How to Run AI Use Case Discovery: A Four-Week Playbook for Finding AI Opportunities That Survive the Pilot

What AI use case discovery is

AI use case discovery is the structured work of finding where AI could create measurable value by examining how work actually happens, not by brainstorming what AI could do. It produces a shortlist of framed, evidence-backed candidate use cases, each tied to a real process, its data, and a baseline. It comes before prioritization and long before a pilot.

It is the first stage of the AI use case identification process: discover, frame on a canvas, test with a framework, score on an impact–feasibility matrix. This article covers the discovery stage only, because it is the one most teams skip and the one that decides whether the later stages rank anything real.

Discovery versus the things it gets confused with

• AI use case discovery — How work happens today, from the people doing it Output: A shortlist of framed candidates with baselines Where it goes wrong: Treated as a workshop instead of fieldwork
• AI ideation workshop — What AI could do Output: A long list of ideas Where it goes wrong: Ideas nobody can size or verify
• Process discovery — A process, end to end Output: A validated current-state map Where it goes wrong: Stops at the map; never asks where AI fits
• AI readiness assessment — The organization's data, skills and governance Output: A maturity score Where it goes wrong: Says whether you could do AI, not what to do
• Vendor-led demo — The vendor's product Output: A use case that fits the product Where it goes wrong: Solves the vendor's problem

Discovery borrows from process discovery and feeds the canvas. If you already have validated current-state maps, discovery is mostly done; if you only have an ideas list, it has not started.

The four-week plan

• 1. Scope — Pick three to five processes where cost, volume or error rate is visible Who: Sponsor, ops lead Output: A scoped list with a baseline number for each
• 2. Elicit — Interview the people who do the work; collect the documents and screenshots they actually use Who: Facilitator plus 2–4 people per process Output: Interview notes, exceptions list, undocumented rules
• 3. Frame — Write one canvas per candidate; mark which boxes are validated Who: Process owner plus facilitator Output: 8–15 canvases in a register
• 4. Validate and shortlist — Play the canvases back to the people interviewed; fix what they reject; score Who: Same participants, then the steering group Output: A shortlist with impact and feasibility scores

Week 2 is where discovery is won or lost. Most teams compress it to a single workshop; the step-by-step guide to generative AI use cases puts current-state mapping at step five, after framing, which is fine when maps exist and fatal when they do not. Interview first.

The twelve discovery interview questions

Ask these of the person who does the work, not their manager, in this order. The first six find the use case; the last six find the reasons it will fail.

1. Walk me through the last one you did, start to finish, including anything you did outside the system.
2. What do you check before you can start, and where do you check it?
3. Which step takes longest, and what are you waiting on during it?
4. What decision do you make that a new hire gets wrong?
5. What do you do when the standard path doesn't apply? How often?
6. What do you re-key, copy, or look up in a second place?
7. Where is the rule for that written down? If nowhere, who else knows it?
8. What would a wrong answer cost, and who would notice?
9. Which of these steps changed in the last year, and did the documentation change with it?
10. What would you refuse to let a system do without you looking first?
11. Who else touches this before it is finished?
12. If this step disappeared tomorrow, what would you do with the time?

Question 4 produces more good use cases than any brainstorm. Question 7 produces the blocker list: the undocumented rules an AI would need and cannot get from a system export. Question 10 writes the human-in-the-loop box on the AI use case canvas for you.

When this applies, and when it doesn't

Run discovery when the AI programme is at the "which use cases" stage; when an ideas list exists but nobody can size the ideas; when a pilot stalled because the process turned out to be different from the diagram; or when a COO or AI strategy lead has been asked for a portfolio rather than a project.

Skip it, or shorten it to week 1 only, when the use case is already fixed by an external requirement (a regulator, a contract, a platform migration), when the process is fully instrumented and a current-state map exists and was validated in the last six months, or when you are evaluating a single vendor product for a single team and the decision is buy or don't.

Discovery does not replace an AI readiness assessment; it tells you what to be ready for. And it is not process mining. Mining reads event logs from systems and shows you paths and frequencies; discovery elicits the parts of the work that never hit a log, which is where most of the answers to questions 4 through 7 live. The AI agent readiness framework explains why those undocumented parts are what agents trip on.

Doing week 2 without forty interviews

The honest constraint is facilitator time. Three to five processes, three people each, an hour per interview, plus write-up: two to three weeks of one person's time, which is why it gets cut to a workshop. ClearWork's automated discovery exists for that week. Clarity, its discovery agent, runs the interviews asynchronously with each person, ingests the documents and recordings they already have, spots where two people describe the same step differently, and lists what is still unknown. Every finding waits for a human to confirm it before it enters the record, and the output is a source-linked current-state map and requirements set that fill the canvas's "how the work happens today" and "data and knowledge" boxes. It does not choose your use cases; it makes sure the ones you choose are built on how the work is actually done.

Frequently asked questions

How long should AI use case discovery take?

Four weeks for three to five processes with one facilitator, most of it in week 2. With asynchronous interviews and document ingestion the elicitation week compresses; the framing and validation weeks should not, because they are where the people doing the work correct the record.

Who should lead it?

Someone who owns operations or transformation, not the AI team alone. The AI team knows what is buildable; discovery is about what is worth building, and that requires access to the people doing the work and enough credibility that they answer question 5 truthfully.

What is the output of discovery?

A register of eight to fifteen framed canvases, each with a baseline, a validated description of how the work happens today, a list of undocumented rules the AI would depend on, and a first-pass impact and feasibility score. Not a roadmap; that comes after prioritization.

How is this different from process discovery?

Process discovery produces a validated current-state map of a process. AI use case discovery uses that kind of evidence to find where AI fits and writes it up as candidate use cases. If you have recent validated maps, most of week 2 is already done.

Can we do discovery with a survey instead of interviews?

Surveys find complaints; interviews find decisions. A survey can scope week 1 (where is the pain), but questions 4 through 7 need a conversation, because the answers are things people do not know they know until they walk through the last case.

Run the four-week discovery plan and bring validated use cases to your next AI steering meeting.

Discovery is the stage that decides whether the matrix ranks anything real. If week 2 is the bottleneck, see how ClearWork's automated discovery runs the interviews and document review and holds every finding for human validation. Start a free 14-day trial and run it on the first process in your scope list.

‍