The Use-Case Discovery Canvas
Nobody needs another list of AI use cases. What a board needs is the method that produces the list: take a workflow your staff actually run, write down the steps as they perform them today, and tag each step with what AI is really good at. The candidates fall out on their own — "if it's really good at routine work, then here's where that lands in our shop." Load a worked example to see the whole move, then run it on a workflow you oversee.
Five things it's really good at
These don't change from workflow to workflow — they're the same five capability buckets from the capability-patterns session. Everything on this page hangs off them.
Click a bucket to trace it — the steps you've tagged with it light up below.
Load a worked example — or start your own
Three ordinary insurance workflows, walked and tagged properly. Load one, see how the logic runs, then swap in a workflow from your own shop. The examples are illustrations; the method is the product.
The workflow, as staff run it today
Not the flowchart from the procedures manual — what a person actually does. Each step starts with a verb and names an action a human takes. Then tag it: which of the five buckets fit that step?
Pick something a real person is paid to do, start to finish.
How to run this with a board
35 minutes, projected. The whole session rides on one reframe: they are not here to receive use cases, they are here to learn the machine that finds them.
- Frame it (2 min).
Say"We are not going to hand you a list of AI use cases. Lists go stale and they're always someone else's. We're going to hand you the machine that makes the list — and it runs on workflows you already oversee."
- Put up the five buckets (3 min). One sentence each, no more. Remind the room they've already met these five — same taxonomy as the capability-patterns session. New vocabulary here would be a bug, not a feature.
- Walk one worked example (10 min). Load "A claim, from first notice of
loss to settlement." Read each step aloud and ask the room: what is the person
actually doing here? Then tag it together and let them watch candidates appear in
the right-hand panel as the tags land.
Say, when the panel fills"Notice nobody brainstormed a use case just now. We walked the steps, and the logic did the work. If it's really good at routine work — and it is — then the settlement letter drafts itself onto this list."
- The room's turn (15 min). Hit "Start your own workflow" and ask for one a director actually oversees: preparing the board pack, the audit committee's review cycle, a rate filing, onboarding a new broker. Type the steps as they call them out. Two rules: every step names a person's action, and "the system does it" is not a step — ask who touches it and what they do.
- Close on the cheapest pilot (5 min). Read the pilot nudge aloud,
then hand the framework over.
Say"This canvas goes home with you. The list on the screen is the least valuable thing here — the method is the deliverable. Run it on any workflow you oversee, and it will keep producing candidates long after today."
Traps. Someone proposes "AI should just handle the whole claim" — bring them back to one step: "Which step first?" Someone debates whether AI can really do a step — tag it anyway; deciding that is what a pilot is for. Steps come out as system behaviour — re-anchor on the human: this framework only works on what people do.
Saved in this browser only — nothing leaves the page. The examples are deliberately generic: swap in your own lines of business and the method doesn't change.