While it is true that some technologies need further development, others (such as traditional machine learning) are quite mature and have been available in one form or another for decades. Even newer technologies such as deep learning are based on research conducted in the 1980s. New research is being conducted all the time, but the mathematical and statistical foundations of current AI are well established and known. The implementation of a deep learning-based system is tied to a series of steps that as such take time, a factor that must be thought through very well.
- Time for development.
First of all, artificial intelligence systems must be developed. And since they would add very little value to the company if they were completely generic, time is needed to adapt and configure them to one’s business and the specific knowledge within it. If the AI you are adopting employs machine learning, you have to accumulate a considerable amount of data for its training, and if it analyzes images-as in computer vision applications-it can be even more difficult to make the systems work.
There is a lot of taxonomy and local knowledge that has to be incorporated into the AI system, similar to the old “knowledge engineering” activity for expert systems. AI of this kind is not just a software coding problem: it is a knowledge coding problem. It takes time to discover, disambiguate, and disseminate knowledge.
- Time for integration.
Once systems are in place, they need to be cross-integrated from the production line. Adaptation with business processes and IT architecture requires significant planning and adaptation time. The transition from pilot and prototype projects to production systems can be difficult and time-consuming.
Even when you are able to move pilot projects and prototypes to production, it becomes necessary to redesign business processes to fully impact your business and industry. In most cases, artificial intelligence supports individual activities and not entire processes, so these must be redesigned as well as new human activities around it.
- Time to create human-machine interactions.
Finally, there are human challenges to overcome. While in many cases AI-based systems are fully autonomous, in others they are focused on enhancing the capabilities of human workers. The introduction of AI-based applications means new skills or even new roles for the humans working alongside them, and some time is usually required to retrain workers on the new processes.
Although the goal for an artificial intelligence system is to be fully autonomous, some time is normally required for training and refinement. During this period, a critical part of machine learning occurs through interaction between the system and its human users and observers. Called “interactive learning,” this is a critical step in understanding how the system interacts with its ecosystem. During this time, new data sets can be collected to begin turning them into algorithms, a task that often takes months or years, depending on the case.
- Time for governance.
Because artificial intelligence systems are optimized to deliver results on an exponential scale, their governance requires a broader approach than traditional controls or testing approaches. The effectiveness of algorithms can alter over time because they are based on a mix of historical data and recent knowledge. Systems can be updated as the machine learns from patterns in new data, but they must be properly monitored to ensure that they correctly interpret any changes. For example, if a system is trained to check the quality of parts manufactured by a certain process, and that process is heavily modified, one must ensure that the appearance of the same parts manufactured by the new process has not been distorted, at least in the eyes of the vision system.
Governance must also contemplate surveillance of possible cheating by users. As technology becomes smarter, so do their users, who may try to bypass systems with incorrect data and activities. Monitoring and preventing this requires the right precautions and regular human monitoring.
Winner takes all.
Developing and implementing artificial intelligence-based systems could therefore take a long time, and there are few shortcuts to the various steps needed. Once these have been successfully undertaken, scaling can be very rapid, particularly if one has an abundance of data and adequately masters knowledge engineering. By the time a follower has completed all of the necessary preparation, early adopters will have taken on considerable market share: they will be able to operate at significantly lower cost with better performance. In short, the winners may “take it all” and the followers may never catch up.
Of course, some steps could be accelerated by waiting for a company willing to surrender its unique knowledge and ways of conducting business. Some vendors are developing “prepackaged” knowledge models for machine vision applications. If one exists for your specific industry or problem, and you are willing to adopt it with little modification, you may be able to accelerate the AI adoption process. But you might also lose your distinctive expertise or competitive advantage if the system is not modified to fit your context.
The obvious conclusion is that if you want to succeed with AI and you think there might be a threat from competitors or AI-driven newcomers, you should start adopting it in your company now.