How to choose AI use cases is the first real decision in any AI program, and many teams make it backwards: they pick a tool, then look for work to give it. This guide gives you a method that starts with the work. It includes a scorecard for value, feasibility, data readiness and risk, a worked example with six real-world candidates, and a pilot plan with a success threshold and a stop rule.
How do you choose AI use cases?
List candidate tasks, each with an owner, a metric and a current baseline. Drop any that fail a few pass-or-fail gates. Score the rest from 1 to 5 on value, feasibility, data readiness and risk, weight the scores, and pilot the top one or two with a clear success threshold and a rule for stopping. Choose few: companies that get the most from AI concentrate on a small number of use cases.
Score tasks, not departments. "Use AI in finance" can't be scored. "Draft variance comments for the monthly close" can.
Run the gates before the scores. No owner, no baseline, no permission to use the data, or AI deciding about people alone: the idea is not ready.
Data is the usual blocker. In a Gartner survey of 1,203 data management leaders (July 2024, published February 2025), 63% said their organization lacked, or wasn't sure it had, the right data management practices for AI.
Fewer is better. In BCG's October 2024 survey of 1,000 executives, the companies furthest ahead pursued about half as many AI opportunities as others and expected more than twice the return.
Every pilot needs a stop rule, written down before it starts.
[02]
What makes a good AI use case?
A good first AI use case is a task, not a job. It has these traits:
Frequent. It happens many times a week, so small gains add up.
Checkable. A person can tell quickly whether the output is right.
Built on text or documents. Emails, tickets, notes, forms and reports are where today's AI tools are strongest.
Owned and measured. Someone runs it today and knows how long it takes or how often it goes wrong.
Painful. There is a backlog, a delay, overtime or a recurring error.
Forgiving. One mistake is cheap, or a review step catches it before it matters.
Warning signs are rare tasks, tasks where nobody agrees what good looks like, one-off decisions with high stakes, and problems that a process fix or plain automation would solve without AI.
Look beyond the obvious places. MIT NANDA's State of AI in Business 2025 report (July 2025), based on interviews with 52 organizations, observed that budgets favor visible sales and marketing projects while back-office automation often paid back faster. It also found that successful companies let budget holders and frontline managers bring forward use cases, rather than leaving the search to a central AI team.
[03]
When to use a scorecard (and when not)
Use a scorecard when you have more ideas than people or budget, when several departments compete for the same team, or when you need to explain a choice to leadership. It turns opinions into a ranked list everyone can argue with.
Skip it when there is one obvious candidate with a willing owner: run a small pilot instead. And don't use a scorecard to pick a tool. Choose the use case first, then decide whether to build or buy.
[04]
Step 1: Collect candidates in a form that can be scored
Ask each team lead to submit candidates in the same shape. One line each is enough:
Field | Example |
|---|---|
The task | Reply to customer emails asking where a shipment is |
Who does it today, and how often | 30 support agents, about 1,800 emails a week |
The metric it would move | Average handling time per email |
Today's baseline | About 6 minutes per email |
The data it needs | Shipment status from the transport system; past replies |
Who would own the pilot | Support team lead |
What a wrong output costs | A wrong delivery date in a reply; caught if a person reviews |
A request like "use AI in customer success" goes back to its sender until it names a task, a metric and an owner.
[05]
Step 2: Apply the gates
Before you score anything, check five pass-or-fail gates. A candidate that fails one waits until it passes.
Owner: a named person will run the pilot and own the result.
Baseline: there is a current number for the metric, or you can measure one within two weeks.
Data permission: the data may be used this way under your company's existing data rules. Ask the data owner or security team.
People decisions: AI will not decide alone about hiring, performance, credit, health, legal or financial matters for people. If it would, HR or legal must be involved in the design first.
Checkability: a person can check the output, or the action can be undone.
Some uses of AI are regulated in some places. Ask your legal and compliance team what applies to you before you roll this out.
[06]
Step 3: Score value, feasibility, data readiness and risk
Score each candidate from 1 to 5 on four criteria. Use the descriptions so that a 3 means the same thing to everyone.
