Skip to main content
Industry

The art of implementing machine learning at scale

Machine learning holds the promise of saving time and money while increasing accuracy and efficiency across a variety of business use cases – but there can be pitfalls on the road to implementation. Companies often find artificial intelligence projects surprisingly easy to launch but then deceptively difficult to scale, from both the tech and organization sides. We at Tata Consultancy Services (TCS), a global IT services, consulting, and business solutions organization with almost half a million consultants, have built our own roadmap for navigating these challenges. Whether it’s increasing predictive strength for powerful insights into your manufacturing process or exponentially improving efficiency in document reviews, we hope our experience and learnings will help you reap the benefits of this emerging technology.

What’s different about machine learning?

Machine learning requires and generates large quantities of data. Without a solid AI strategy, having much more data without knowing how to use it effectively is not necessarily a good thing – especially for smaller companies. Software implementations are often viewed as a package of work where the roadmap and the target outcome are clearly determined at the start. With machine learning, that’s simply not the case: The exact end-state can be unclear. You know the final impact you want, (for example, cost reduction), but you may not have a clear picture of the output needed to get there. This requires ideation, debate, and fine-tuning throughout the process. And that means that failure along the way is likely! It is critical to be able to evaluate outcomes mid-process and course correct as necessary.

A start-up state of mind

We have embraced a hybrid model that incorporates the best of DevOps thinking based on a framework of lean start-up principles. In this model, there is always room for debate and iteration in an environment that is fast, flexible, and explainable. Learning must happen continually for constant improvement. Even after delivery, you can monitor the development and view the learning as an ongoing process. We have developed a process life cycle for these projects, with nuances in the various phases and activities. Modularizing development is critical for advancing what works while being willing to drop what doesn’t.

What keeps the model working is a simultaneous focus on people, process, and technology. Various roles must work together, and each part of the project requires its own diverse, non-overlapping skillset. This includes data scientists, engineers, and web developers as well as people on the business side, all interacting to make this work who must be willing to have challenging conversations.

For example, not every forecasting model is simple. The opportunity statement may be one line, but sometimes it’s challenging to put together scattered data or data with doubtful integrity. A lot depends on the conversation with the business partner. You start with a high-level statement, but it’s key to understanding what they really want, and to keep re-evaluating the results to see if you’re getting there. Within this structure, it is important to have a target architecture along with the right tools, templates, and automation for accelerated builds and deployments.

Real-world benefits

We have had success with several types of machine learning implementations. Machine learning can save a tremendous amount of time in document processing and allow skilled analysts to apply their talents where they are most needed instead of wasting time on routine or unnecessary tasks. One cognitive visual learning (CVL) project reduced the number of contracts reviewed annually from 3,200 to 1,800, saving approximately 10,000 hours over the course of the year and allowing manual reviews to focus on the most important cases. In a forecasting project, we reached a mean absolute percentage error of less than 2% for engineering group headcount forecasts, compared to manual forecasting. Machine learning helped us forecast over 1,000 combinations in less than three hours.  Overall, having a front-end UI that’s easy to understand helps democratize the power of machine to every Finance user, not just those who are more tech savvy.

New mindset, new results

The key aspect of our work on developing forecasting models for Microsoft finance was that we could not use the standard business intelligence delivery methodology to run the machine learning models and cognitive projects. Machine learning implementation requires a growth mindset to simulate, learn, react, cycle, and continuously improve the intelligent models. Once we found the holistic strategy that worked, we have been able to continually improve our machine learning offerings, with each tool we build enabling us to take the next one forward. Yes, the challenges surrounding machine learning implementations are real, but the promise is more powerful and the tools more accessible with every iteration.  What will you start building?