The days of asking if your company needs artificial intelligence (AI) are over. The answer, across nearly every industry and spanning the globe, is a resounding yes.
The research firm Gartner last year estimated that global AI business value would reach $1.2 trillion by the end of 2018, up 70% from the previous year, and will more than triple by 2022. Taking advantage of greater computing power and developments in machine learning, business leaders are tapping the power of AI to enhance customer experience, create new revenue and reduce costs.
So the issue is no longer whether to adopt AI, but how to do it. And that’s a question of whether to build or to buy.
There’s a simple answer, and a more complicated one. For companies that need AI to power their core business or to ensure strategic success, building is the way to go. Think Uber and autonomous vehicles, or Netflix’s sophisticated recommendation engine. For the majority of business needs, such as improving non-core activities like human resources, finance and accounting or customer service, buying one of the many well-tested, off-the-shelf AI products is sufficient.
“That’s the high-level and basic answer,” says Thomas Malone, founding director of MIT’s Center for Collective Intelligence. “It’s based on the same factors that apply to any build-or-buy decision. It comes down to how strategic and unique to your company are your applications of AI likely to be?”
From there, it gets complicated.
Learn more about how companies are leveraging AI today.
The Decision Tree
Think of the choice as a standard decision tree involving a few essential questions. The first one concerns your chief objective: Do you want AI for a big, transformative “moon shot” that will define your company? Or do you want AI for the low-hanging fruit—fairly easy-to-accomplish enterprise applications that will deliver immediate value?
If the goal is an AI project that handles routine activities and delivers immediate value, it’s almost never a good idea to build your own AI, says Thomas Davenport, professor in management and information technology at Babson College. Even using open source tools, build-it-yourself AI can cost millions of dollars, and it can take months to train a machine learning algorithm to do what most vendors have already accomplished.
If the goal is an AI project that handles routine activities and delivers immediate value, it’s almost never a good idea to build your own AI, says Thomas Davenport, professor in management and information technology at Babson College.
Machine learning excels at repetitive back-office administrative tasks such as sniffing out redundant customer records or checking supplier invoices to verify shipments, and tools that do these things can be purchased ready-made. Davenport, who is also a senior adviser at Deloitte Analytics, found in a study of 152 AI projects that these were also the most successful.
Davenport cites the work of NASA, which has deployed robotic process automation (RPA) technologies to handle back-office tasks, streamlining work in the accounts payable, IT and human resources departments. The bots can also open email, paste and copy text, move file folders and screen job applicants. In the HR application alone, the bot proved it could handle 86% of transactions without a human involved. The agency is now rolling out similar bots across its organization.
Buying offers a host of other advantages. Vendors take over the tricky issues of integrating new AI applications into an existing IT environment and training workers to use the tools. They also offer specialized algorithms for tasks like image recognition, a massive chore that requires feeding tens of thousands of images into a program in order to teach it to recognize objects.
“It’s a cheaper solution than spending millions of dollars hiring data scientists,” says Davenport.
Once a company has chosen to buy, the next question is whether to purchase from a specialized AI vendor or to go with enterprise software enhanced with new AI abilities.
The advantage of going with a specialist vendor often comes down to cost and time savings. The vendor has already hired the data scientists and software engineers and has put them to work on the labor-intensive task of training up machine learning models.
Still, there are risks of going with a specialist, especially a newcomer, Davenport warns. Many AI companies are relatively new startups, and may not last. There are also questions about the amount of support they provide and what the annual costs might be.
Because of these uncertainties, many companies turn to their existing software vendors, who increasingly are adding AI capabilities to systems that their workers already use. Not only do such systems have access to massive data sets—often the buyer’s own data—they can also be used without requiring specialized know-how or additional training.
When Deloitte surveyed 1,100 mostly C-suite decision makers for its 2018 State of AI in the Enterprise, it found that 60% of them had chosen this type of vendor-supplied solution. Deloitte calls this “the easiest path: using enterprise software with AI ‘baked in.’”
One example is Salesforce Einstein, which offers predictive lead scoring that points salespeople to clients most likely to close on a deal. It even advises the best time of day for a sales call. It’s a good use of salespeople’s time, requires no training and can be paid for in one sale.
Deloitte predicts that such packaged AI services, programmed for specific job-related tasks, will continue to grow. Among such services Deloitte envisions are AI-powered cash-management services, video analysis for brand managers and trouble-ticket analysis in customer service.
Off-the-shelf AI can be the safer route, but its rewards can be limited, and it doesn’t offer the disruptive competitive advantage that AI has the potential to deliver.
What if an organization needs something more than a packaged AI tool but doesn’t want to splurge on creating their own AI system, with the required investment in hardware, software and expensive data-science talent?
There’s a hybrid solution: combine purchased AI technology with internal expertise. Open source AI tools, like Google’s TensorFlow or Microsoft’s Cognitive Toolkit, have made it easier and cheaper to build AI from scratch.
Cloud-based AI services also make it easier for companies that want to build their own AI tools. Companies like Google, Amazon and SAP offer a variety of ready-to-use AI tools, including customer service chatbots, speech recognition and image recognition software.Such cloud providers, Deloitte notes, offer AI as a pay-as-you-go model, allowing companies to rent the computing power needed to run machine-learning applications and, in some cases, access AI applications on the platforms. In Deloitte’s survey, 39% have chosen “to acquire advanced technologies such as AI through cloud-based services.” Only 15% say they have chosen to go the fully on-premise route.
The upshot? Buying or renting AI tools are the best options for most applications today, and are likely to become even more so as time goes on. Indeed, Deloitte forecasts that the need for companies to build their own AI will decline as more AI-enhanced software enters the market.
DIY Pros And Cons
Still, many organizations will need to build their own AI systems to accomplish their most challenging and disruptive objectives—their moon shots. Building in-house offers greater flexibility, and gives the creator ownership over the intellectual property that can come from building something entirely new.
Building in-house offers greater flexibility, and gives the creator ownership over the intellectual property that can come from building something entirely new.
The risks are greater as well, and sometimes the hoped-for value doesn’t materialize.
Davenport points to the case of MD Anderson Cancer Center, which spent $62 million between 2013 and 2017 to develop a way to use IBM’s Watson AI supercomputer to diagnose and recommend treatment for certain cancer patients. The project was eventually put on hold without ever being used on a patient.
Meanwhile, Anderson found success using AI in another way. With customer-engagement tools from CognitiveScale, the center helps visiting families make hotel reservations and identifies patients who need help paying bills. The projects improve patient satisfaction and the center’s financial performance.
“AI is not great yet at breakthroughs,” Davenport says. On the other hand, he adds, “A really hard problem, like cancer, no one else has solved. So you really don’t have much choice but to do it yourself.”
Learn more about how companies are leveraging AI today.
