AI: Bridging the Gap from Potential To Impact

I believe that everything is an AI problem. Almost any problem can be addressed in some way using AI—better healthcare, food security, wildlife conservation—you name it. However, there is currently a huge gap between the potential of existing technologies and the impact they have in the real world. How do we bridge that gap? And how do we do it at scale?

I won't suggest that we have it figured out, but there are a few important lessons I have learned over the last twenty years of applying machine learning and artificial intelligence to industry problems.

First, success requires real collaboration and communication between the problem owners and the AI experts. It is incredibly easy for the problem owner to specify their goals in a way that is intractable, while a small variant of these objectives might be much easier to solve. Similarly, it is very easy for the AI experts to solve a formalization of the problem that doesn't actually address the business challenge. Oftentimes, most of the business value can be provided with state-of-the-art technologies that exist today. 

Second, technologists should focus on simple solutions that are easy to implement. In the trade-off between cheap, fast, and good, I believe that cheap and fast should usually win. That is, we should get to a working system as quickly and cheaply as possible and iterate from there. There are a couple of reasons for this: it clarifies the objectives, and it’s often very hard to answer the question "how good is good enough" before you have some working system. All too often, we can overshoot on the accuracy or performance of a solution, or we can fail to realize that perfection is required but is unattainable.

Third, if we're looking to replace or improve an existing process, we should understand how well that process is working now. We should have a baseline measurement of performance and we should be happy with how we made that measurement. This is essential to enable concrete arguments about the value of the solution. In addition, it is often surprising to see the error rates in human processes that we have assumed to be perfect.

Finally, while some of the underlying AI technologies are technically deep, these projects can and should be measured and assessed in the same way that we measure and assess any other project: quantitatively and objectively. And keep in mind that transformational change is hard. AI success will not be determined by the technology alone; rather, it will be largely determined by human factors such as leadership; cross functional collaboration, communication, and trust; and user acceptance.