MODULE 2 · LESSON 1
Free — no login requiredSign in to track progress, save quiz attempts and enrol in the full course.
Sign in to track progress / enrolA Method for Finding Use Cases
The four signals
A promising AI candidate almost always shows all four. Missing one is a warning; missing two means stop.
Signal 3 is the one that most often turns out to be absent, and it is the one nobody checks before getting excited. A task can be high-volume, repetitive and low-risk, and still be unusable because the decision was never recorded in a field anybody can query.
When that happens, the finding is valuable rather than disappointing: the first project is to start recording it. That costs little and produces an asset in months.
The session
Ninety minutes, four to six people who actually do the work, and deliberately no senior stakeholders in the room — their presence changes what people admit to.
Round 1: the tedium round (25 minutes). Ask each person: what did you do last week that a reasonably bright new joiner could have done after two days of training, and that you did more than ten times?
Phrase it exactly like that. It bypasses the two failure modes of the usual question. "Where could we use AI?" gets you what people have read about. "What's difficult?" gets you the genuinely hard things, which are the worst candidates. What you want is the dull things, and people will not volunteer those unless you specifically ask.
Round 2: the friction round (25 minutes). What do you wait for? What do you re-key from one system into another? What do you have to chase? What do you check twice because you do not trust the first source?
Re-keying between systems is a particularly strong signal — it usually means an extraction or matching problem, and it is measurable.
Round 3: the wish round (20 minutes). If you had ten more hours a week and no extra headcount, what would you do that you currently cannot? This is where new capability candidates come from, and they will not appear in the first two rounds because people do not think of absent work as work.
Round 4: the veto round (20 minutes). Go through what has emerged and ask, for each: what would go wrong if a system did this and got it wrong one time in twenty? This kills the unsuitable ones early and, more usefully, tells you what the exception path must look like for the survivors.
The paper trail sources
Sessions surface what people remember. Records surface what actually happens, and the two lists differ.
- Complaint and escalation logs. Categorised complaints point straight at repeated failures. Where the same category recurs, either a process or a judgement is failing consistently.
- Rework and correction records. Anything reversed, amended or re-done. High rework means a judgement being made badly or inconsistently.
- Where the queue backs up. The step where work waits is either under-resourced or too slow. Both are candidates, though the fix is not always AI.
- Overtime and seasonal peaks. Work that requires temporary staff at quarter-end is by definition high-volume, repetitive and time-bound.
- The handover documents. What someone writes down when going on leave is a precise description of their repeatable work — the parts that can be explained are the parts that can potentially be automated.
Three bad sources
The vendor's suggestion. Vendors propose the use case their product serves. That is not dishonest, it is simply not an assessment of your operation. Use vendor conversations to learn what is possible, never to select what to do.
The competitor's press release. You are seeing an announcement, not a result. Companies publicise launches, not quiet retirements eighteen months later. Announcements are also chosen for how they sound to investors.
The executive's weekend reading. The most difficult to handle politically and the most common. The productive response is not refusal but translation: "That's interesting — the version of that which would work here is probably X, because we have the data for it. Shall I test that?" You have kept the sponsorship and replaced the project.
A 200-person specialist insurance broker ran this session with five people from operations and client service.
The tedium round produced: re-typing policy details from insurer PDFs into the broking system; categorising inbound emails; checking renewal dates against a spreadsheet; producing the same monthly client report with different numbers; chasing insurers for quote confirmations.
The friction round produced: waiting for underwriter responses with no visibility of where a submission sat; re-keying the same client data into three insurer portals; double-checking premium calculations because the spreadsheet had been wrong before.
The wish round produced something nobody had raised in ten years of planning: we only review a client's cover against their changed circumstances at renewal, because that is all we have time for. If we could review continuously, we would spot under-insurance and cross-sell gaps all year.
The veto round eliminated two candidates immediately. Automated premium calculation was rejected — an error goes straight to a client and is a regulated matter. Automatic email categorisation was kept, but with the requirement that anything from a named list of large clients bypasses it entirely.
The final shortlist, in the order they chose:
- Extracting policy details from insurer PDFs. All four signals present, and the recorded outcome existed because the typed-in values were in the broking system — every historical PDF had its answer key already attached. This was the strongest candidate and nobody had proposed it before the session, because it was too boring to think of as a project.
- Continuous cover review. New capability, higher value, harder. Deliberately second.
- Email categorisation. Useful, modest, with a clear exception rule.
Two observations. The winner came from the tedium round, which is why that round is phrased so specifically. And the highest-value idea came from the wish round from a person who had never been asked the question — it had simply never occurred to anyone that the constraint was removable.
A candidate task is high-volume, involves repeated judgement, and errors are recoverable — but the decision that was made has never been recorded in any queryable field. What should the manager conclude?
The four signals
Click to flipVolume, repetition of judgement, a recorded outcome, and tolerable error. All four should be present; the recorded outcome is the one most often missing and least often checked.
Click to flip backFind candidates from the people doing the work, using four rounds that deliberately surface tedium, friction, absent work and risk — then cross-check against complaint logs, rework records and handover documents. Test every candidate against four signals, and check the recorded outcome first, because it is the one most often missing. Never let a vendor, a competitor's announcement or an executive's reading choose the project for you.