Five Mistakes Companies Make With Machine Learning

eWEEK DATA POINTS RESOURCE PAGE: The development of ML technologies has given rise to the ability to extract more information and intelligence​ from the wide range of content in the enterprise, whether structured or unstructured.


Two forces are driving a surge in the use of machine learning technology and other artificial intelligence-enabling technologies, according to industry analysts: the astounding growth of unstructured content and the use of robotic process automation (RPA) to automate content-related processes.

Cognilytica says that between documents, images, emails, online data and videos, up to 90% of the content in the enterprise is in the form of unstructured data, which is growing at an astounding 55% to 65% per year.

Go here to see a listing of eWEEK's Top Predictive Analytics Companies.

Consequently, Everest Group Research says intelligent automation technologies are using ML where RPA intersects and interoperates with content-related processes. The development of ML technologies has given rise to the ability to extract more information and intelligence​ from the wide range of content in the enterprise, whether structured or unstructured. 

Machine learning refers to software that enables machines to “learn” in both a supervised and unsupervised way, improving accuracy and performance. In a process involving capturing documents and processing with RPA, machine learning and other AI technology learn from potentially thousands of variations of documents, such as processing invoices or handling vendor orders.

Nonetheless, according to ABBYY Chief Innovation Officer Anthony Macciola, organizations are making five common mistakes when working with ML solutions. ABBYY is a global provider of Digital IQ technologies and solutions.

Data Point No. 1: Overly Complex ML Capabilities

Organizations risk working with ML tools that require a large amount of data in order to train for the most basic unstructured content use cases. Use proven ML tools that contain advanced algorithms that can be trained using a small data set and can run in full production in just a few hours, rather than needing hundreds of thousands of documents in a sample set to get a project up in running, which can take weeks or even months to occur.

Data Point No. 2: Relying Too Much on RPA

RPA is acclaimed for boosting efficiencies by connecting to legacy systems and external data sources. It can be rapidly deployed and its digital workers are easy to configure – and once in place, they perform work just like humans. The big difference between RPA and ML technologies is RPA is focused on repetitive structured work while ML is designed to understand structured and unstructured content. RPA needs ML technology to give its digital workers content intelligence, thereby giving them the cognitive skills to extract useful information and gain intelligence, learn from various forms of content, derive meaning and intent of documents, and add decision-making capabilities.

Data Point No. 3: Assuming They Know Where to Apply ML Technology

When starting an automation project, the right processes to start with are not always selected. That’s because many companies are compartmentalized in organizational process knowledge. Additionally, top management is not involved in the day-to-day workflow and lack process documentation, making it increasingly difficult to truly discover what processes are ready for automation. Incorporating process intelligence before a project will give you a comprehensive understanding of where to apply RPA and ML solutions, and their expected value and savings to the organization – all based on data not opinion or bias.

Data Point No. 4: Missing Out on High-Value Business Cases

Typically, a company will rely on conventional wisdom and select a task that occurs the most frequently because it has the appearance of delivering great results. However, this ad hoc approach to process selection may ignore other business opportunities that lead to better ROI opportunities. While it’s completely acceptable to start in areas that have the least amount of disruption to the organization or interaction with end users, you should have in mind how you can “land and expand” ML quickly and easily throughout the company.

Data Point No. 5: Thinking It’s One and Done

The job isn’t over just because you’ve trained your algorithms and deployed your digital workers. It is absolutely critical to enable continuous improvement by monitoring and measuring automation’s up and down stream impact to ensure ongoing protocol compliance and prevent the bottleneck shift and the possibility of negatively impacting the process in other places. Monitoring the digital workforce as well as the entire end-to-end process post-implementation is just as essential as the planning and execution.

If you have a suggestion for an eWEEK Data Points article, email [email protected].