Artificial intelligence is coming to the enterprise, or so we're told. But what does enterprise AI actually refer...
to, and how do CIOs get a handle on how to incorporate AI technologies into their IT strategies?
In this SearchCIO Q&A, Tom Davenport, President's Distinguished Professor of Information Technology and Management at Babson College and well-known analytics expert, explains why he's sharpened his focus on enterprise AI and how CIOs can get started. Davenport, who was traveling internationally, provided his responses via email.
Let's start with definitions. When you talk about artificial intelligence, what are you talking about?
Tom Davenport: It's common to say that AI handles tasks that were previously only addressable by humans. But I think to exclude more traditional forms of automation, you also have to define it as performing tasks requiring a high level of expertise, insight or perception.
You've been in the analytics space for a long time. What made you turn your attention to enterprise AI?
Davenport: I found that while many companies are still doing a good bit of work in analytics and big data, they are less interested in hearing or reading about those topics. And to me, AI and cognitive technologies are a straightforward extension of analytics in most cases. Most of the models are statistical in nature, and analytical people are logical candidates to push AI forward in organizations.
Davenport: Most of the early adopters of deep learning were Silicon Valley companies with a history of making software available in open source versions. So that kind of discourages vendors from charging for proprietary deep learning software. There is a deep learning component (API) available in IBM Watson's collection of them, but it was only recently announced. Salesforce.com also announced that some deep learning capabilities will be available for image classification as part of its Einstein offerings.
CIOs should know that deep learning is very helpful for categorizing unstructured data -- typically images, speech or sounds. And unless the capability is already embedded in a system that you buy from a vendor, it will take some pretty sophisticated data science skills to work with it.
You recently wrote about how companies can bring AI into the enterprise. You suggested three avenues: mostly buy; mostly build; some buy, some build. How can CIOs determine which avenue is right for them?
Davenport: As of now, the 'mostly buy' options are largely from vendors that will either configure a solution for you (e.g., IBM Watson), or that have embedded AI capabilities into their existing software (e.g., CRM from Salesforce.com). Early adopters of Watson had to train it on each new application, but now there are some solutions that require less development. You can also mostly buy robotic process automation for back-office digital tasks -- they are easy to configure.
The 'some buy, some build' capabilities are basically either IT automation capabilities, chatbots for customer service, or some low-hanging fruit AI applications that do tasks like recommending products or services. I would also put statistical machine learning capabilities -- often used for highly segmented marketing models, for example -- in that category.
The 'mostly build' work involves open source libraries for machine and deep learning. And as I mentioned above, they often involve classification of unstructured data, like images.
At EmTech, you called Watson a "high price" option. But isn't it a cloud service? What makes it so expensive?
Davenport: It's true that all the Watson APIs are available through the Bluemix developer cloud, and that's a low-cost -- even free for a while -- way to access Watson. But customers tell me that IBM is mostly marketing Watson as a 'transformative' solution for big problems and issues in business and healthcare. That means some work upfront to identify the best place to use Watson, some solution architecture work to figure out what APIs are needed, and some consulting work to train the system and fit it to a business process. I think it often adds up to a high-cost, high-value approach. But as members of the Watson ecosystem use the APIs in their own solutions, I think that prices for some Watson services will drop. If you don't want to be the first in your industry to do something with Watson, it will be cheaper, too.
In your book Only Humans Need Apply, you describe a future workplace enhanced by "augmentation." How do you define augmentation, and how can CIOs begin to prepare for this future?
Davenport: Yes, we (my co-author Julia Kirby and I) believe that augmentation -- smart humans working alongside smart machines -- is both the most likely outcome and the best one for long-term success. If CIOs want to take the lead in introducing AI to their organizations, they should begin to identify which business processes have cognitive bottlenecks, need fast and accurate decisions, involve too much data for humans to analyze or now have human advice that is just too expensive.
Then, you identify the technologies that can help in that area and start a pilot or proof of concept. At the same time, you can begin to do some work design to think about what the machines should be doing and what the human workers should do. And then IT organizations should probably partner with HR to let employees know what kinds of skills are going to be available in the future. They need some time to prepare themselves for a future of working closely with smart machines.
Davenport at EmTech
Davenport on Keeping Up with the Quants
Davenport on Big Data at Work