Start with operating evidence
Ask “what repeats?” before “what can ChatGPT do?”
- Where does someone copy information from one tool to another?
- What gets rewritten from the same source material?
- Which queues need classification or routing?
- Which recurring reports require the same preparation?
- Where does someone research the same kind of entity repeatedly?
Create a workflow inventory
Capture the workflow, owner, frequency, minutes, systems, source data, output, exception rate and consequence of error. That inventory becomes the raw material for prioritization.
Score before you build
Assess repetition, data availability, tolerance for error, review cost, current burden and implementation dependencies. The goal is to eliminate weak ideas early.
Validate with a small pilot
Choose one bounded stage, measure the pre-AI baseline, run a controlled pilot, and compare time, quality and exception handling. Evidence beats enthusiasm.
Frequently asked questions
Who should be involved in finding AI opportunities?
Include the people who perform the work, the workflow owner, someone who understands the systems/data, and the person accountable for the outcome.
How many workflows should I assess at once?
Start with 10–20 candidates, measure a smaller shortlist, then deeply assess the top three to five.
Should employees self-report time spent?
Self-reported estimates are useful for discovery but should be validated with samples, logs or direct observation before using them for investment decisions.