"What is an AI use case?" Owners ask it after the third vendor in a month has promised to "identify your use cases", or after a consultant's deck arrives with a slide called "use case portfolio" and forty boxes on it. The honest answer is short. A use case is one job in your company, named in plain words, where a tool does part of the work and a person does the rest, with a number attached that you already care about. That is all it is. It is a sentence about your own business, and you can write it this afternoon.
Once you have one use case, the next question is how to test it; the page on what an AI pilot is covers the trial, the written test of success and the end date.
I was COO and then President of a public company that closed more than forty acquisitions, and I ran the integration on most of them. The initiatives that happened were the ones a named person could describe in one sentence with a number at the end. I am a lawyer by training, and nothing here is legal advice. Now, at 013 Labs, I write fixed-fee AI plans for CEOs of private-equity-backed and owner-led mid-market companies, and every plan starts with one use case written this way.
The short answer
A use case names a job, not a technology. "Draft the first answer to every customer email about an order" is a use case. "Use AI in customer service" is not; it is a wish. The difference matters because a job has an owner, a volume, a before and after, and a way to tell whether the tool helped. A wish has none of those, which is why wishes get bought and then sit unused.
A good use case has five parts, and they fit on half a page: the job, in the words the people who do it would use; the person who owns it; what the tool may touch and what it may never be given; what a person still does before anything counts; and the number you will look at in thirty days, written down before you start. If you cannot fill in all five you have an idea, not a use case, and ideas are not what a vendor should be pricing.
The number is the part owners skip and should not. Hours a week on the task. Mistakes a month. Days from a job finished to an invoice sent. Calls that went to voicemail. Pick one you already know or can count in a week. A use case without a number cannot be read, only argued about.
What makes a good one, and a bad one
A good first use case is boring. It is a task that happens many times a week, is already written down somewhere (a checklist, a template, a spreadsheet everyone complains about), is reversible when the tool gets it wrong, and is owned by someone who wants it off their desk. Drafting the first reply, pulling the figures for the weekly report, matching the paperwork to the job, writing the reminder before the renewal, turning the site notes into the first version of the report. None of these will make a headline. All of them give you a number inside a month.
A bad one is exciting. It is the vendor's demonstration, which was built for a company that is not yours. It is a one-off project nobody will repeat, so nothing is learned. It is a step that is a promise, a payment, a price or a hire, where a wrong answer costs real money or a real person. It is a problem nobody has measured, so nobody will be able to say whether anything improved. And it is anything that needs a tool to be right on its own, with no person in the loop, on the first day.
"Use case portfolio" deserves a plain translation: it is a list. A company with fifty to five hundred people does not need a list first. It needs one use case done honestly, read after thirty days, and either kept or stopped. The list can come after, and it will be a better list, because it will be written by people who have done one. The page on picking your first AI project covers how to choose when several look equal.
The data rule
Every use case touches information, and the sentence about what the tool may touch is the one that keeps you out of trouble. Decide in writing, before the tool is switched on, what may go in at all: not the full customer list, not pricing, not payroll, not drawings or recipes, nothing about an employee, nothing under a confidentiality agreement, and nothing from a client that the client would be surprised to find in a tool they have never heard of. Then decide which company account the tool runs on, because a use case living in someone's free personal account is the quiet problem the page on shadow AI describes, and it leaves with them.
The data rule is one or two sentences everyone involved can repeat, written on the same half page as the job. If a vendor's tool cannot work without the information you have ruled out, that is your answer about the vendor, not a reason to change the rule.
Five checks before you commit to one
First, say the job in one sentence without the word AI in it. If the sentence is about technology rather than work, start again.
Second, name the owner and ask them. A use case assigned to someone who did not want it is a project; a use case claimed by someone who is tired of the task is a change.
Third, write the number before you start, and write how you will count it. If you cannot count it in a week with what you already have, choose a different use case or a different number.
Fourth, mark the step where a person checks. Decide what the tool may finish on its own and what waits for a named person, and write that name down. The page on human in the loop is about exactly this decision.
Fifth, decide the stop. Pick the date you will look at the number and the result that means you switch the tool off. A use case with no stop written down becomes a subscription nobody remembers approving.
Check yourself
Seven yes or no questions. Each no is a decision still to make.
- Can you say the job in one sentence that a new employee would understand?
- Does it happen often enough that thirty days will show a result?
- Is there a number for it today, or one you can count in a week?
- Has one named person agreed to own it?
- Is it written down which information the tool may touch and which it may never be given?
- Is there a named person who checks before anything leaves, gets paid or gets filed?
- Have you written the date and the result that would make you stop?
