Start where work already piles up
The best first candidates announce themselves as backlogs. A queue that never empties, a report that is always late, an inbox that needs triage before anyone can act — each is a signal that demand outstrips the manual capacity behind it.
Walk the operation and list the places where work waits. A step with a persistent queue is more promising than a step someone imagines could be clever, because the value is already measurable as the delay you are trying to remove.
Follow the copy-paste trail
Watch how information moves between systems. Every time a person reads from one screen and types into another, reformats a document to fit the next tool, or re-keys the same customer detail, there is a candidate for structured extraction and controlled write-back.
These handoffs are easy to miss because they feel like normal work. Ask the team to narrate a case out loud; the moments where they switch windows are usually the moments worth automating.
- Reading from one system and typing into another
- Reformatting documents to fit the next tool
- Re-entering the same detail in several places
- Answering the same lookup question repeatedly
Prefer decisions with reachable evidence
A workflow is only automatable to the degree the system can reach the information a good decision needs. Before committing to a candidate, list the policies, records, and documents involved, and confirm each one can actually be retrieved at the moment of the decision.
Candidates whose evidence lives in accessible systems move quickly. Candidates that depend on knowledge held only in someone's head, or in a tool nobody can integrate, are worth deferring until that access exists.
Score candidates before you commit
Once you have a shortlist, rank it. Weigh how often the work repeats, how bounded the variation is, whether the evidence is reachable, and how reversible a mistake would be. A frequent, bounded, evidence-rich, low-blast-radius task is the safest place to start.
This scoring keeps the first project honest. It steers you toward a workflow that will prove value and build trust, rather than the most impressive demo that quietly depends on data you cannot yet reach.
Treat discovery as risk work, not just upside
Opportunity discovery is also where you decide what could go wrong and who is accountable when it does. Naming the failure modes early — missing evidence, low confidence, an unusual case — turns a vague idea into a workflow you can actually operate.
Recognized frameworks for managing AI risk exist precisely so this step is not improvised. Using one keeps discovery grounded in governance from the start, rather than bolting controls on after the build.