Why Companies That Wait to Adopt AI May Never Catch Up
While some companies—most large banks, Ford and GM, Pfizer, and virtually all tech firms—are aggressively adopting artificial intelligence, many are not. Instead they are waiting for the technology to mature and for expertise in AI to become more widely available. They are planning to be “fast followers”—a strategy that has worked with most information technologies. That likely won’t work. It can take a long time to develop and fully implement AI systems, and there are few if any shortcuts to the necessary steps. Once they have been successfully undertaken, scaling—particularly if the company has a plentiful supply of data and the knowledge engineering mastered —can be very rapid. By the time a late adopter has done all the necessary preparation, earlier adopters will have taken considerable market share—they’ll be able to operate at substantially lower costs with better performance. In short, the winners may take all and late adopters may never catch up.
While some companies — most large banks, Ford and GM, Pfizer, and virtually all tech firms — are aggressively adopting artificial intelligence, many are not. Instead they are waiting for the technology to mature and for expertise in AI to become more widely available. They are planning to be “fast followers” — a strategy that has worked with most information technologies.
We think this is a bad idea. It’s true that some technologies need further development, but some (like traditional machine learning) are quite mature and have been available in some form for decades. Even more recent technologies like deep learning are based on research that took place in the 1980s. New research is being conducted all the time, but the mathematical and statistical foundations of current AI are well established.
System Development Time
Beyond the technical maturity issue, there are several other problems with the idea that companies will be able to adopt quickly once technologies are more capable. First, there is the time required to develop AI systems. Such systems will probably add little value to your business if they are completely generic, so time is required to tailor and configure them to your business and the specific knowledge domain within it. If the AI you are adopting employs machine learning, you will have to round up a substantial amount of training data. If it manipulates language — as in natural language processing applications — it can be even more difficult to get systems up and running. There is a lot of taxonomy and local knowledge that needs to be incorporated into the AI system —similar to the old “knowledge engineering” activity for expert systems. AI of this type is not just a software coding problem; it is a knowledge coding problem. It takes time to discover, disambiguate, and deploy knowledge.
Particularly if your knowledge domain has not already been modeled by your vendor or consultant, it will typically require many months to architect. This is particularly true for complex knowledge domains. For example, Memorial Sloan Kettering Cancer Center has been working with IBM to use Watson to treat certain forms of cancer for over six years, and the system still isn’t ready for broad use despite availability of high-quality talent in cancer care and AI. There are several domains and business problems for which the requisite knowledge engineering is available. However, it still needs to be manipulated to a company’s specific business context.
Integration Time
Even once your systems have been built, there is the issue of integrating AI systems into your organization. Unless you are employing some AI capabilities that are embedded within existing packaged application systems that your company already uses (e.g., Salesforce Einstein features within your CRM system) the fit with your business processes and IT architecture will require significant planning and time for adaptation. The transition from pilots and prototypes to production systems for AI can be difficult and time-consuming.
Even if your organization is skilled at moving pilots and prototypes into production, you will also have to re-engineer the business processes to have full impact on your business and industry. In most cases AI supports individual tasks and not entire business processes, so you will have redesign business processes and new human tasks around it. If you want to affect customer engagement, for example, you will need to develop or adapt multiple AI applications and tasks that relate to different aspects of marketing, sales, and service relationships.
Human Interactions with AI Time
Finally, there are the human challenges of AI to overcome. Very few AI systems are fully autonomous, but are rather focused on augmentation of and by human workers. New AI systems typically mean new roles and skills for the humans who work alongside them, and it will typically require considerable time to retrain workers on the new process and system. For example, investment advice companies providing “robo-advice” to their customers have often attempted to get human advisors to shift their focus to “behavioral finance,” or providing advice and “nudges” to encourage wise decisions and actions in investing. But this sort of skill is quite different from providing advice about what stocks and bonds to buy, and will take some time to inculcate.
Even if the goal for an AI system is to be fully autonomous, it is likely that some period of time in augmentation mode will be necessary. During this period, a critical piece of machine learning occurs through interaction between the system and its human users and observers. Called interaction learning, this is a critical step for organizations to understand how the system interacts with its ecosystem. They can often gather new data sets and begin to bake them into algorithms during this period — which often takes months or years.
Governance Time for AI Applications
While AI systems are geared to provide exponential scale and predictions, they will need a much broader governing approach than the classic controls and testing driven approach. The efficacy of AI algorithms decays over time because these are built based on historical data and recent business knowledge. The algorithms can be updated as the machine learns from patterns in new data, but they will need to be monitored by subject matter experts to ensure the machine is interpreting the change in business context correctly. Algorithms will also have to be continuously monitored for bias. For instance, if an AI system is trained to create product recommendations based on customer demographics and the demographics change dramatically in new data, it may provide biased recommendations.
Governance will also include watching for customer fraud. As the systems become smart so will the users. They may try to game the systems with fraudulent data and activities. Monitoring and preventing this will require sophisticated instrumentation and human monitoring in the context of your business.
Winners Take All
It may, then, take a long time to develop and fully implement AI systems, and there are few if any shortcuts to the necessary steps. Once they have been successfully undertaken, scaling —particularly if the company has a plentiful supply of data and the knowledge engineering mastered— can be very rapid. By the time a late adopter has done all the necessary preparation, earlier adopters will have taken considerable market share — they’ll be able to operate at substantially lower costs with better performance. In short, the winners may take all and late adopters may never catch up. Think, for example, of the learning and capability that a company like Pfizer — which has, according to one of the leaders of the company’s Analytics and AI Lab, more than 150 AI projects underway — has already accumulated. Tech companies like Alphabet have even more learning; that company had 2700 AI projects underway as far back as 2015.
Admittedly, some steps can be accelerated by waiting if a company is willing to compromise its unique knowledge and ways of conducting business. Vendors are developing a vast variety of knowledge graphs and models that use techniques ranging from natural language processing, to computer vision. If one exists for your industry or business problem, and you’re willing to adopt it with little modification, that will speed up the process of AI adoption. But you may lose your distinctive competence or competitive advantage if you do not tweak it to fit your context and build everything around it.
The obvious implication is that if you want to be successful with AI and think there may be a threat from AI-driven competitors or new entrants, you should start learning now about how to adapt it to your business across multiple different applications and AI methods. Some leading companies have created a centralized AI group to do this at scale. Such central groups focus on framing the problems, proving out the business hypothesis, modularizing the AI assets for reusability, creating techniques to manage the data pipeline, and training across businesses. One other possibility may be to acquire a startup that has accumulated substantial AI capabilities, but there will still be the need to adapt those capabilities to your business. In short, you should get started now if you haven’t already, and hope that it’s not too late.
Vikram Mahidhar is the leader of artificial intelligence business at Genpact. He was an entrepreneur and built AI Solutions business at Rage Framework (acquired by Genpact).
Thomas H. Davenport is the President’s Distinguished Professor in Management and Information Technology at Babson College, a research fellow at the MIT Initiative on the Digital Economy, and a senior adviser at Deloitte Analytics. He is the author of over a dozen management books, most recently Only Humans Need Apply: Winners and Losers in the Age of Smart Machines and The AI Advantage.
Why Companies That Wait to Adopt AI May Never Catch Up
Research & References of Why Companies That Wait to Adopt AI May Never Catch Up|A&C Accounting And Tax Services
Source
0 Comments
Trackbacks/Pingbacks