Continuing my search for examples of Practical AI, I found a great one in a recent discussion with Vipin Gopal, chief data and analytics officer at Eli Lilly, the $28.31 billion (2021 revenues) pharmaceutical company. It’s practical because it demonstrates how the recent advancements in machine learning, specifically deep learning’s success at analyzing images, can be shown to quantifiably improve a traditional process, in this case in manufacturing.
Gopal reported that inspecting the quality of the syringes Lilly produces used to be done manually. The process was then automated using a vision system to find defective syringes. But now, with deep learning algorithms, the overall accuracy of the system improved, with about 20% of previously discarded syringes proven to be non-defective. This resulted in reduced downtime and did not even require replacing the existing hardware component of the system.
Another example of improving the efficiency and quality of operations Gopal mentioned was related to the core business of the company, drug development. It usually takes about ten years and $2.5 billion to bring a new drug to market. Aiming to cut down drug development time wherever possible, “we looked at various decisions that are made as part of the operationalization of clinical trials,” says Gopal. Which sites to work with in conducting the clinical trial proved to be a critical decision. It turned out Lilly had internal and external historical data on the execution of clinical trials by different sites, such as the time it took to enroll patients in the trial and the time it took to address protocol amendments. Analyzing the data on site performance led to improved site selection.
The third example from Lilly’s experience with AI has to do with the first step its research organization takes in developing a new drug. “AI enables model-driven drug discovery, to narrow down the funnel to a handful of molecules,” says Gopal. The data from previous drug discovery efforts and from previous clinical trials is analyzed by the new AI tools, speeding up to process of identifying the right molecules from which to create the required compounds.
With the new AI tools—and other initiatives—Lilly has managed to cut down drug development time by about three years. “Human intuition plus the power of machine learning coming together to produce the best solution,” Gopal sums up the experience.
AI is learning from data. And Practical AI is the successful, measurable, business use of learning from data. My discussion of these Practical AI examples with Gopal occurred a couple of weeks after he participated in the “How Artificial Intelligence Powers Digital Ecosystems” panel at the recent MIT CIO Symposium. The panel moderator was Tom Davenport, distinguished professor at Babson College, and the other panel participants were Michelle Hoiseth, chief data officer and senior vice president at Parexel (one of the largest clinical research organizations) and Marshall Van Alstyne, professor at Boston University.
Hoiseth described Parexel’s Practical AI examples which, somewhat similar to Gopal’s examples, fall into “three buckets”: Efficiencies driven by Robotic Process Automation (RPA); pattern detection—finding fraudulent activity in a study, for example; and predicting outcomes—using AI, for example, “to look across heterogenous population data to understand which patient will respond differently to therapeutic interventions so we can direct the study to different types of patients.”
The panel discussion was all about learning from data created inside and outside of enterprises. Davenport highlighted what he called “the virtuous circle of AI”—capturing more data from customers to improve the AI model which improves customer service and eases customer friction, leading to acquiring additional customers which provide more data.
More data—and more learning from data—comes from the creation of a digital ecosystem, the sum total of an enterprise’s business relations. Van Alstyne pointed out the “value of externalization,” the network effects of data created and used across a business ecosystem.
To prepare the enterprise for the abundance of valuable data that exists externally, a necessary step, said Gopal, is to “create an ecosystem within the company where data is flowing freely” and “data is managed consistently across the organization.” The challenge is “the tremendous amount of data that is resident in various parts of the organization, largely disconnected,” observed Gopal.
Hoiseth agreed that “to get our own internal data interoperable” is a “thankless task.” Employees create data for their own needs and don’t appreciate the need of other people in the company to use that data. To break open these data silos, the consumers of data must create value for the data creators, said Hoiseth.
Sharing data across the enterprise and integrating it with external data requires a “data-driven culture” that can be achieved by making all employees educated about data and analytics. Or as Gopal put it, “Getting the organization as a whole to be performing in an AI-enabled ecosystem.”