The latest insights from your peers on the latest in Enterprise IT, straight to your inbox.
Putting Artificial Intelligence to work in the Enterprise
Nine companies share their experiences
Read all about the business promise and benefits of AI along with an insight into the ongoing progress in putting Artificial Intelligence to work in the enterprise, only on Straight Talk.
We’ve been here before.
In 1970, artificial intelligence pioneer Marvin Minsky said in a Life magazine article he was certain that within eight years, and maybe as few as three, “we will have a machine with the general intelligence of an average human being.”
Despite his optimism and years of encouraging progress, artificial intelligence research foundered. It was another 27 years before IBM’s computer program Deep Blue defeated reigning world champion Garry Kasparov in a chess tournament. And it wasn’t until 2004 that a competition showcased the capabilities of autonomous vehicles—though none of the driverless cars finished the 150-mile route through the safely remote Mojave Desert.
Now the buzz around artificial intelligence and the prospect of machines taking over human tasks has again reached a fever pitch.
Virtual assistants, from Apple’s Siri to Microsoft’s Cortana to Amazon’s Alexa helpfully respond (well, most of the time) to our commands. At this year’s CES global consumer electronics show in Las Vegas, attendees could converse with chatbots embedded in lamps (GE), robot nannies (Mattel), vacuum cleaners (Samsung), and cars (Ford).
But artificial intelligence is about much more than consumer products. AI, including associated technologies such as machine learning and deep learning, promises to transform business by making organizations both more efficient and more innovative.
As in the past, the excitement about AI’s potential is accompanied by trepidation about its consequences, particularly its uncertain effect on people’s jobs. In newspaper articles, policy statements, and conference discussions, AI is once again portrayed as either savior or destroyer of humanity. Garry Kasparov, who accused Deep Blue of cheating when it defeated him in 1997, wrote recently in the Wall Street Journal that intelligent machines, disruptive as they may be, “are not a threat to humankind but a great boon, providing us with endless opportunities to extend our capabilities and improve our lives.”
So is the buzz justified this time around? Given that the term was coined in 1955, and that the field has since then gone through multiple periods of optimistic anticipation followed by deep disappointment and vanishing interest, what accounts for today’s flurry of interest?
In a word: data.
REALIZING THE BUSINESS PROMISE OF AI
Artificial intelligence is a subfield of computer science in which computers perform tasks that traditionally have been considered intrinsic aspects of human intelligence—speaking, reading text, identifying objects, learning from experience, predicting future actions.
So-called “general AI” refers to the idea of computers possessing broad human-like intelligence. “Narrow AI” is where we are now, with computers performing one cognitive task or a limited combination of cognitive tasks.
Machine learning is an application of narrow AI. While computers are typically told what to do by programmers developing clear and specific rules, spelled out in computer programs, the machine learning approach relies on the computer’s statistical analysis of data to find patterns in it or to categorize the data into different buckets. This enables the computer to “learn”(for example, continually improve the performance of a certain task or the accuracy of a result) and “predict” (for example, anticipate the nature of new data that is fed to it).
The machine learning expert, one who is an expert in both computing and a specific business domain (typically known today as a “data scientist”), develops models and algorithms for training the computer on a specific data set. Over the years, machine learning has been applied successfully to such tasks as identifying spam emails, hand-writing recognition, language translation, fraud detection, and product recommendations.
As machine learning has evolved it has become less dependent on specific, domain-based models, broadening the range of practical applications that computers can learn and perform. In so-called “deep learning” vast amounts of data are moved through many layers of hardware and software arranged as an “artificial neural network” (so-called because it is influenced by a computational model of neural networks in the human brain), with each layer coming up with its own representation of the data and passing what it “learned” to the next layer.
So what does this all mean for business? Well, web-native companies such as Google, Amazon, and Netflix have built their fortunes with the help of modern machine learning algorithms. With the Internet-enabled explosion of available data, these companies have proved how machine learning can use lots of data from a variety of sources to, for example, predict consumer behavior. Using lots and lots of data makes predictive models even more robust and predictions even more accurate.
