Identifying high-value GenAI opportunities — question bank
35 standalone questions on finding the opportunities worth having — visibility bias, the right altitude, the six shapes, the five trails, and honest scoring.
Audience: senior business leaders. No technical background assumed.
Source: the opportunity-identification handout — visibility bias, the six capability shapes, the five trails, the qualifying questions and the scoring framework.
How to use this: every question stands alone. Pick an option, then read the answer. The nine sections run in the order you would actually work: see the bias, get to the right altitude, hunt, qualify, score, then build a portfolio.
A · Step 1 — See the misallocation
Q1Research into enterprise GenAI found that close to half of all budgets go into one area, while the largest returns were found somewhere else. Which way round?
Answer: (2)Roughly 50% of spend to the front office, and the money to be made sitting quietly in the back. That gap is the whole reason opportunity identification is a discipline rather than a brainstorm.
Q2What is the name for the mechanism that produces that gap?
Answer: (2)Naming it is what makes it correctable. Sales and marketing is where AI is visible — to customers, to the board, to the market, in the annual report. Invoice reconciliation is not.
Q3A man searches for his keys under a streetlight. He dropped them in the grass, but "the light is better here." What is the organisational equivalent?
Answer: (2)The budget goes to the lit area and the keys stay in the grass. The diagnostic question is blunt: name your three most-discussed AI initiatives, then your three most expensive repetitive processes. How much overlap is there? For most organisations the answer is none.
Q4Around 95% of GenAI pilots produce no measurable return, while the median among those that did succeed is about 188%. What kind of problem does that distribution describe?
Answer: (3)The shape of the distribution matters more than any single number. A few large wins and many zeros is not a technology verdict; it is evidence that choosing well is where the value is decided. And selection is a leadership job, not a technical one.
Q5Between 60% and 80% of in-flight AI initiatives have no documented baseline and no re-measurement plan. What does that imply?
Answer: (2)Without a baseline there is no way to know whether you are in the 95% or the 5%. Note also the typical payback window for disciplined deployments — month 14 to month 28 — which is far longer than most business cases assume.
B · Step 2 — Get to the right altitude
Q6A leader asks "where can we use AI?" and the answer comes back as "customer service" or "HR" or "legal." What is wrong with that answer?
Answer: (2)You do not automate "being an accountant." You automate "reconciling 400 invoices against purchase orders every Tuesday." The right altitude is the task: something with an input, an output, and a person currently doing it a known number of times per week.
Q7Which of these is stated at the right altitude?
Answer: (3)It has an input (an NDA), an output (a marked-up first pass), a standard to compare against, and — crucially — a countable weekly volume. The other three cannot be baselined, so they cannot be costed or later evaluated.
Q8Why does the wrong altitude reliably produce the missing-baseline problem?
Answer: (2)Every one of the 60–80% with no baseline started at the wrong altitude. The test to apply in a room: pick the function you are keenest on and name one task inside it with a weekly volume. Anyone who cannot do that in a minute has just found their real problem.
C · Step 3 — Recognise the six shapes
Q9Rather than asking "could AI do this?", the better question is "does this task look like one of six recurring shapes." Which set is correct?
Answer: (2)Long in and short out; finding the relevant thing in a large corpus; checking something against a standard; messy input to clean fields; producing a first draft; helping someone solve a problem. Use it as a shopping list, not a taxonomy.
Q10Which two of the six turn out to be the most prevalent in practice — and why does that matter?
Answer: (2)A useful corrective, because it points somewhere unexpected: the bulk of real value is in helping a person do the task, not in removing the person. Programmes designed around replacement are aiming at the smaller half of the opportunity.
Q11"Turning 40 supplier statements into one reconciliation sheet." Which shape?
Answer: (3)The output is structured fields rather than prose. Compare with "60-page report → board summary," which is summarisation, and "expense claims against policy," which is reviewing work.
Q12What is the practical advantage of handing a workshop the six shapes rather than running an open "AI ideation" session?
Answer: (2)Fewer ideas, better ones. Open ideation produces enthusiasm at the wrong altitude; the shapes force people to match a real task against a real capability, with a weekly volume attached.
D · Step 4 — Walk the five trails
Q13Which trail is described as the highest-yield and the least glamorous?
Answer: (3)Anything you currently pay an outside firm to do at volume — data entry, first-line support, document processing, basic content production. The cost baseline already exists in your accounts, so you cannot argue about whether the saving is real; it is a line item.
Q14Why does "follow the complaint" — what your own people grumble about weekly — matter so much?
Answer: (2)This was the insight behind one of the best-documented adoption results in the field: advisers hated that they could reach only a fifth of their own firm's research. The AI removed a friction everyone already resented, so nobody had to be sold on using it.
Q15Which trail is described as the most reliable, and takes an afternoon to walk?
Answer: (2)Every time somebody copies from one screen and pastes into another, that is a task with a structured input, a structured output, and no intelligence in between. It requires no budget, no vendor and no analysis — only an hour of watching.
Q16What do all five trails have in common?
Answer: (2)Opportunity identification is an operations exercise wearing an AI hat. None of the five begins with a model, a vendor or a capability — they begin with where work actually sits, waits, repeats or annoys.
