AI use cases: how to choose the first one worth doing
A scoring method and a one-week exercise for choosing the first AI use case your organisation should commit to.
BY STEFAN MARITZ··7 MIN READ
TL;DR
- A good first AI use case is frequent, bounded, owned by one business leader, fed by data you already have, and cheap to get wrong. Visibility and novelty are poor reasons to pick one.
- Poor use-case selection and unclear business value are the most common reasons generative AI projects get abandoned. The choice of first project matters more than the choice of tool.
- Score each candidate on five criteria: value, frequency, data readiness, cost of an error, and whether someone will own the change. Weight cost of error heavily for a first project.
- Keep decisions about people, credit, or anything customer-facing and high-stakes off the first list. Start with internal work where a human checks the output.
- A leadership team can produce a ranked shortlist in one working week using the exercise at the end of this post.
Ask a leadership team for AI use cases and you will get 40 ideas in an hour. Ask which one they would bet a quarter on and the room goes quiet. The list was never the hard part.
This post is about the second question. It covers what makes a good first use case, where to look for one in operations, marketing, and HR, a scoring method you can run with your own team, and the kinds of use case I would keep off the first list.
What makes a good first AI use case?
A good first AI use case is one where success is likely, measurable, and visible to the people doing the work. It earns the trust you need for the second and third projects. A spectacular idea that stalls in month two costs more credibility than it was ever going to earn.
In practice, strong first candidates share six traits:
- It happens often. A task done 200 times a month pays back faster than one done twice a year, and people learn the new way quickly.
- It has clear edges. You can describe the input, the output, and what "good" looks like.
- The data already exists. If the first step is building a data pipeline, it is a data project with AI attached.
- A human reviews the output. Drafting, summarising, classifying, and searching all fit this.
- One leader owns it. Shared ownership usually means nobody changes the process.
- An error is cheap. If the model gets it wrong, someone notices and fixes it before it reaches a customer.
Why does the first choice matter so much?
Because the failure modes are well documented, and most of them are locked in at selection. At least 30% of generative AI projects were expected to be abandoned after proof of concept by the end of 2025, due to poor data quality, weak risk controls, escalating costs, or unclear business value. Poor use-case selection combined with a lack of business value sits at the top of the list of causes.
Scale tells the same story. Only 22% of companies have advanced beyond the proof-of-concept stage, and only 4% are creating substantial value. My view is that most of the projects in the other 78% were technically possible and the wrong first bet.
Where should you look for AI use cases?
Look where skilled people spend time on work that is repetitive, text-heavy, and checked by someone anyway. That pattern shows up in every function, with different specifics.
- Operations: turning emails and forms into structured records, drafting standard responses, summarising incident logs, and searching internal procedures, all of which show up among the strongest AI use cases in operations.
- Marketing: first drafts of briefs, repurposing long content into shorter formats, and summarising customer research, which is where AI fits most naturally in a marketing team.
- HR: drafting job descriptions, answering policy questions from an approved handbook, and structuring interview notes. Screening and ranking candidates is a different matter, covered below.
- Across the business: a searchable company knowledge base, so people stop asking the same three colleagues the same questions.
How to score and prioritise AI use cases
Score each candidate from one to five on the five criteria below, and multiply by the weight. The weights reflect a first project, where learning safely matters more than maximum upside. For later projects you can shift weight from cost of error towards value.
| Criterion | Question to ask | Weight for a first project |
|---|---|---|
| Value | If this works, what number moves, and by roughly how much? | x2 |
| Frequency | How many times a week does this task happen? | x1 |
| Data readiness | Can we get the inputs today without a new integration? | x2 |
| Cost of an error | What happens if the output is wrong and nobody catches it? (score five for "very little") | x3 |
| Ownership | Is there one leader who wants this and will change the process? | x2 |
The scores are a conversation tool, and a precise total matters less than the discussion it forces. When two executives score data readiness differently, you have found something worth knowing before any money is spent.
Which AI use cases should you avoid first?
Keep anything that makes or shapes decisions about people off the first list. Recruitment screening, performance ratings, and credit decisions carry legal weight in most jurisdictions. Under the EU AI Act, CV-sorting software for recruitment is classed as high-risk, and in South Africa, POPIA restricts decisions based solely on automated processing that profile a person. These can be good later projects with proper governance. They are poor places to learn.
I would also hold back on customer-facing chat where a wrong answer reaches a customer directly, and on anything where the honest fix is a better process. Some candidates will turn out to need rule-based automation instead of AI, which is cheaper and more predictable.
A one-week exercise to choose yours
Choosing a first use case works like choosing which station on a production line to improve first. You pick the one with steady volume, a clear specification, and an operator who wants the change.
- Monday: collect candidates. Ask each functional lead for their three most time-consuming repeatable tasks, described in plain language.
- Tuesday: watch the work. For the top 10, spend 20 minutes with someone who does the task. You will learn more about data readiness here than in any workshop.
- Wednesday: score independently. Have three or four leaders score each candidate on the table above without conferring.
- Thursday: compare and argue. Discuss every criterion where scores differ by two or more points.
- Friday: pick one and name the owner. Agree the success measure, the people whose workflow changes, and the date you will decide whether to continue.
The adoption step hides in Friday. If the team doing the work is not part of the decision, the use case will be technically successful and practically ignored, which is how AI pilots fail most often.
Frequently asked questions
How do you choose the right AI use case?
Choose a task that happens often, has clear inputs and outputs, uses data you already have, has one owner, and is cheap to get wrong. Score candidates on value, frequency, data readiness, cost of error, and ownership, then discuss where scores disagree.
What are the most common AI use cases in business?
Drafting, summarising, classifying, and searching text are the most common starting points, because a person can check the output. Examples include standard email responses, meeting summaries, policy questions, and content repurposing.
How many AI use cases should we start with?
One or two. A small number lets the team learn how to change a workflow properly, and that skill transfers to every project after it.
Should our first AI use case be customer-facing?
Usually not. Internal use cases with a human reviewing the output let you learn where the model fails before a customer sees those failures.
What is the difference between an AI use case and an AI strategy?
A use case is a single task or workflow where AI is applied. A strategy decides which use cases matter, in what order, and what has to be true across data, people, process, and tooling for them to succeed.
If you want a scored starting point before you run the exercise, the AI Readiness Diagnostic is free and shows which of the five readiness dimensions is likely to limit your first use case.
WORK WITH PRAXES
Ready to move from reading about AI to getting your organisation ready for it? Start with the diagnostic.
Take the AI Readiness Diagnostic