But the business value of artificial intelligence and machine learning, isn’t limited to companies born on the Web and the really big data they are able to collect. All enterprises in all industries can increasingly realize the data-driven benefits of modern machine learning because of two further technology trends: cloud computing and open-source software.
BROADENING THE BUSINESS BENEFITS OF AI
The rapid adoption of cloud computing gives enterprises access to newly created pools of big data by breaking down an enterprise’s traditional data silos. The cloud, as a shared computing and storage resource, allows employees from different functions and departments to benefit from all the data, private and public, that is available to the enterprise.
The cloud also reduces the cost of processing and storing data, allowing enterprises to accumulate more data and to do more with the data they have accumulated. The providers of public cloud resources, such as Amazon, Microsoft and Google, have also made machine learning services an important component of their services.
In addition, these companies and other leading users of modern machine learning, in order to attract scarce talent, create an academic-like environment where AI theories can be tested and the results shared with the world through conference papers, academic journals, and blog posts.
More important, they frequently share open source software that can be used by both newcomers and experienced machine learning practitioners, adding to the extensive library of machine learning tools long supported by the open software community.
The availability of tools, resources, and talent have enabled many enterprises to take advantage of the most recent advances in machine learning and apply its power to a wide range of activities, interactions, and transactions.
Here are some ways that enterprises are leveraging modern machine learning to improve operations, increase productivity, and delight customers.
PREDICTING INFRASTRUCTURE NEEDS
“Machine learning is a safety net,” says Catalyst Paper CIO Paul Einarson. “It helps me, as a CIO, sleep better at night.” Why? “It’s difficult for humans to predict,” says Einarson—in his case, how much the amount of data generated and consumed by Catalyst Paper, one of North America's largest producers of mechanical printing paper, will grow over the next few months. It’s something that the IT staff needs to get right in order to ensure the availability of adequate computing and storage capacity. Machine learning algorithms, monitoring and analyzing consumption patterns, alert Einarson when he needs to add capacity.
With data growing by leaps and bounds, accurately forecasting the needs of the IT infrastructure is crucial to its optimal management—to say nothing about improving the sleeping patterns of CIOs. This is even more challenging when your IT infrastructure serves millions of consumers and businesses. The amount of data flowing over AT&T’s network has grown by 150,000% since 2007, according to Mazin Gilbert, Vice President of Advanced Technology at AT&T Labs.
“Our network has been put together over the past century using a lot of different technologies,” says Gilbert. “A few years ago, we started to transform it into a pure software-defined network that rides on a commodity cloud hardware.” The software-defined network allows AT&T to embed intelligence within the network, and create new infrastructure management applications overnight.
Just as machine learning tools allow Catalyst Paper to predict future required capacity, they help AT&T predict a business customer’s desired bandwidth, based on usage patterns. In addition, the software learns the patterns of normal activity on the network, helping it detect and respond to cyberattacks before they become a major issue. “We call it closed-loop automation,” says Gilbert. “These AI software agents are continuously monitoring and taking action.”
In a manufacturing environment where data is collected and analyzed, “an automation system can self-learn or self-tune and provide alarms—‘I’m going to fail, come and fix me’,” says Srinivas Nidamarthy, CTO at ABB Robotics Systems. In fact, he says, machine learning has given rise to an entire service industry on top of manufacturing automation, one that provides monitoring and proactive equipment maintenance.
Prith Banerjee, until recently Executive Vice President and CTO at Schneider Electric, has used machine learning for proactive maintenance on the systems it maintains for its customers. “Instead of waiting to replace an asset after it has failed,” says Banerjee, “we are using machine learning algorithms to predict when assets such as drives, PLCs, breakers, or transformers will fail.” Rather than waiting for a customer to report a failure, the equipment “calls home” with an alert about an emerging or imminent issue. Being able to predict when an asset will fail, “we can send a replacement part and a repair person before that happens.”
