Why the usual method disappoints
The standard approach to AI use-case selection looks rigorous. Run workshops across the business, collect a long list of ideas, score each from one to five on value and feasibility, plot them on a two-by-two grid, and fund the top-right quadrant.
It produces a tidy slide. It also produces three predictable problems.
The value scores are guesses. In a workshop, nobody has measured the baseline: how long the task takes today, how often it is done, what an error costs. A "5 for value" usually reflects the enthusiasm of whoever proposed the idea.
The feasibility scores measure the demo, not the work. "Feasible" tends to mean someone tried the task in a chatbot and the answer looked good. That says little about the full range of real cases, about integration, or about who checks the output.
Each idea is scored alone. Ten use cases that all depend on the same unglamorous foundation, such as clean customer master data or a governed document collection, are evaluated as ten separate bets. The foundation never gets funded, and most of the ten stall.
The result is the pattern described in From AI Experimentation to Enterprise Value: many pilots, few that reach the P&L.
What the evidence says about where AI helps
Two of the most rigorous field studies of generative AI at work give a sharper picture than any vendor survey.
Figure 1
The same tool helps or harms depending on the task
Results from two large field studies of generative AI at work
Consultants · task inside the frontier
- More tasks completed+12.2%
- Faster task completion+25.1%
- Higher quality, blind-rated+40%
Consultants · task outside the frontier
−19 pts
less likely to reach the correct answer than consultants working without AI
Customer support agents · issues resolved per hour
In a field experiment with 758 Boston Consulting Group consultants, researchers led by a Harvard Business School team found that on tasks within the model's capabilities, consultants using GPT-4 completed 12.2% more tasks, worked 25.1% faster and produced work rated more than 40% higher in quality. On a task deliberately chosen to fall outside those capabilities, consultants using AI were 19 percentage points less likely to reach the correct answer than those working without it. The researchers called this boundary the "jagged technological frontier": capability is uneven, and from the outside it is hard to see where the edge lies.
A separate study of 5,179 customer support agents, in its NBER working-paper version (later published in The Quarterly Journal of Economics), found that an AI assistant raised issues resolved per hour by 14% on average and by 34% for novice and low-skilled workers, with little effect on the most experienced agents.
Two practical lessons follow for selection:
- The same tool can help or harm depending on the task. Selection has to happen at task level, and it has to include a test on hard cases that sit near the edge of the model's ability, not just typical ones.
- Gains concentrate where the gap between novice and expert performance is large. Work with high staff turnover, long ramp-up times or wide quality variation between people is often a better target than work already done well by a stable expert team.
Start from the work, not the ideas
Instead of collecting ideas, pick a workflow that matters, such as claims handling, supplier onboarding, tender evaluation or customer complaints, and decompose it into its component tasks. Then assess each task, not the workflow as a whole.
Figure 2Illustrative
Assess tasks, not workflows
Example: complaint handling decomposed into six tasks
- Classify and route incoming complaintVolume: HighTo verify: CheapError cost: LowAutomate
- Retrieve account and order historyVolume: HighTo verify: CheapError cost: LowAutomate (integration)
- Summarize complaint and prior contactsVolume: HighTo verify: CheapError cost: MediumAI drafts, agent checks
- Decide remedy or compensationVolume: MediumTo verify: CostlyError cost: HighHuman decides; AI suggests policy
- Draft customer responseVolume: HighTo verify: CheapError cost: MediumAI drafts, agent edits
- Handle regulatory escalationVolume: LowTo verify: CostlyError cost: HighLeave with specialists
Four properties decide whether a task is a good AI candidate:
- Volume. How many times a month is this done? High volume turns small per-task gains into material value.
- Language and judgment content. Does the task involve reading, writing, summarizing, classifying or reasoning over text, documents or code? That is where current models are strongest.
- Cost of an error. What happens if the output is wrong and nobody notices? A misrouted ticket is cheap. A wrong figure in a regulatory filing is not.
- Verifiability. How quickly and cheaply can a person, or a system, confirm the output is right?
The last property is the one most selection methods ignore, and it is often the deciding one.
The question most grids miss: can it be checked?
Generative AI changes the economics of producing a first draft. It does much less to the cost of checking it. So the most valuable use cases share a pattern: the output is expensive to produce from scratch but cheap to verify.
Figure 3
Cost to produce vs. cost to verify
Generative AI cuts the cost of producing a draft, not the cost of checking it
Sweet spot
AI drafts, people check. Biggest returns.
Code with tests · contract field extraction · reply drafts · sourced summaries
Proceed with care
Value depends on how review is designed and costed.
Investment memos · legal opinions · clinical summaries
Automate at volume
Worth it only where volume is high.
Ticket routing · document classification · data entry
Avoid
AI adds checking cost to cheap work.
One-off judgment calls · short, high-stakes answers
Cost to verify →
Code with a strong test suite is the canonical example. Writing the code is slow, while running the tests is fast and objective, which is a large part of why software engineering has become one of the clearest AI success areas. The same logic applies to extracting fields from contracts against a known template, drafting customer replies that an agent reviews before sending, or summarizing long documents with links back to the source passages.
