AI Advice for Enterprises: Avoid Moonshots, Focus on Low-Hanging Fruit | Straight Talk


The latest insights from your peers on the latest in Enterprise IT, straight to your inbox.

A Conversation with Thomas Davenport

Business executives are increasingly investing in the potential of artificial intelligence to accelerate their digital transformation initiatives. A Deloitte survey of 1,100 IT and business executives, conducted in the US in late 2018, found that 37% have invested $5 million or more in AI technologies. “Early adopters are ramping up their AI investments, launching more initiatives, and getting positive returns,” concluded Deloitte’s report on the survey, State of AI in the Enterprise, 2nd Edition.

One of the authors of the report, Tom Davenport, is the President's Distinguished Professor of IT and Management at Babson College and a Digital Fellow at the MIT Initiative on the Digital Economy, in addition to acting as an independent Senior Advisor to Deloitte's Analytics and Cognitive practice. Davenport has been directly involved in the evolution of enterprise data and its analysis, from the development of early “expert systems” in the 1980s to automated decision making and knowledge management starting in the late 1990s to documenting over the last fifteen years the success of data-driven businesses and their big data analytics initiatives. He has published numerous influential articles and 18 books, most recently The AI Advantage: How to Put the Artificial Intelligence Revolution to Work, just out in a paperback edition from MIT Press.

What is the state of AI in the enterprise in mid-2019? Gil Press, a CTO Straight Talk Contributing Writer talked recently to Davenport about the present condition and future prospects of the adoption of AI technologies by businesses worldwide.

Let’s start by defining what exactly we are talking about when we use the term “artificial intelligence” today.

AI employs knowledge, insight and perception to solve narrowly defined tasks. We typically include in it activities that heretofore were done only with human brains. It’s not a perfect definition because AI also includes activities that go beyond the capabilities of human brains, like large scale data analysis. There are several underlying technologies that are typically included in discussions of AI, ranging from robotic process automation, or RPA, which at least up to now hasn’t been terribly intelligent, to very advanced methods of statistical analysis such as machine learning and its most complex form, deep learning.

Where in the enterprise do you see these AI technologies deployed today?

AI technologies are making an impact in all facets of the business today, externally and internally. For example, in the area of customer relations, enterprises are trying to improve their interactions with customers by using conversational AI. Capital One offers its customers Eno, an intelligent assistant that helps with such tasks as fraud alerts and balance inquiries. Telefonica is using conversational AI in call centers to interact with customers calling to ask about their balance in their pre-paid accounts or to make sure they have international service—mundane but useful-to-automate kind of activities.

What about internal functions and activities?

IT is the most active user of AI. In the surveys I’ve done with Deloitte, there is a higher percentage of IT departments using AI than any other function. Now people call it “AIOps,” automating activities such as rebooting servers, supplying people with their passwords, even software testing. It’s really just more RPA being applied to IT, but RPA that is slowly getting smarter.

RPA, which used to be the dumbest form of AI, is getting more intelligent mostly when it’s integrated with automated machine learning. This combination of technologies is ushering in an era of many intelligent automation applications. It’s particularly useful for structured, information-intensive “swivel-chair” tasks such as pulling information out of one system and putting it in another.

Improving and optimizing internal processes have been the focus of operations research for quite some time.

Indeed, but it has always been done on a relatively small scale, using individual models with only a few variables. Advanced machine learning helps companies today in taking this to the next level in both breadth and depth. A good example is the steel manufacturer startup Big River Steel. Working with AI consulting firm, Big River has successfully applied machine learning in six areas: Demand prediction, sourcing and inventory management, scheduling optimization, production optimization, predictive maintenance, and outbound transportation optimization.

Most promising for Big River—and for other companies—is the significant business benefit that comes from integrating these applications of machine learning, creating an “end to end” optimization of the performance and profitability of the firm.

Have you seen any surprising use of AI by an internal function?

Yes, in Human Resources. A function that might have been considered a laggard in terms of the use of technology, especially a technology like AI that is just emerging, is doing a lot of interesting things with it.

Looking at the results of a survey I conducted with Oracle, I was most surprised by how much usage of AI there is in HR. When asked which types of analytics they were using, 31% made “artificial intelligence” their first choice, and 58% put it in their top three choices.

What exactly is HR doing with AI?

