Let's talk about 2 subjects: Machine Learning and AI. Machine Learning is a subset of AI and is the current darling of the industry.

IMHO, Machine Learning is now a commodity. Hundreds of thousands of people know how to do it, some are good and some are not. Most of the well tested algorithm are open source. There are freelancers offering to do ML from $20 per hour and up. SO, again, IMHO, a company should pay little more per hour than a Java programmer costs for a ML coder. It's just not that hard. Industry pundits would like you to think it is esoteric rocket science. They are lying their AO.

Further, Machine Learning algorithms are not intelligent. They are just sophisticated nonlinear statistical algorithms to draw conclusions from prior cases, sort of like advanced regression. We call it AI, but its not really smart. On the other hand, it is very highly useful and valuable if you have a data set with valuable intelligence in it. So therefore, it does not matter that machine learning algorithms are both commodities and not intelligent, because they are extremely useful. For Pete sakes, O'Reilly has published an undergraduate level book on it that any C programmer can read in a couple days and deploy. Don't be conned by the arrogant. Even competing neural nets are a known commodity.

Now, the real challenge in ML is picking and scrubbing data (really boring but important work). Any company can do it through employees or contractors for $75 per hour or less. Don't you ever dare pay an AI professional $500,000 a year again. If you do, your idiocy.

Now for AI. AI is much broader than ML, consists of many unsolved problems like consciousness, true inductive planning that is useful, curiosity and motivation. These problems are not solved. There are some other useful AI topics that can be deployed: machine vision and natural langue and speech recognition are coming along, but can also improve.

Use cases now are a whole lot less about AI and a whole lot more about data science. Do you have data that can be scrubbed and mined? Good, then you can easily setup use cases. But keep your eyes on the prize. the prize is not AI. It is actionable intelligence from data. Don't get lost in the hype. Reality will hit hard in 2018. In summary, there is very very high value that can be obtained from big data sets with known AI algorithms. The AI part is commoditized, generally the algorithms are not smart, but it doesn't matter. Good data science is where its at.

Let me disagree with the previous answer a little. The existence of frameworks etc. make it SEEM easy to apply ML and it is if you are happy with mediocre results. If you want good results, then the person(s) working for you must know BOTH the mathematics and the domain that provides the data quite well. It is the domain knowledge portion that most companies ignore at present, at their great cost. How the data is collected, where it comes from, how dirty it is and so on matters a lot!

As far as user intent is concerned, I would go about this as with any other project. Does it matter if ML is under the hood or not? All the user wants is the question answered or the problem solved and that solution to be displayed nicely, right? So I would recommend getting a small user group to guide an agile development project to give them what they need. For this, leave out ML for the time being and focus on getting input/output right first. Then introduce ML as the cherry on the cake later on.

We do this all the time in practice and the majority of user feedback is on usability of the interface and such. The actual math is the heart of it, true, but it is expected to run in the background and does not influence the user experience much at all; unless the answers are poor of course.