Criterion (weight) | 1 | 3 | 5 |
|---|---|---|---|
Value (30%) | Saves a few hours a month; nobody would notice | A clear gain for one team on a metric it tracks | Moves a number leadership tracks, at volume |
Feasibility (25%) | Needs new systems, long unsupervised work, or abilities today's AI lacks | One integration or a few weeks of setup; others have done similar work | Runs in a tool you already have, with little or no setup |
Data readiness (25%) | The data doesn't exist, is on paper, or may not be used | The data exists and may be used, but is scattered or messy | The data is in one place, current, and approved for this use |
Risk (20%, higher is safer) | Errors are hard to spot, hit customers or money directly, and can't be undone | A review step catches errors before they matter | Errors are harmless or obvious; internal use only |
Run the scoring as a 60 to 90 minute session with four to six people: the owner, someone who does the work, someone from IT or data, and someone from security or risk. Each person scores alone first. Then discuss any criterion where scores differ by more than one point, and agree a final number. This catches the optimism of the person who proposed the idea.
[07]
Step 4: Weight, rank and apply the floors
Combine the scores into one number from 20 to 100:
Score = 20 × (0.30 × Value + 0.25 × Feasibility + 0.25 × Data readiness + 0.20 × Risk)
Then apply two floors, because a weighted average lets ease hide a lack of value:
Value floor: a funded pilot needs a value score of at least 3. Easy, low-value ideas go on a "quick tools" list, usually solved with an AI assistant you already have.
Data floor: a data readiness score of 1 means fixing the data comes first.
Change the weights if your priorities differ, for example a higher risk weight for customer-facing work. Agree the weights before anyone scores, not after.
[08]
Worked example: six candidates at a logistics company
Harbor Logistics is an example company: a freight forwarder with 400 employees. Team leads submitted six candidates. Here is the scored result:
Candidate | Value | Feasibility | Data | Risk | Score | Decision |
|---|---|---|---|---|---|---|
A. Draft replies to shipment-status emails | 4 | 4 | 4 | 4 | 80 | Pilot now |
B. Summarize carrier contracts for sales | 2 | 4 | 4 | 3 | 64 | Quick tools list: value below 3 |
C. Enter data from scanned bills of lading | 4 | 3 | 3 | 3 | 66 | Next quarter, after improving scan quality |
D. Screen job applicants' CVs | n/a | n/a | n/a | n/a | Gate failed | Parked: AI would rank people; HR and legal must design any version first |
E. Forecast weekly volume per route | 4 | 2 | 3 | 4 | 65 | Hand to the analytics team; a forecasting model fits better than a chat tool |
F. Answer expense-policy questions | 2 | 5 | 5 | 5 | 82 | Quick tools list: use the assistant staff already have |
To check one row: candidate A scores 20 × (0.30 × 4 + 0.25 × 4 + 0.25 × 4 + 0.20 × 4) = 20 × 4.0 = 80. Candidate F scores 20 × (0.6 + 1.25 + 1.25 + 1.0) = 82.
Three lessons from this example:
The top score was not the right pilot. F ranked first because it is easy and safe. The value floor caught it.
The gate did its job. D never reached scoring. Screening CVs is a decision about people, so it needs HR and legal before any design, and AI should not make that decision alone.
Not every good idea is an AI chat idea. E has real value, but forecasting volume is a job for statistical models and the analytics team.
[09]
Step 5: Design the pilot
A pilot is an experiment with a decision at the end, not a soft launch. Here is Harbor Logistics' plan for candidate A:
Item | Plan |
|---|---|
Hypothesis | Drafted replies cut handling time for shipment-status emails from 6.0 to 4.5 minutes or less, without lowering quality |
Baseline | Two weeks before the pilot: 1,800 emails a week, 6.0 minutes average handling time, quality score 4.2 out of 5, reopen rate 8% |
Who | 8 of the 30 support agents use the drafts; the other 22 are the comparison group |
Duration | Two weeks of baseline, then six weeks of pilot |
Success threshold | Handling time 4.5 minutes or less, quality score 4.2 or higher, reopen rate 8% or lower |
Stop rule | Stop or redesign if handling time falls by less than 10% by week 4, or after any data incident |
Cost cap | $3,000 in tool and usage costs, plus four hours a week of the team lead's time |
Review | 15 minutes every week; a decision meeting in week 8: scale, fix or stop |
Owner | The support team lead, with a named IT contact for the email integration |
The comparison group matters. Handling time often improves during a pilot simply because people pay more attention. Comparing against colleagues who work without the tool separates the AI's effect from that. How to measure the ROI of AI continues this example and turns the results into a return.