Mostly no? That is the normal starting point. The AI readiness scorecard asks ten questions like these across the whole company and shows which of the three decisions is missing.
Finding one in an afternoon
You do not need a workshop. Three conversations and one walk through the inbox will do it.
Ask the two or three people in the company who already use these tools quietly what they use them for. There are always some, and they are usually the ones colleagues walk over to when something new appears. What they are already doing without permission is a list of candidate use cases, with volunteers attached.
Ask the people who run the office which task they would pay to never do again, and listen for the one that is frequent, dull and already half written down. Then ask what goes wrong when it is done badly, which tells you whether a wrong answer is reversible.
Look at the spreadsheet everyone complains about and the inbox that is never clear. Both are lists of jobs waiting for an owner, and the one with a number already attached is the one you will be able to read.
By the end of the afternoon you will have three or four candidates. Pick the one with the clearest number, the most willing owner and the smallest cost of being wrong.
Writing it on half a page
Five lines; a line can be a short paragraph.
The job: what gets done, how often, by whom today, in the words the team uses.
The owner: one name, and one line on what they decide. Who owns the question across the company is a different decision, covered in who owns AI in a company; this is the owner of one job.
The data: what the tool may be given, what it may never be given, and which company account it runs on.
The check: which steps the tool may complete, which wait for a named person, and what that person looks at before saying yes.
The number and the date: the figure today, the figure you would call a success, the day you will look, and the result that means you stop.
Date it and sign it. When the vendor's contract, the staff's questions and the next board meeting all point at the same half page, the use case is real. Show it to any vendor before signing anything; a good one will say which lines they can meet and which they cannot, and the page on evaluating AI vendor claims has the questions to ask if the answer is a demonstration instead. The invoicing example on this site, can AI do my invoicing, is that half page written out for one job, with the check placed where the money is.
Picture it going wrong
A made-up example, to show the shape. A sixty-person commercial printing company buys a tool after a demonstration in which it answered a customer's question about paper stock in seconds. The owner signs for a year. Nobody writes down the job: the sales team assumes it will quote, the office assumes it will chase late payments, the owner assumed both. Three months later it does neither well, two people have quietly stopped using it, and the renewal notice arrives with no number to point at, so the argument about whether it worked is about feelings.
The same company, a different start. The office manager says the job she would pay to never do again is writing the first answer to the forty or so customer emails a day that ask where an order is. That is the use case: draft the first answer to order-status emails from the job record; she reads and sends every one; the number is the time from a customer's email to the first answer, measured for a week before and for the month after; the tool is shown the order record and the customer's own message and nothing else; the stop is a reply sent unread, or no change in the number after thirty days. A month later there is a number, a log and a decision to keep it. The tool is the same in both stories. The use case is the difference.
Mistakes that repeat
- Buying the demonstration instead of describing the job.
- Choosing the biggest problem first, where a wrong answer costs the most.
- Starting with a list of twenty instead of one done properly.
- No number written before the start, so the thirty-day read becomes an opinion.
- An owner assigned rather than a volunteer found.
- The data rule left to the vendor's default settings.
- No stop date, so a trial becomes a renewal nobody chose.
- Letting the tool finish a step alone in the first month because it was right the first ten times.
Frequently asked questions
What is an AI use case?
One job in a company, named in plain words, where a tool does part of the work and a named person does the rest, with a number attached that the owner already cares about. It is a sentence about the business, not a type of software.
What is a good example of an AI use case for a small business?
Drafting the first answer to routine customer emails from the order record, with a person sending each one. Pulling the week's figures into the standard report for a manager to check. Writing the reminder before a renewal from dates already in the calendar. Each is frequent, written down somewhere, reversible and easy to count.
How many use cases should a company have?
One, done honestly and read after thirty days, before any list. A company of fifty to five hundred people learns more from one real result than from a portfolio of twenty candidates nobody has tried.
What is an AI use case portfolio?
A list of candidate jobs, usually produced by a consultant or a vendor. It is useful after the first use case has been run and read, because then the people who did it can tell which items on the list are real.
Who should own an AI use case?
The person who does or supervises the job today and wants it changed. Not the IT provider, not the vendor, and not the owner of the company unless the job is theirs.
How do I know if the use case worked?
The number you wrote down before you started moved in the direction you expected, nothing left the company unread, the log shows what the person changed and why, and the owner of the job would choose to keep it. If any of those is missing, you do not know yet.
Can a use case involve payments, prices or hiring?
Not as a first use case, and never without a named person deciding each one. Those are the steps where a wrong answer costs real money or a real person. Start where a mistake is cheap and reversible, and let the tool draft while a person decides.
Not sure where you stand? Take the 3-minute AI readiness scorecard.