Enterprises spanning virtually all industries and markets are jumping onto the artificial intelligence bandwagon. Yet as they climb on board for a ride into the future they soon discover that stock AI tools aren't capable of meeting their exact business needs.
However, organizations don’t need to spend large amounts of time and money building their own AI tools to fix this problem, suggested Kevin McMahon, executive director of mobile and emerging technologies at digital tech consultancy SPR. All they have to do, he noted in an interview, is to customize already existing tools to fit their exact business needs.
To be effective, AI software must be tailored to address specific business problems, and this is where generic algorithms often fall short, said Sanjay Srivastava, chief digital officer at professional services firm Genpact. An AI solution that reports adverse events related to drug safety, for instance, is different from one that predicts risk in a lending portfolio. "Both can use the same core AI platform, but they require different datasets, semantic understanding, process knowledge, and associated ontologies," he explained. "It’s all about training the AI the right way to solve a particular business problem."
This can be a challenge, however, for organizations that lack access to a large set of industry-specific data. "Without relevant data or context to fine-tune algorithms, organizations will struggle with, and sometimes fail, when implementing AI solutions," Srivastava observed.
Even if an organization manages to gain access to the necessary data, it will likely face a long and complex development process. "In most cases, unless it is a simple input/output prediction task, it can take many weeks if not months," noted Dokyun Lee, an assistant professor of business analytics at Carnegie Mellon University's Tepper School of Business.
Enterprises beginning their AI journeys often rely on the services of the software provider or an AI development company to provide necessary customization. Some organizations, however, attempt to tackle the work in house, often with mixed results. "Having internal AI capability -– a combination of talent, platforms, tools, knowledge, relationship, and data -– offers the option of doing it internally versus outsourcing," said Monika Wilczak, an advisory managing director in artificial intelligence at business services advisory EY. "The stronger the internal AI capability, and more mature the enterprise is around the application of AI as a strategy for growth, the more likely it is to use their own data scientists and application engineers for customization," she explained.
Still, even enterprises with full-fledged AI development teams can find customization to be an expensive and time-consuming undertaking. "Customization of vendors’ AI products requires data class inclusiveness, controls to avoid data bias, and the availability of a sufficient volume of labeled data," Wilczak said. "There's also still a need to understand the business challenge at a high level of granularity in the context of the existing technology," she noted. Developers also need to pay close attention to business and human elements "to not only build a successful prototype, but to ensure it can be operationalized in adherence with requirements to activate value realization."
AI customization should be a collaborative process. "While the client needs to delineate their business challenges, they also need a partner that understands the data science component and has domain expertise," Srivastava said. A service provider that has access to a broad base of industry datasets is best positioned to provide an optimal solution, he noted. On the customer side, "bilingual' staffers -- individuals fluent in both data science and industry needs and goals -- are the best people to train and fine tune algorithms to deliver maximum value, he observed.
It's also important to remember that unlike many types of conventional enterprise software, sophisticated AI applications aren't "set it forget it" products. "Supporting deployed AI solutions requires the use of diagnostics as well as data scientist intervention to ensure the performance and stability of machine learning data science models," explained Nate Regimbal, senior manager, innovation and digital transformation, at professional services firm Grant Thornton.
The pitfall that most frequently complicates or derails AI development efforts, CMU's Lee said, is a lack of vendor transparency and auditability. If the algorithms offered by the vendor are black box, hidden or mysterious, there's little a customer can do if operations go awry other than rely on the vendor for support. "This can pose a significant threat and risk," he warned.
Making off-the-shelf AI work for business needs also requires at least a certain degree of experimentation. "AI continues to be an emerging technology," EY's Wilczak noted. "Off-the-shelf AI products offer potential ways to accelerate the development of custom solutions, yet as with any other application of AI they do not guarantee getting to an acceptable outcome."