E · Step 5 — Qualify the candidates
Q17A task is irritating, manual, and done eleven times a month. Should it be a GenAI candidate?
Answer: (2)Volume is the first qualifying question. At 40,000 queries a day even modest improvements compound into real money; at eleven times a month nothing does. Irritation is not the same as opportunity.
Q18Your input is already clean rows and columns in a database. What follows?
Answer: (2)GenAI's genuine edge is over unstructured input — language, documents, images, audio. Structured data was well served long before any of this, and the older approach is usually better on all three of cost, accuracy and explainability.
Q19The third qualifying question is "what happens when it is wrong?" Which half of that question matters more?
Answer: (2)Errors you detect are manageable. Errors that look exactly like successes are how organisations end up with a metric that looks fine while the best customer relationships quietly degrade. Detectability, not severity alone, is what determines whether a failure mode is survivable.
Q20What is the single governance rule described as the most valuable in this whole topic?
Answer: (2)If you cannot state today's cost, cycle time or error rate, you cannot prove improvement and you will not get credit for it. Two weeks of measurement before a build has saved more AI programmes than any technology choice.
F · Step 6 — Know where not to go first
Q21Which of these is the worst shape for a first project?
Answer: (2)Infinite input space, brand and legal exposure, and the failure is public. The other three are bounded, internal, and fail quietly enough to be fixed.
Q22"Anything requiring data you do not have or cannot access." Why is that a first-project trap?
Answer: (2)And the sponsor who approved an AI project will spend a year receiving updates about data pipelines. The best-documented failure in the field died on exactly this — never integrated with the records system it needed.
Q23A proposed use case would apply a model to a task where a person could write the rule down in a sentence or two. What is the verdict?
Answer: (2)GST calculation, eligibility thresholds, licence checks, format validation. You would be paying per call, forever, for an answer an IF statement gives you exactly and free — and you would lose the ability to explain the result.
Q24Why is "no named business owner" listed as a disqualifier alongside technical and legal risks?
Answer: (2)The failure is not at build time; it is at the moment the thing has to become part of somebody's operation. An unowned project has nobody whose job gets harder if it stops working — which means nobody's job gets easier if it does.
G · Step 7 — Score honestly
Q25In the scoring framework, which dimension carries the largest weight?
Answer: (2)Value 40%, feasibility 30%, time-to-value 15%, reuse 15% — then a risk penalty of up to 1.0 subtracted from the total. A 5 on value means large, quantified, and landing on a P&L line somebody owns.
Q26Most prioritisation frameworks omit "reuse" entirely. Why does it earn a full 15% here?
Answer: (2)Selection is where a platform is either created or prevented. Scoring reuse is how the platform argument enters the decision at the only moment it can — before anything is built.
Q27Four candidates are scored. A public-website chatbot scores 4 on value but 2 on feasibility with a −0.80 risk penalty, finishing last at 2.30. Call recordings into structured CRM notes finishes first at 4.05. What is the point of the exercise?
Answer: (2)The chatbot is customer-facing, demonstrable and makes a good slide. On honest scoring it comes last, and not narrowly. Run it live and have the room predict the ranking before you reveal it — the gap between the intuitive answer and the scored one is the bias.
Q28The winning candidate won on a dimension nobody would have discussed in a boardroom. Which?
Answer: (3)It scores 5 on reuse. That is the leverage argument arriving through the scoring sheet rather than through an architecture debate — and it is invisible to anyone ranking on value alone.
H · Step 8 — Build a portfolio, not a bet
Q29How many slots does a healthy first portfolio have, and what are they?
Answer: (2)A single project is a bet. Three of the right shape is a strategy. Quick wins buy credibility and teach the organisation how to do this; the platform builder is unexciting and high leverage; the strategic bet is protected from quarterly pressure.
Q30Why include a project you expect to fail?
Answer: (2)Its output is knowledge, and you say so in advance, in writing. Naming one project as a learning project at the outset is the cheapest way to buy an organisation permission to be wrong.
I · Step 9 — Close the hours-saved trap
Q31Nearly every GenAI business case is built on estimated hours saved. In how many ways can those hours actually become money?
Answer: (3)Exactly three, and no others. Make the sponsor name which of the three, in advance, in writing, before the project is approved.
Q32A sponsor says the freed hours will let people "focus on higher-value work." What three things must you ask?
Answer: (2)Without those three answers the hours dissipate — absorbed into the working day, entirely invisible in the P&L. That is the mechanism by which a 188%-ROI case quietly becomes one of the 95%.
Q33A table reports back with a task, a weekly volume, and no baseline. What is the correct facilitator response?
Answer: (2)It is the most common failure and the easiest to be soft about. A baseline built after launch is not a baseline; the "before" no longer exists to be measured.
Q34During the exercise, every table picks something customer-facing. What has just happened?
Answer: (2)The bias does not disappear because it has been named ninety minutes earlier. Seeing the room reproduce it in real time, after being warned, is what makes it stick.
Q35Three questions to take away. Which is not one of them?
Answer: (4)Vendor choice is downstream of every question that matters here. The other three all point at the same discipline: look where the returns are rather than where the light is, and refuse to start anything you cannot later prove worked.