Adoption doesn’t always equal success, however. A 2020 Gartner study found that only 53% of AI proof-of-concept projects make it into production. One common challenge is that it’s easy to get overwhelmed with the scope and complexity of AI initiatives. That’s partly a result of misunderstanding exactly what AI and ML can do for your organization, according to Scott Tease, vice president and general manager for High Performance Computing and AI at Lenovo. “When AI is done properly, you don’t call it AI,” he says. “It’s so integrated into the offering and solution, it just becomes part of the value-add.”
In that regard, AI can be viewed as just another tool, like a hammer or a screwdriver or a calculator, Tease says. This approach puts an AI project in the context of achievable outcomes, rather than “moon shot” initiatives that tend to overpromise and under-deliver.
AI pioneers recommend starting with discrete projects designed to enhance existing functionality or optimize existing business operations rather than attempting to overhaul the entire organization. For Tease, that means starting with an application for which you can easily measure results.
Getting started with AI and ML
“One of the biggest missteps we see is that customers don’t start with a good KPI measurement of what success looks like based on what they’re actually trying to do,” Tease says. “A lot of times, we see people deploying AI just for AI’s sake.” That might result in an interesting project that doesn’t return value, he explains.
Tease recommends starting a project not with AI as the focus but with the business value one hopes to achieve. “If AI is part of that, fantastic,” he says. “If it’s not, you can find other ways to achieve that same goal.”
From there, Daryl Cromer, chief technology officer for PCs and Smart Devices at Lenovo, recommends focusing more on the data that AI needs to get results, rather than on the AI tools themselves. “You can find algorithms,” he says. “There are a lot of tools that allow you to get a good start. The data set is a little more complicated.” Consequently, Cromer’s group at Lenovo spends more time curating data and training ML models than writing software to make use of it.
“You do not have to create all of this IP and all these algorithms yourself,” Tease affirms. “There are thousands of startups that have unique and interesting IP to go do certain tasks.” In the area of retail alone, for example, off-the-shelf, AI-powered software can help in areas such as inventory management, theft minimization, self-checkout, improvements to prevent missed scans, and more, he says.
To get good data for AI and ML to work with, Tease recommends following established practices around data management, including tearing down data silos to foster
data sharing across the organization, identifying the right data governance framework for your organization, and engaging data stakeholders early on.
As for what kinds of projects to start with, Cromer suggests chatbots, for one.
Use cases for AI
Chatbots are a widely adopted use case for AI and ML, and part of the reason Gartner predicts that emerging technologies will drive 70% of customer interactions by 2022.
They’re also a place to gain some quick wins with maximum return on investment, according to Cromer. That’s because they encompass a relatively limited domain, where developers can predict questions and gauge answers with relative ease, helping them measure success. “A chatbot allows you to move forward and build a data set more easily because you have a good frame of reference,” he explains.
Supply chain management is another potential area for AI and ML to return high value, especially for manufacturers, according to Cromer.
For example, Lenovo has to predict demand for millions of computers around the world and orchestrate component supplies as needed. AI helps the company’s supply chain managers do exactly that. “It seems fairly complex given the number of machines and different pieces,” Cromer says. “But AI can do a really good job of helping provide that guidance.” What’s more, the system Cromer’s colleagues developed provides daily feedback on how it’s working, allowing them to adjust the data set or algorithm on the fly. “We like to see things like that, where there’s a clear goal, a measurable outcome that you can use to tune the data very quickly.”
Tease points to predictive maintenance as another task that’s ideal for AI. For example, data on temperature, changes over time in storage devices’ read/write speeds, and even fluctuations in electrical current can help AI predict when equipment such as servers in a data center will fail. “Having a tool look for trends that predict that this fan or this power supply or this hard drive might be about to go allows you to take action proactively,” he says.
AI can also help product developers and marketers get a read on customer sentiment. Cromer cites as an example a system at his company that scans millions of comments on social media sites and other sources for intelligence on customers’ likes and dislikes. “We have a database of over 50 million comments,” Cromer says. “Once a month, we go through all this data and look to see the trends. It’s a really good way to get unfiltered feedback from our customers.”
The importance of ethical AI
With all the good AI and ML can do, experts caution that if misused, these tools, like any other, have the potential for harm, for example, through biases that favor one group of people over another. That puts a responsibility on AI developers to create systems that don’t operate as black boxes making decisions without accountability, Tease says.
“We need to be able to explain to people how the AI is operating so that it’s doing things that are just and legit,” Tease says. “Machines may incorporate biases that we’re not aware of. So, doing things in a responsible, explainable, repeatable fashion is critical.”