When you're kicking off a project, clarity is everything. If you don't know exactly what you're trying to accomplish, how it benefits your customers, and how you’ll measure success, you’re setting yourself up for failure. If this is an AI project the cost is way higher and it takes longer so you don't want to screw up defining the use case.
Cassie Kozyrkov, data scientist and former Chief Decision Scientist at Google, put it best: "The fastest way to derail enterprise value is to throw AI at problems that are poorly defined or better solved with non-AI solutions."Â
She also came up with my favorite test for a use case Here’s my favorite test from also from Cassie Kozyrkov: Imagine AI as a "Drunk Island" challenge—people on an island with no ability to follow written instructions, only examples. If your task works there, it's a great candidate for AI.
What makes for a good AI use case?
It’s clear that your project offers value to both the customer and the business. But before you pat yourself on the back, ask yourself: Have you tested this idea with actual customers? Can you articulate the value in a way that resonates with your team and stakeholders? Chances are, there’s still work to do.
Here’s a quick test: Explain your use case to someone completely unfamiliar with the project. If they can understand it and explain it back in their own words, you’re on the right track. If not, it’s time to refine your message.
When it comes to AI, focus on automating tasks that make humans want to crush your soul and steal your energy. Think invoice processing, customer support triage, or labeling images—repetitive, mundane is what we're looking for.
The strength of AI is that it learns from examples. When my kids were toddlers, they loved books filled with animals so they could compare them—AI learns in a similar way. The AI pioneer, Fei-Fei Li, accelerated AI's progress by creating a dataset of 14 million labeled images. If you have a lot of transactions, sensor readings, assessments, logs, images and text that is promising.
AI is great for tasks that are easier to demonstrate than explain. For instance, training a robotic arm by showing it videos of humans performing tasks works better than writing out complicated instructions. This is where AI shines—learning by example.
If a problem can be solved with traditional software, do that. AI should be reserved for problems that can’t be solved with straightforward rules. For instance, I was once offered an option get my mammogram imagery evaluated for heart disease risk markers. You couldn’t write step-by-step instructions for that, but AI thrives on the vast amounts of data from about 37 million mammograms performed in the US each year.
Finally, make sure you have clear success metrics. Are you measuring for accuracy, speed, precision, or cost savings? Whatever your goal, establish a baseline so you can assess whether your AI solution is truly better, faster, or cheaper than the alternatives. Here's a useful guide from Google.
Focus on what sets you apart. What unique data do you have access to? How can you leverage that data, along with your deep understanding of your customers and industry, to create real value? For example, Akasa developed AI specifically tailored to the complexities of healthcare, using their deep industry expertise to address specific challenges and add significant value. (Disclosure: my husband works for them)
What's not a good use case?
Simple lookups like checking the time in another country or converting currencies are quick, straightforward and based done with traditional software engineering. If we know the answer it's not a good AI project.
The real strength of AI lies in identifying patterns. If there’s no clear relationship between inputs and outputs—if the data is too random or lacks structure—AI will struggle to produce meaningful results.
If you don’t know how your current process performs, you won’t have any way to measure whether AI is truly improving things. If you aren't measuring performance today, start with a project to get those baseline metrics. Without a baseline, you're operating in the dark, with no means to gauge progress or success.
AI is probabilistic, which means it operates with a certain degree of uncertainty. According to Jan Van Looy if your task demands absolute precision—like certain medical diagnoses—AI might not be the right fit, some use cases simply don’t allow for mistakes.
If AI isn’t going to provide a significant improvement, or if the cost to implement is too high for the expected benefit, the juice isn't worth the squeeze. I’ve met many entrepreneurs who found a use case but the cost of using AI was so high and the implementation so challenging that their client determined they were happy doing this the old fashioned way.Â
So how to I create a use case?
In my next post, I’ll explore how AI/ML teams approach use cases differently from product and design teams. I’ll also show you how to take your ideas and research and translate them into an use case your AI team will be excited to get.
Resources
Image: One of my daughters on the walkway between the east and west buildings of the National Gallery of Art in Washington, DC.