AI is everywhere. We only have to look. From smartphone apps to that self-checkout queue at that neighbourhood store to that web series that got recommended to us, there is data that we as consumers provide and AI-driven personalized experiences that we receive in return.
Despite its ubiquitous presence, there are many more avenues for AI to shine through that are as yet untapped. For AI’s true potential to be fully realized, companies have to adopt AI holistically, infusing it in all lines of businesses, across both front office and back office functions, from the physical to the digital. In other words, companies have to view AI as a critical enabler of digital transformation itself.
Choosing the right use case
Leaders need to start on the AI journey with the right use case. A use case that has manual execution or at best is realized using unsophisticated software, is a significant contributor to or enabler of business, and has potential for an outsized impact (e.g. in revenue/profitability numbers, or as competitive differentiation) is a good candidate to solve for using AI.
Initially, the trendsetter in AI adoption were those scenarios that required cognitive decision making, something that humans were quite good at with little to no training but that had befuddled machines & software. That was until the breakthrough applications of a particular design of AI systems called as deep learning happened a little over a decade ago.
In process-driven sectors such as manufacturing as well as process-driven functions such as the back office of service enterprises such as banking and insurance, AI has led to tangible cost savings and increased speed of processing. This has often had ripple effects going all the way to increase in business by way of reduced turn-around times for workflows. In particular, mundane tasks such as automated paperwork processing tend to be ideally suited for AI to do well. Most companies view such automation as a necessary first step in their move towards launching digital offerings.
Getting the data in order
“Data is the new oil”, said Clive Humby, a mathematician, while working on Tesco’s loyalty card. While this quote was from 2006, the early times of big data, it’s more modern cousin meant for the AI world could very well be “Labelled data is gold dust”.
Labeling is the process of attaching well-formed, informative labels to data in a consistent manner that guides machines in their learning. When done correctly and for large enough data, modern AI algorithms (like the deep learning ones) can often learn well and produce wondrous results that seemingly demonstrate a mastery at times surpassing human ability.
Care must be taken to avoid training with unbalanced labelled data that could prejudice the learning. Common causes of imbalance are unbalanced data itself and biases in labelling, among others.
Trying it out
Actions speak louder than words. Having done the hard work to engender AI-driven alternatives in the lab, many companies hesitate to take the plunge and take action. These companies are often stuck at this stage for fear of failure or, worse, lack of perfection.
By its very definition, an AI system learns with more experience (and corresponding labelled data). It is therefore impractical to wait for the perfect system to emerge before rolling it out on the field. On the contrary, it is best practice to launch trials, first at a limited scale, gather real-world feedback, iterate, then try out at a larger scale, and so on. After a few iterations, the improvement in performance can be assessed to see how many more iterations are required and at what cost in order to make the decision of going live or going back to the drawing board.
In summary, AI is a crucial enabler of digital transformation. Companies can be more successful in their transformation agenda by choosing the right set of use cases that AI is well suited for. For the chosen scenarios, collecting adequate amounts of labelled data in the conduct of the business is necessary. Finally, for AI-infused systems, imperfect action is better than perfect inaction. After all, much like the organization itself, a learning machine does best when it continues to learn.