IMPROVING DECISION MAKING
Embedding AI technologies in business processes may provide a solution to a problem faced by most organizations: Knowledge of how to perform a certain task or make a specific decision walks out the door every time an employee moves to another company or retires. Even when this tacit knowledge is captured, codified, and stored in a database, it may not be accessible to the people who need it, when they need it.
Ibrahim Gokcen, Chief Digital Officer at Maersk Group, owner of the world’s largest container shipping business, says that, with the help of machine learning, it is possible to capture in software the knowledge people acquire on the job and enable employees who do not have the right experience, “to have a head start in making a decision instead of starting from zero.”
The digital retention of organizational memory provides the context for decision making that often times is missing from the real-time data at hand. “We use AI technology to make the best possible decision at a specific point in time, using all the data available to the employee,” says Gokcen.
Beyond decisions regarding businesses processes and other internal operations, modern machine learning can support decisions concerning the enterprise’s external environment. With businesses such as Manheim, an automobile auction business, Autotrader, an online marketplace for new and used cars, and Kelley Blue Book, a vehicle valuation and research organization, Cox Automotive captures massive amounts of data about the car buying and selling business. “We have all of this digital exhaust over the lifecycle of a car, up to 2 terabytes a day,” says Steven Hatch, manager of enterprise logging at Cox Automotive. “Machine learning can draw a picture and analyze events, trends and activities, providing insights we couldn’t even imagine before.”
MAKING CUSTOMERS HAPPY
Mazin Gilbert joined Bell Labs in 1990 and four years later published a book on using artificial neural networks for speech recognition. “How to get AI to speak like a human?” is the way he describes the focus of his research during that period. Now, as the head of AT&T Labs (the successor to Bell Labs), he applies state-of-theart speech recognition and other AI tools to help AT&T “stitch together the customer journey.”
AT&T customers today reach out to it through many channels—retail stores, phones, email, social media. Analyzing data from across these channels helps AT&T pinpoint hot issues and determine who can address those issues and how. This understanding is especially important when a customer contacts the call center, as customer satisfaction “goes down significantly if we don’t address the issue with the first contact,” says Gilbert. AI technologies help the agents at the call center understand the customer’s issue, identify the best agent to address the problem, and select which channel should be used to respond—this, as an alternative to transferring the customer from one agent to another. “We are not trying to force a bot on our customers,” says Gilbert, “we are trying to expedite the resolution of a problem.”
AI tools help AT&T avoid “resetting the clock every day at the contact center. Instead, every minute of the day we are learning so that with every new transaction we can react faster and better.” Adds Gilbert: “Going forward, we would like to be able to predict what customers will do next, predict their issues so we can fix them before they contact us.”
Serving customers today often means interacting through a dizzying array of channels. “We have 100,000 mentions a week on social media,” says Tjalling Smit, Senior Vice President of Digital at KLM Royal Dutch Airlines. And the increase in interactions with customers through social media, which simply lowered the barrier for customers to interact with the company, hasn’t been offset by a decrease in phone or in-person interactions with customers. “We handle around 15,000 customer service cases a week, and we answer our customers 24/7 in 10 different languages,” Smit says.
To handle this daunting growth in the volume of customer interactions, KLM is using AI and machine learning to help its customer service agents respond to routine in-person requests and inquiries. In many situations, AI-powered software will suggest to the KLM agent the right response to a customer’s query. Instead of automating customer interactions, AI is actually “a facilitator for the continued human interaction between our brand and our customers, Smit says. In the future, “we are going to also implement AI in the call center. We will use it to create a personalized and relevant experience for an individual customer.”
MAKING REGULATORS HAPPY
Founded in 1899, Equifax is the oldest of the three main credit reporting agencies in the U.S. It collects and analyzes data on more than 820 million consumers worldwide, providing its customers with detailed reports on the personal credit and payment history of individuals.
