Andrew Ng, Computer Science professor at Stanford and one of the leading lights in the field of Artificial Intelligence (AI) has been quoted as saying “Artificial Intelligence is the new electricity”. The implication is that this technology will leave almost no industry unchanged and the transformative effects of AI could be felt for decades to come. Add to this the constant barrage of AI related news from self-driving cars to voice assistants, and dire predictions of potential evil effects from job losses to AI-powered weapons in wrong hands. This has created a sense of urgency and panic even in industrial enterprises that view hardware as their key product, and data, if any, as incidental to their business.
Thus, across the spectrum, companies are now trying to figure out how best they can use the new AI technologies via “digitalization initiatives”. The data to be crunched for AI can come from a variety of sources such as sensors, SCADA (Supervisory Control and Data Acquisition) systems, enterprise management software etc. This is the meeting ground – and in some cases battleground – of traditional engineering and the ‘new’ disciplines. On one hand are tools, techniques, and people trained in the known ways of doing things, using relatively well-understood technologies. On the other hand is the brave new world of “digital technologies” with ever changing acronyms, terminologies, and constant attention in the popular press. This poses some fairly non-obvious challenges for implementation of AI initiatives in an industrial enterprise. The following is an indicative list of challenges culled from the first-hand experience of many colleagues in different industries, apart from my own.
At the outset, there is a challenge of justifying ROI on AI initiatives. Three key things promised by AI, namely efficiency improvements, gaining new insights into the data, and enabling new business models, are not easy to deliver in the time frames demanded by the management. It takes time and sustained efforts, not to mention upfront investments, to realize the benefits.
The second challenge for companies that are not “digital natives” concerns the data. It is usually not digitized and generally not in a format that can be fed directly into the AI models. Sometimes the data is in paper records, with hard to decode annotations/abbreviations, since they were meant to be read by trained humans. Even if it is machine data, the data needs to be ‘curated’. This can entail huge upfront efforts.
Next, the way the AI models are developed and implemented do not make them a natural fit in the existing workflow process of the internal clients. Packaging the AI models in such a way as to fit seamlessly into the existing workflow can sometimes take more time and effort than the development of AI solutions! Modification of the existing processes can also be expensive and time-consuming due to regulatory, safety, or operational requirements.
The fourth challenge revolves around the question of mismatched expertise pools within the organization. The existing workforce, though well versed in the domain, is typically not AI savvy. The new hires, generally young graduates who have acquired proficiency in AI/ML and related areas, lack the domain knowledge of the industry. Not surprisingly, both groups tend to view each other with a bit of suspicion and sometimes hostility. This can lead to missed opportunities and delays, not to mention heartburn.
Another recurring theme concerns the fact that AI tends to give superior results when the complexity and ‘dimensions’ of the problem are very large. But training AI models for a complex problem also requires large amounts of data that is properly curated. Data collection, labeling, and validation take time and efforts. (Someone has said ‘data scientists spend 80% of their time in preprocessing the data, and the rest 20% complaining about it’!). Also, validating the results in an industrial set-up (on say on a manufacturing line) can be both costly and time-consuming.
Above, a few challenges for injecting AI into ‘brick-and-mortar’ companies are outlined. However, by experimentation and some deft handling of organizational viscosity, some of the challenges can be addressed. First is to have AI teams work closely with the people who have the knowledge of the context, own the data, and actually are looking to have some specific problem solved. Secondly, AI teams also need to recognize that the legacy data will never be ‘clean’ and hence should try to develop suitable solutions to address this. AI models that detect and correct anomalies in the input data could be one solution. Next, the analysts also need to fully understand the workflow of the internal clients and define their work scope to include the injection of AI models into the workflow as the part of their mission. And finally, sponsorship of the top management in the initial period till success can be demonstrated is absolutely needed to make a lasting impact.
-Ardhendu G Pathak, Ph.D.
At Koinvent, we help leaders think deeper, work smarter and manage better. If you would like to learn more about Koinvent, please visit our website at www.koinvent.com or send us an e-mail at firstname.lastname@example.org.