The loudest idea wins
In most organisations the first AI use case is chosen by volume. Someone senior saw a demo, or a vendor offered a free trial, or one team had budget left in the last quarter. The project starts because it can, and nobody asks whether it was the right place to begin.
That choice matters more than it looks. The first project is the one the rest of the company watches. If it saves a visible amount of time on a task people dislike, the second project gets volunteers. If it produces a clever output that nobody needed, the second project gets polite excuses.
The 2024 Work Trend Index from Microsoft and LinkedIn found that 79% of leaders agree AI adoption is critical to staying competitive. The same report found that 60% say their company lacks a vision and plan to implement it. Urgency without a plan is how accidental first projects happen.
Five tests for a first use case
I use five tests when I look at a list of candidate tasks. None of them needs a data scientist.
Frequency comes first. A task done 200 times a month gives a small gain many chances to add up, and it gives you enough volume to see whether the gain is real. A task done twice a year does neither.
The second test is whether the task has clear rules. Matching a supplier invoice to a delivery note has a right answer. Deciding whether to extend credit to a struggling customer involves judgement and relationships. Start with the first kind.
The third test is the cost of an error. Choose work where a mistake is caught by the next step and takes minutes to fix. Leave anything that touches a regulator, a customer's money or someone's employment until you have experience.
The fourth is data access. If the information the tool needs sits in one system the team already uses, you can start this month. If it needs three departments to agree on an export, the project will spend its first quarter in meetings.
The fifth is ownership. One named person must want the result and have the authority to change how their team works. A use case with a committee behind it has no owner.
What a ranked list looks like
The output of this exercise is a short table that fits on one page. Take a fictional mid-sized freight company and its finance operations team. Three candidates might survive the five tests.
Matching supplier invoices to delivery notes ranks first. It happens daily, the rules are clear and the accounts payable lead owns it. The effort is medium because the two documents arrive in different formats. Drafting replies to payment-status emails ranks second, with low effort and credit control as the owner. Summarising the weekly cash position ranks third, also low effort, owned by the finance manager.
Notice what is missing. There is no customer chatbot, no forecasting model and nothing with the word transformation in it. The list is dull, and each item has a person who will be glad when it works.
Three mistakes that repeat
Three mistakes come up again and again.
The first is starting with the customer. Customer-facing AI carries the highest cost of error and the most reputational risk. Internal work is a safer place to learn.
The second is choosing a task because the tool is impressive at it. Writing a strategy memo looks good in a demo and saves almost nothing, because nobody writes 200 strategy memos a month.
The third is picking ten use cases at once. Attention is the scarce resource. Ten projects with a tenth of the attention each will all stall, and the organisation will conclude that AI does not work here.
Ask the people who do the work
The best source of candidates is the team itself. Ask each person which task they would hand over tomorrow if they could, and how often it comes up. The answers are usually specific and unglamorous: retyping figures from one system into another, chasing the same three approvals or writing the same email with small changes.
In my experience a list built this way is shorter and more accurate than one produced in a leadership workshop. It also arrives with owners already attached.
Write the list down
A first use case does not need a programme. It needs one page with the ranked list, an owner beside each line and a date to review the first one. With that page the choice becomes a decision that someone made and can defend.
This is the work I do in an AI Opportunity Assessment, which is the Start stage of the four I use. It is a fixed scope at US$1,500, planned at eight founder hours, and it ends with that ranked list and a readout. You can also do a rough version yourself with the five tests and an honest hour with the people who do the work.
Whichever route you take, pick the first project on evidence. The organisation will judge every later AI proposal by how that one went.