The Boring Truth About High-Value AI Projects

Artificial intelligence may be the leading edge of advanced analytics, but it has not yet reached mainstream deployment in the enterprise. If you haven't deployed it, you aren't behind. Yet. Organizations that have deployed it are often in industries where they need it to remain competitive. Those deployments will continue to expand to other industries.

Use cases of existing deployments can provide a lot of insight into the best, highest-value projects for first time AI deployments in enterprises. A new report by McKinsey Global Institute (MGI) is the result of sifting through the data about hundreds of AI use cases to find that value. The report, Notes from the AI Frontier: Insights from Hundreds of Use Cases, provides a both a high-level perspective and a deeper dive into techniques to help every kind of enterprise get started.

When McKinsey Global Institute researchers took a step back from looking at each individual use case and looked at the whole project, they noticed something surprising about the results, according to Michael Chui, an author of the report and a partner at MGI.

"The vast majority of the potential of AI is in increasing the performance of existing use cases where other types of analytics could already be used," he told InformationWeek in an interview. "The core parts of the business are actually where the core benefits of AI would be."

But those effective, value-laden use cases aren't the ones you may envision after watching sci-fi films or even futuristic keynote addresses at technology events. They are more down to earth, or even boring, according to Chui.

For instance, if you work in a business where marketing and sales are a core part of the business you may think that having bots call customers and try to sell them things would be a good potential use case for AI in your business.

But Chui says that the place where you may find the biggest initial value is in a recommendation engine for customers -- just using AI to recommend to each customer the next product they may want to buy.

"You could increase the predictive power of that next product recommendation," Chui said. "You could make it better by using some of these AI techniques."

Better predictions and better recommendations translate to higher sales.

If your business is more focused on operational efficiencies that go to the bottom line, predictive maintenance may be the right use case for your organization.

The McKinsey Global Institute report notes that in 69% of the use cases studied, deep neural networks can be used to improve the performance beyond that provided by other analytic techniques. The cases in which only neural networks could be applied -- called "greenfield" cases by McKinsey, made up just 16% of the total. In the remaining 15%, neural networks provided limited additional performance over other analytics techniques, often due to data limitations that made them unsuitable for deep learning, McKinsey said in the report.

Data and data limitations are an important component of any AI implementation you may be planning, Chui said. Businesses that have been able to implement effective AI that delivers value tend to be businesses that are either digital natives or digital giants. What these kinds of company have in common is lots of quality data.

Organizations that have undergone a digital transformation may find that they are collecting more data, and that's an essential first step before you pursue initial AI projects.

"It's hard to leapfrog to AI without moving to digital first because you need a lot of training data," Chui said. And then once you train your model, you need to be able to communicate those results throughout your network in order to achieve any kind of benefits at scale, according to Chui. You need a digital connection to change behavior -- to change ow a service representative interacts with a customer.

"The more digitized a sector is, the easier it is for superior insights to get past the first mile problem," Chui said.

AI will touch many facets of your organization, so Chui didn't necessarily recommend coming up with an initial proof-of-concept use case. Rather, his recommendation is to review the opportunity against multiple dimensions and examine all the possible use cases in your organization. One of the first things to look at is all the types of data you have available to you.

That means your organization needs to get serious about data management and governance for the long term, Chui said.