The High-ROI Use Case Filter

Use case selection is where AI budgets go to die. Not because teams pick bad ideas, but because they pick impressive ones — moonshots that demo well and never survive contact with the org chart.

We score every candidate against five questions. Anything under four out of five does not get built.

01The five questions

• ▸Frequency — does this happen at least weekly? Automating a quarterly task saves theater, not time.
• ▸Friction — does someone visibly hate doing it? Pain is adoption fuel. If nobody minds the task, nobody adopts the fix.
• ▸Fallibility tolerance — can a human catch a mistake cheaply? Drafting an email: yes. Wiring money: no. Start where errors are survivable.
• ▸Footprint — does the input data already exist digitally? If the project starts with "first we digitize…", that is a different (earlier) project.
• ▸Finish line — can you count the result? Hours saved, tickets deflected, days shortened. If you can't measure it, you can't defend it.

02What scores well

The unglamorous middle of the org chart is where the ROI lives:

• ▸Triage: routing tickets, leads, candidates, and invoices to the right person with context attached.
• ▸Drafting: first versions of replies, summaries, proposals, and reports that a human finishes.
• ▸Retrieval: answers with citations from the company's own documents and history.
• ▸Reconciliation: matching records across systems that don't talk to each other.
• ▸Watching: flagging anomalies, risks, and aging items before a human would have noticed.

03Scoring in practice

Three real-shaped candidates, scored honestly:

• ▸Invoice matching (5/5). Daily, hated, human-checkable, fully digital, countable in hours saved. Build it first.
• ▸"Executive strategy chatbot" (1/5). Used quarterly at best, no unit, no finish line. It will demo beautifully and die quietly.
• ▸Automated contract negotiation (2/5). High friction, but errors are expensive and hard to catch, and half the inputs are relationship context that lives in nobody's system. Bench it; revisit after a drafting wedge earns trust.

04What to reject

• ▸The everything-chatbot. A bot with no workflow is a demo with a login page.
• ▸The moonshot. If it requires the org to change before the tool works, the org will win.
• ▸The zombie process. Some workflows shouldn't be automated — they should be deleted. Automating waste just makes waste faster.