A variety of applications and uses: identifying talented individuals who are prime candidates for leaving the organization, predicting which job applicants are likely to be high performers, developing workforce plans, finding best-fit candidates with resume analysis.

You have argued for some time that AI is largely an analytical technology and that, for most organizations, it is a straightforward extension of what they do with data and analytics.

Machine learning in particular is a highly statistical activity. Predictive analytics is very similar in the sense of predicting an outcome when you don’t know it. Lots of companies I talk to about AI are saying “we have been doing it for years, we just didn’t call it machine learning, we called it analytics or statistical analysis.” But what we see today in what we call “AI” are more complex types of algorithms and methods such as deep learning that people didn’t use in the past.

Another recently much-improved AI method, NLP, or Natural Language Processing, some of which is statistical and some of which is semantic in nature, we didn’t do before under the analytics banner. All of these new approaches are challenging but we are making good progress.

Thinking about AI as an extension of analytics must also help enterprises meet the challenge of the AI skills shortage.

I say that a lot. It’s why you want to combine your AI center of excellence with your analytics staff. Much of the work for both activities—AI and analytics—is getting the data ready and prepared and a lot of the technologies and tools are the same. But for some, the idea of AI as an extension of analytics is “hype deflating.” People want to pretend it’s something entirely new but it’s not.

What are other key challenges that companies implementing AI typically encounter?

The big challenge is scaling AI, getting it to production deployment. Usually it’s things like retraining people, work redesign, integrating AI with the current technology architecture. All of these tend to slow down implementation, which is why you don’t see widespread production deployment today. It’s similar to what happens with analytics in many organizations, where lots of models are developed but not deployed. I’ve been telling companies to calculate their deployment score—what percentage of the models get deployed. It’s worse with AI, because it’s getting easier to generate the models but maybe even harder to deploy them fully.

Is the culture of the organization a big factor contributing to the difficulties in deployment you have observed?

People say “culture” but I don’t think that this umbrella term really helps. The real issues are that we haven’t done enough with employees to prepare them for AI, we haven’t thought about what their job would look like in the near future when it’s assisted by AI, and who would be a good fit for working in these new environments.

In a global survey I did with ServiceNow, the workers were not that concerned about AI taking their jobs, but they were concerned about not getting re-trained and nobody was telling them what their job was going to be like in the future. Companies need to think about these issues—how to prepare people for jobs that involve working closely with machines, including when the machine is in a supervisory role. I recently talked to Farmers Insurance and they have an AI system that tells call center representatives whether they are using simple language that’s easy for customers to understand and act on. As usual, it’s the people on the front line that get added supervision from new technologies, not senior executives.

Another challenge often mentioned is companies not having enough data to feed the AI algorithms.

Yes, this is an issue, especially with deep learning, which relies on lots of data to develop and train the models. I think this will lead to the creation of data clubs or associations. You are starting to see different groups come together to combine their data sets in specific domains. For example, radiology images where a group of hospitals combine their data to use it for training their deep learning models.

Given all these challenges, what’s your number one advice to enterprises considering or experimenting with AI?

Avoid moonshots and focus on the low-hanging fruit. Singapore-based DBS Bank, the largest bank in Southeast Asia, tried to use IBM Watson to make investment recommendations to the bank’s relationships managers and their customers. They failed. The technology wasn’t quite ready for this ambitious task. But they had great success with applying AI to smaller tasks such as knowing when the ATM is going to run out of cash, predicting the churn of salespeople, and detecting fraud in trading. Each one of these is not dramatic, but if you combine them in a specific area of the business they could end up being transformational.

What do you see happening with enterprise AI in the near future?

We will see AI applied in the context of more strategic questions—how do we plan, how do we match demand and supply, the type of issues and decisions that are not highly repetitive. It has been easier to start by applying AI to address and improve repetitive tasks but now it’s time to tackle less frequently performed but more strategic questions.

In addition, a lot of the AI technologies that are now stand-alone are going to be merged and almost all applications will be a combination of them. RPA is an example—it is increasingly going to be combined with deep learning. There will be a lot of AI pieces to choose from and they will be integrated. A lot of companies will find it appealing to buy AI capabilities from their existing vendors—Salesforce is betting on it in a big way. Users will not think about AI as a new and separate application because it will be integrated with other enterprise software.