If one were to listen to the current debates on AI, we are either promised a world of limitless possibilities or the threat of destruction. This hyperbole is amplified by the media, while all too often the real opportunities to drive business decisions and automate processes are overlooked.
As with any emerging technology, the future of AI is still largely unwritten. We are nevertheless entering the world of Business 4.0, and opportunities are plentiful for the wise. In an age of data abundance, deep learning promises better and more robust performance. But there are crucial steps that must be taken before we can get the maximum benefits from the data-driven world we are now entering.
In the past few years, deep learning techniques and technologies have almost taken over large parts of machine learning. Apart from improved predictive powers, deep learning allows for dealing with a variety of data, for example, flat, temporal, textual, or image-based, via a common approach.
In this same timeframe, software products in the data science/statistics space have rapidly become outdated and have been replaced by a proliferation of open-source tools for deep learning as well as easy data visualization.
Meanwhile, there is increasing insistence from business to drive AI adoption, while overlooking the primary importance of data science. Adopting such technology without the necessary organizational maturity to formulate, implement, go live, and continuously improve use cases based on machine learning is like looking at the problem from the wrong end of the telescope.
Focusing on imagining the right business problems and opportunities while being aware of both the possibilities as well as limitations of data science is a better strategy. The bigger challenge, I would say, is balancing business expectations while measuring and delivering real value.
And while we talk about data, it’s important to look past the initial lure of public source data and mine valuable internal data, something that is far harder to access in spite of considerable corporate spend on data lake platforms and so on.
Looking through the right end of the telescope means starting from the business (or social) problem that one wants to impact and only then focusing on the technology. One should also ask further questions: is the goal process automation, or is it to amplify human effectiveness with contextual insight, or is it digitalization of activities requiring computational as well as predictive power, like with real-time offers on e-platforms?
The business-first approach requires assembling a team that is capable of close collaboration, leaps of imagination, readiness to experiment, and demonstrates not just potential but actual value. Really good data scientists are a rare commodity, but finding one that can lead a team, ideally one that includes a statistician, some data visualization and machine-learning/deep-learning engineers, and one or more business domain experts, is optimal.
What is critical is for the leader of such a team to have the ability to amalgamate these skill sets to both, form hypotheses and solve business problems, as well as communicate the new data-driven culture across the whole unit or organization.
Equally important to consider is the type of IT framework required. What are the parameters of the business problem? Is it automation of tasks, amplification of skills, or digitalization? Answering this helps define whether a data-driven or AI solution is needed.
For example, automating or eliminating a business process using a machine learning model can be implemented within an IT workflow system where certain steps are automatically executed.
Robotic process automation tools allow such automation using hand-crafted rules; plugging a machine learning-based decision into such scripting engines can be easily done.
Amplification of employee effectiveness and knowledge synthesis might require, for example, an internal social network such as Slack, into which an intelligent chatbot is deployed.
The Human Touch
Finally, digitalization of a process like creating personalized offers or experiences for online consumer traffic would require a different framework altogether; for example, a streaming engine and in-memory database, or even a parallel execution platform. For all such ventures, developing a culture of ‘deploy first, then measure effectiveness, and continually improve’ is best.
As with any venture into unknown waters, there will be false positives, and they may carry a cost, but they will go down and the system will improve. Machine learning systems get better and better over time – this has to be experienced to be believed.
And one final point: we must not forget to celebrate the role that human intuition and creativity play in imagining and validating business problems – something that machines will take a long time to learn.