Two practical points for the build. The quality of the drafts depends heavily on the instructions the tool is given, which is what prompt engineering is about. And the team should test the drafts on 50 to 100 real past emails before the pilot starts, checking especially for invented delivery dates, a common type of AI hallucination.
[10]
The AI use case scorecard to copy
Copy this into a spreadsheet. One row per candidate.
Use case | Owner | Metric and baseline | Gates passed (yes or no) | Value 1-5 (×0.30) | Feasibility 1-5 (×0.25) | Data readiness 1-5 (×0.25) | Risk 1-5 (×0.20) | Score (×20) | Value ≥ 3 and data ≥ 2? | Decision |
|---|---|---|---|---|---|---|---|---|---|---|
… | … | … | … | … | … | … | … | … | … | Pilot, next quarter, quick tools, hand off, or park |
… | … | … | … | … | … | … | … | … | … | … |
In a spreadsheet, the score formula is =20*(0.3*E2+0.25*F2+0.25*G2+0.2*H2) if value sits in column E.
[11]
Common mistakes
Mistake | Why it fails | Fix |
|---|---|---|
Starting from the technology ("we need an agent") | Use cases get invented to fit the tool | Start from a list of painful, frequent tasks |
Scoring vague ideas | "AI for HR" has no metric or owner | Require a task, a metric, a baseline and an owner |
Running many pilots at once | People and budget spread thin, and nothing scales | Run one to three pilots at a time |
Ignoring data until the build | Many projects stall on data access or quality | Score data readiness; fix a 1 before anything else |
Letting the proposer score alone | Optimism inflates every number | Score independently, then discuss gaps |
Picking the easiest idea | Low value proves nothing to leadership | Apply the value floor |
No stop rule | Pilots drift on for months | Write the stop rule before the pilot starts |
Changing weights after scoring | The ranking bends to fit a favorite | Lock the weights first |
[12]
Frequently asked questions
What is AI use case prioritization?
AI use case prioritization is the process of ranking possible uses of AI so that you fund the ones most likely to deliver value. A common method scores each candidate on value, feasibility, data readiness and risk, applies a few pass-or-fail gates, and pilots the top one or two with a measured baseline.
What are good first AI use cases?
Good first use cases are frequent, text-heavy tasks where a person can check the output: drafting replies to common emails, summarizing tickets or calls, extracting data from documents, and answering questions from internal documents. Avoid first projects where one error is costly, or where the AI would decide about people.
How many AI use cases should we pilot at once?
One to three, depending on the size of your team. Each pilot needs an owner, a baseline and weekly attention. In BCG's 2024 survey, the companies furthest ahead pursued about half as many AI opportunities as their peers. Finish, measure and decide before adding more.
What criteria should we use to evaluate AI use cases?
Use four weighted criteria: value (30%), feasibility (25%), data readiness (25%) and risk (20%), each scored from 1 to 5 with written descriptions. Add pass-or-fail gates for an owner, a baseline, permission to use the data, checkable output, and no AI decisions about people made alone.
How long should an AI pilot run?
Long enough to beat the novelty effect and see real volume: usually two weeks of baseline and four to eight weeks of pilot. Set a check-in at the halfway point with a stop rule, and a decision meeting at the end where you scale, fix or stop.
Does the scorecard work for AI agents too?
Yes, with stricter risk scoring. An agent takes actions instead of drafting, so score risk on the worst action it could take, and require a person's approval for anything you can't undo. AI agents explained for business leaders covers what agents can safely do today.
[13]
Next steps
Understand what agents can and can't do: AI agents explained for business leaders and what is an AI agent?
Decide how to source the tool: build vs buy AI agents.
Decide which steps a person approves: what is human in the loop AI?
Prove the pilot worked: how to measure the ROI of AI.
Bring the team along: how to roll out AI to your team.
Hand your technical team a testing method: how to evaluate AI agents.