Ask your leadership team for AI ideas and you will have thirty by Friday. Sales wants proposal drafts. Finance wants faster close. Operations wants scheduling help. Customer service wants a chatbot, and somebody's nephew has a tool he swears by. Nobody is wrong. The problem is that nobody can pick, so nothing starts.
This article is about how to pick your first AI use case. It is the decision itself, made on paper, with the people who know the work. It is not a technology comparison and it is not a project plan. You need a list of ideas, a pen, and about an hour with two or three people who run the work day to day.
Why the first pick matters more than the best pick
You are not looking for the most valuable idea on the list. You are looking for the safest one that teaches you something.
The first use case sets the tone for every one after it. If it works and the team trusts it, the second one is easier. If it fails loudly, the whole effort gets the label "we tried AI." Some of the ideas on your list are high-value and also high-risk. They can wait. Your first pick should be boring on purpose.
Most of what AI does well in a mid-market company right now is narrow, repetitive work with a person nearby to check it. We lay out what AI can do for a mid-market company now in a separate piece. Here the job is to sort your own list.
Step one: write the list and strip it to plain verbs
Put every idea on one sheet, one line each. Then rewrite each line so it names a task, not a technology.
"Use AI in customer service" is not a task. "Draft a first reply to inbound support emails" is. "Make finance smarter" is not a task. "Match vendor invoices to purchase orders and flag mismatches" is.
If you cannot rewrite an idea as a task a person does today, set it aside. It is a goal, not a use case. Goals matter, and your use case should connect to one. We cover that link in building an AI strategy that connects to results. But you cannot screen a goal. You can only screen a task.
Step two: how to pick your first AI use case with four screening questions
Now run each task through four questions. Write yes, no, or "not sure" beside each. Be strict. "Not sure" counts as no until someone finds out.
1. Is the input stable?
Does the work arrive in a form that looks about the same each time? An invoice, an intake form, a call transcript, a standard customer email: stable. A pile of handwritten notes, scanned PDFs in six layouts, or a spreadsheet only one person understands: not stable.
AI handles variety better than old software did, but messy input produces messy output, and you will spend the pilot cleaning data instead of learning. If the answer is no, the fix may be to tidy the input first. Our pieces on documentation before AI and AI data readiness cover that ground.
2. Is the output defined?
Can you describe a good result in a sentence or two, and can two experienced people look at it and agree it is good? "A reply that answers the customer's question, uses our return policy correctly, and sounds like us" is defined. "Better insight into the business" is not.
A defined output is what lets you measure. If you cannot say what good looks like before you start, you will not be able to say whether it worked after.
3. Does a human confirm before it matters?
Look at the point where the result leaves the building or touches money, a customer, an employee, or a record you rely on. Is there a person who reviews it first?
The strongest first use cases have a natural checkpoint already. A manager approves the quote. An accountant reviews the entry. A rep reads the draft before sending it. AI does the first pass. The person decides.
If the only way to use the idea is to let AI act with no one looking, mark it no. That does not mean never. It means not first. The verification step is also where trust gets built, which we discuss in the verification layer for AI reliability.
4. Is it reversible?
If the output is wrong and gets through anyway, can you undo it? A bad internal summary is easy to correct. A wrong email to a client is harder. A wrongly denied refund, a wrong payment, or a wrong note in an employee file is harder still.
Reversibility is your margin for error. It is why the first pick should be work where a mistake costs an hour, not a relationship.
Step three: read the sheet before you pick your first AI use case
Count the yeses on each line.
- Four yeses: a candidate for your first use case.
- Three yeses and one "not sure": a candidate, once someone answers the question.
- Two or fewer: park it. Write down which question failed, because that tells you what would have to change.
Two things usually happen. First, the ideas people are most excited about tend to fail question three or four. Customer-facing, autonomous, high-stakes ideas are exciting for the same reason they are risky. Second, a few unglamorous ideas score four yeses: summarizing meeting notes into action items, drafting first-pass responses to routine requests, sorting and tagging inbound documents, preparing the weekly report from data that already exists.
Those are your candidates. Do not apologize for them.
Step four: break ties with two plain questions
If three or four ideas pass, you still need one. Ask two more things.
Who does this work, and do they want help? A use case succeeds or fails with the people doing the work. If the team dreads the task, they will welcome the help. If the task is the part of the job they take pride in, tread carefully. The first pick should give people time back for work that needs their judgment, not make them feel measured against a machine. We wrote about the dignity of work and getting employee adoption right for exactly this reason, and about what employees fear about AI.
Will you notice if it works? Pick the candidate where the change would be visible and countable in your current numbers: hours spent, turnaround time, error rate, backlog. You do not need a projected result. You need a baseline you can already read today. Write down what the number is now, before anything changes. Without that, you will be arguing from impressions in ninety days. Our piece on measurable goals for AI strategy goes deeper on this.
Pick one. Not two.
Common mistakes
Picking the idea with the loudest advocate. Enthusiasm is useful for adoption. It is not a screening question.
Picking the biggest prize. The largest opportunity usually touches the most systems, the most people, and the most risk. It is a good second or third project, once you have learned how your organization handles this kind of change.
Picking a tool before a task. If the first sentence is "we bought a license for...", you have a tool looking for a job. Reverse it.
Skipping the baseline. If you do not record the current number, you cannot show what changed.
Treating the first pick as permanent. It is a small, deliberate step. You can be wrong about it and lose very little. That is exactly why question four exists.
A word on stewardship
There is a reason we put reversibility and human review into the screen rather than leaving them as afterthoughts. You are responsible for the people who work for you, the customers who trust you, and the resources you have been given to manage. Starting small and checking the work is not timidity. It is what careful stewardship looks like when the technology is new and your people are watching how you handle it. We take that up more fully in the stewardship framework for AI.
What happens after you pick
Once you have a single use case, the work shifts from deciding to doing: who owns it, what the pilot covers, how you will know it worked, and when you stop. The pilot charter is where that gets written down, and we will cover it separately. For the stretch after the decision, see what the first 90 days after deciding to adopt AI look like.
Do this today: how to pick your first AI use case on one page
Take thirty minutes. Write your team's AI ideas on one page, one task per line. Add four columns: stable input, defined output, human confirms, reversible. Fill in yes, no, or not sure. Circle anything with four yeses and send the page to the two people closest to that work with one question: "Is this right?"
You will learn more from their answers than from another brainstorm.
Bring the sheet to a Discovery Call
If you do this exercise and end up with three candidates that all look the same, or none that pass, that is a normal place to be. It is also a good use of thirty minutes with us. Bring the marked-up sheet to a Discovery Call. It is thirty minutes, no cost, no obligation. We will go through it with you, and we can tell you if the answer is "this one, not that one yet," or "smaller than you think," or "fix the input first."