“Several years ago,” says Peter Maynard, Senior Vice President of Enterprise Analytics at Equifax, “we noticed that we were not getting [adequate performance] from our traditional credit scoring methodology.” But there didn’t seem to be a viable alternative to traditional machine learning approaches such as logistical regression. To pass regulatory scrutiny, the results needed to be interpretable—that is, understanding how the learning machine reached its results. And more advanced machine learning approaches, based on artificial neural networks and that promised more accurate results, were problematic because they weren’t interpretable. In fact, artificial neural networks were considered a “black box,” based on a process so complex that even their programmers didn’t fully understand how the learning machine reached the results it produced.
Maynard and his team decided to challenge the conventional wisdom. They developed a revolutionary mathematical proof that made the output of their neural network interpretable for regulatory purposes. “We stripped apart the black box,” says Maynard. The new and transparent neural net has improved the predictive ability of Equifax’s credit scoring model by up to 15%, according to Maynard. And the larger the data set analyzed and the more complex the analysis, the greater the improvement.
The immediate benefit of using neural nets is faster development of the scoring models that identify detailed segments of the population and weights the relative importance of each. Instead of a data scientist building, testing, and manually refining a model over time, this can be done automatically. “It’s similar to search optimization or product recommendations,” Maynard says, “where the model gets tweaked every time you click. In addition, the “attributes” affecting a credit score—for example, the size of an individual’s checking account balance and how it has changed over the previous 6 months—are created based on massive amounts of historical data rather than the hypotheses of data scientists. “Instead of creating thousands of attributes, we can create hundreds of thousands of attributes for testing, with algorithms determining which are the most predictive in terms of the behavior we are trying to model.”
But besides streamlining the process while ensuring regulatory requirements are met, machine learning has produced another positive result: improved access to credit. Analyzing two years’ worth of U.S. mortgage data, Equifax determined that numerous declined loans could have been made safely, promising an expansion in the number of approved mortgages. Similarly, there was good news for consumers who in the past would have had to make a down payment to get a cell phone. “With this model, they don’t need to do that anymore,” Maynard says.
Online advertising is rapidly growing, financing free Internet content and applications and trillions of free gigabytes of data storage for users. To help publishers monetize their content and advertisers get their messages efficiently to more than 3 billion Internet users worldwide, a number of companies offer advertising management tools.
One of these companies is AOL Platforms, a unit of Internet pioneer AOL, now a subsidiary of Verizon Communications. It started building its data science team more than fifteen years ago and has continuously evolved its skills and expertise as required by the rapid changes in advertising formats, practices, and consumption patterns. “The world is always changing,” says Seth Demsey, CTO at AOL Platforms. “Having a strong connection between the content and the consumer, no matter what the channel, and having timely and relevant messaging is key. And you need to use the right tool for the right job. Today, the shiny object is deep neural networks but we use a variety of AI tools to get the job done.”
One area where AOL Platforms employs neural networks is the prediction of demographics for consumers, allowing AOL’s advertisers to better target their ads, according to Rob Luenberger, Chief Scientist and Senior Vice President, Research and Development. “Based on data provided by consumers who have opted in, we have some demographics information, the history of the web sites they visited, zip codes, etc. Using neural networks, we can then predict, for example, the chance of a person falling into a certain age bucket,” he says. “And we’ve found that neural networks perform better than more traditional machine learning models such as decision trees,” the multi-layered analysis of neural networks proving more effective than decision trees’ sequential analysis of decisions and their possible consequences.
The data scientists at AOL Platforms use another AI tool, natural language processing, to automatically analyze the content of a Web page and even assess its mood and tone, helping to match advertising with web content. AI-driven natural language interfaces are also used to automate the answers to questions AOL Platforms’ customers have about their advertising campaigns—for example, queries about the relationship between the number of views that a particular person has of a particular ad and the chance of that person clicking on the ad. The tool “understands a question phrased in a number of different ways and to some extent it understands the context of the content,” Luenberger says. “If I ask how many impressions campaign x got yesterday, it gives the answer. And then if I ask how many clicks, it understands that I’m still asking about the same campaign.” Adds Demsey: “It’s like a natural language BI assistant.”
