tickets/day
AI role        draft a reply from policy + ticket; agent edits and sends
Success metric handle time -25%, customer satisfaction not lower than today
Data           policy wiki (clean), 2 years of resolved tickets (needs PII removal)
Risk           wrong policy quoted -> human review mandatory; no auto-send
Owner          Head of Customer Care

Prefer "assist" over "replace" first

Drafts reviewed by a person are safer, faster to ship and easier to measure than fully automatic decisions. Move towards automation only after the data shows the quality is consistently high.

Quick check: Which task is a strong first AI use case?

• Drafting routine support replies that a human reviews
• Fully automatic final approval of large loans
• A one-time unrepeatable decision
• A task with no way to check quality