The difficult quadrant is where output is expensive to produce and expensive to verify: an investment memo, a legal opinion or a clinical summary. AI can still help, but only if the review step is designed and costed honestly. If an expert must re-derive every conclusion to trust it, the time saved in drafting is spent again in checking.
Practical test: for every candidate task, write down who verifies the output, how, and how long it takes. If nobody can answer, the use case is not ready for scoring.
Score with evidence, not opinions
Once candidate tasks have been identified, score them, but replace one-to-five opinions with evidence that can be checked.
| Criterion | Instead of asking | Ask for |
|---|---|---|
| Volume | "Is this frequent?" | Monthly volume from system logs or ticket data |
| Baseline | "Is this slow?" | Current time per task and error or rework rate, measured |
| Model fit | "Did it work in a demo?" | Results on 50 to 100 real cases, including the 10 hardest |
| Verifiability | Not usually asked | Who checks, how, and minutes per check |
| Cost of error | "Is this risky?" | The worst plausible error and its consequence |
| Data readiness | "Do we have the data?" | Named source systems, owners and access confirmed |
| Adoption | "Will people use it?" | A named business owner and a target user group who have seen it |
Two rules keep scoring honest. First, red lines are not scores. If a task would breach regulation, use personal data without a lawful basis or make a consequential decision without human accountability, it is out, regardless of how valuable it looks. Second, an unknown is not a 3. If the baseline has not been measured, the score is "unknown" and the next step is to measure it.
Build a portfolio, not a ranking
A ranked list treats use cases as independent. In practice they share foundations, and the foundations are where most of the cost and delay sit.
Figure 4Illustrative
Fund the foundations that unlock several use cases
Shortlisted use cases mapped to the shared foundations they depend on
| Use case | Customer master data | Policy and contract documents | CRM integration | Evaluation harness |
|---|---|---|---|---|
| Complaint summarization | ||||
| Reply drafting | ||||
| Contract clause extraction | ||||
| Churn-risk alerts | ||||
| Agent knowledge assistant | ||||
| Use cases unlocked | 3 | 4 | 2 | 4 |
Map each shortlisted use case against the foundations it needs: specific data domains, document collections, integrations, evaluation tooling, and access to the right model platform. Foundations that unlock three or more use cases deserve funding in their own right, even though none of them is a "use case". This is also where AI selection connects to data governance. The work of cataloging, classifying and improving the quality of critical data described in our guide to the NDMO standards is exactly what most AI use cases quietly depend on.
A balanced portfolio usually contains three kinds of work:
- Proven patterns in high-volume tasks with cheap verification, which deliver value within months and build credibility.
- Foundations that unlock several use cases at once.
- A small number of bets on harder, higher-value problems, funded in tranches with explicit stop criteria.
A 30-day selection sprint
For organizations that want to replace the idea-workshop approach, a focused sprint on one or two workflows produces better candidates than a company-wide brainstorm:
- Week 1: Choose and map. Select one or two workflows with high volume and a committed business owner. Map them task by task with the people who do the work.
- Week 2: Measure. Pull real volumes, times and error rates from systems. Collect 50 to 100 representative cases per candidate task, including the hardest ones.
- Week 3: Test at the edge. Run candidate tasks through current models on the real cases. Record where output is good, where it fails, and how long verification takes.
- Week 4: Decide. Score with evidence, apply red lines, map dependencies, and approve a small portfolio with owners, value cases and stop criteria.
The output is not a list of 200 ideas. It is three to five use cases that have already been tested on real work, each ready for the value-case and gate discipline that follows.
Frequently asked questions
What makes a good enterprise AI use case?
A task with high volume, substantial language or judgment content, a tolerable cost of error, and output that is cheap to verify. It also needs a measured baseline, accessible data and a business owner who will change the workflow.
Why is verifiability so important in AI use-case selection?
Generative AI lowers the cost of producing a draft but not the cost of checking it. If verifying the output takes as long as producing it manually, the use case saves little, however good the model is.
What is the jagged technological frontier?
A term from a Harvard Business School and BCG field experiment describing the uneven boundary of AI capability. On tasks inside the frontier AI improved speed and quality; on a task just outside it, AI users were less likely to be correct than people working without AI.
Should AI use cases be ranked by value and feasibility?
Value and feasibility matter, but a two-by-two ranking hides shared dependencies and relies on unmeasured guesses. Score with measured evidence, treat red lines as exclusions, and fund foundations that unlock several use cases.
Sources and further reading
- Navigating the Jagged Technological Frontier, Dell'Acqua et al., Harvard Business School Working Paper 24-013, 2023. Field experiment with 758 BCG consultants.
- Generative AI at Work, Brynjolfsson, Li and Raymond, NBER Working Paper 31161; published in The Quarterly Journal of Economics, 2025. Study of 5,179 customer support agents.