INTRODUCING AI TO THE ENTERPRISE
Artificial intelligence has arrived.
As the various uses in the enterprise described here make clear, though, AI is far from intelligence that is equal or superior to human intelligence but rather enables practical applications, big and small, in a wide variety of industries.
As with any other new-new thing, technology executives must be aware of potential pitfalls and be prepared for the difficulties of—and possible organizational resistance to—adopting new tools and becoming accustomed to new practices.
A first step in putting the power of artificial intelligence to work in your enterprise is to identify specific activities that could benefit from the predictive power of modern machine learning tools. Look for “the thin edge of the wedge of AI in your business,” recommends Catalyst Paper’s Einarson, opportunities to demonstrate immediate impact and a clear ROI for AI.
“Look for real-world examples where the data is of suitable complexity and a human is likely to miss trends,” says Einarson.
Similarly, AT&T’s Gilbert advises looking first at the places where your organization “does things inefficiently.” Machine learning “is not a solution to all problems,” he warns. “But for repeatable problems, there is a huge opportunity for machine learning and AI.” As we have seen, data—and the quantity of data—plays a major role in ensuring a significant return on investment. Says Prith Banerjee: “The challenges have always been training the machine learning algorithms. The more real data you have available, the better the accuracy.”
The challenge for an industry like Schneider Electric’s is that failures are rare, and it is difficult to train machine learning algorithms with very few samples.
Furthermore, besides challenges concerning the quantity of data, there are challenges involving data quality. And the more data you have available, the more vigilant you must be to ensure that the analysis and predictions are based on valid observations.
Another challenge involves the fundamental differences between the familiar—developing traditional software—and the new—developing and managing AI applications. For example, debugging is harder because it’s difficult to isolate a bug in a machine learning program. And unlike traditional software, when you change anything, you end up changing everything.
Most important, the trove of tools and tested processes that have been accumulated throughout the years for software development does not exist for modern machine learning.
Learning from the experience of others and staying up-to-date regarding the latest developments in the practice of machine learning is crucial.
In many enterprises, this expertise exists in the analytics and data science team, so that’s a natural place to incubate and drive early AI initiatives. Depending on the organization, it could be useful to select a senior member of that team to act as a “Chief AI Executive,” with enterprise-wide responsibilities for introducing modern machine learning methods.
As always, your people are the most important element in ensuring the successful introduction of new tools and practices.
And that doesn’t mean only the people managing them but also the people on the receiving end, the employees that must adjust the way they work and understand the potential benefits of using the new tools. Given the bad rap AI sometimes gets in the press—and the ominous-sounding, human-replacing overtones of “artificial intelligence”—carefully introducing it in the organization is even more important than it typically is with other new technologies.
The emphasis should be on “augmentation” of human capabilities not “automation” of them, as well as on the creation of new roles and responsibilities that come with the adoption of AI technologies. Says Maersk’s Gokcen: “In the industrial revolution, certain jobs ceased to exist but other jobs were created. The same thing is happening with the Internet revolution. We will see an impact on certain roles but also the creation of new high-impact jobs and the creation of a lot more value from existing work.”
At AT&T, the focus is on bringing human experts and AI agents together “in a hybrid model,” says Mazin Gilbert, and taking advantage of what each does best: “Humans are very good at understanding context. Machines are good at processing lots of data and finding patterns in it.” Similarly, at KLM, “we don’t replace human agents with technology—we use technology to facilitate the dialog with our customers,” says Smit. “Human interaction will always be key.”
As we succeed in computerizing cognitive capabilities, computers will continue to augment humans, as they have done for more than sixty years. For enterprises, this means working smarter and finding new ways to thrive.