Choosing your first AI use case

Most companies we talk to don’t have a shortage of AI ideas. They have a shared document with forty of them, a few enthusiastic people, and no agreed way to pick one. The first project matters more than the rest, because it decides whether the second one gets funded.

Here is the method we use in our AI readiness assessments. You can run it yourself in an afternoon with the people who know the work.

Write each idea as a task, not a technology

“Use AI for customer service” isn’t a use case. “Draft replies to delivery-status emails, for an agent to check and send” is. Before you score anything, rewrite each idea as a specific task: who does it today, how often, what goes in and what comes out.

This step alone usually cuts the list in half. Some ideas turn out to be three separate tasks. Others turn out to be a reporting problem that a better spreadsheet would fix.

Score each task on value, effort and risk

Give each task a score from 1 to 5 on three questions.

Value: what changes if this works? Count something real, such as hours saved per week, faster turnaround for customers, or fewer errors that cost money to fix. A task done 500 times a month by six people beats a task done twice a quarter, even if the quarterly one feels more impressive.

Effort: how hard is it to build? The model is rarely the hard part. Effort comes from the data and the systems around it. Ask whether the information the AI needs is written down somewhere it can read, whether it lives in one system or five, and whether you’d need to connect to software that has no API.

Risk: what happens when it’s wrong? Every AI system makes mistakes, so the question is what a mistake costs. A wrong draft that a person reviews costs a minute. A wrong payment, a wrong medical instruction or a wrong answer sent straight to a customer costs much more. Score high-risk tasks low, at least for a first project.

Pick from the top-right corner

Add value and subtract effort and risk, or plot the tasks on a simple grid. The best first project is usually high value, low effort and low risk, and it tends to look a bit boring. That’s fine. Good first projects often share three traits:

• A person stays in the loop. The AI drafts, sorts or extracts, and someone checks the result before it counts.
• The input is text the business already has, such as emails, contracts, invoices, policies or tickets.
• Success is easy to measure. You can compare time per task, or error rates, before and after.

Document-heavy back-office work, like reading invoices or answering staff questions from policy documents, fits this pattern well. That’s why so many first projects land there.

Agree what “working” means before you start

Write down the measure of success and the number you need to hit before anyone builds anything. “Cut the time to process a supplier invoice from 12 minutes to 4, with no increase in errors” gives you a clear decision at the end of a pilot. “See if AI can help with invoices” doesn’t.

Agree as well what you’ll do if the pilot misses. A small project that fails quickly and cheaply is a useful result, as long as you learn why.

Check the risks you scored as low

Before you commit, spend half an hour on what could go wrong even in a low-risk task. Will the AI see personal data? Could someone outside the company send it input, for example through an email it reads? Which model provider will process the data, and where? These questions are much cheaper to answer at the start than after launch.

Where we’d begin

If you only do one thing after reading this, take your ideas list and rewrite the top ten as specific tasks. The ranking gets much easier once you can see what each idea actually involves.

If you’d like a second opinion, our readiness assessment gives you a ranked list of use cases for your business, with the risks of each, and a plan for the first one.