Machine learning has the potential to automate many more business processes than are currently automated in enterprise software, based on all the previous generations of software development methods.   

That is a suggestion put forward by Claus Jepsen, chief technology officer at Unit4, an ERP software supplier based in Denmark.    

The degree to which ML can improve the business outcomes is “currently marginal,” he suggests with accuracy of financial forecasts, for instance, sensitive to many greater factors than how well the algorithm can refine itself over time. “If you haven’t got harmonized, accurate and complete data to start with, simply applying ML to it isn’t in itself going to result in better business decisions,” Jepsen stated. 

Defining the business problem is the same challenge that has always faced software developers. “In terms of Gartner’s hype cycle, ML is currently at the peak of inflated expectations,” he stated. “You cannot simply throw ML at a bucket of big data and expect it to magically come up with a perfect business plan.”  

The points in a business process where some judgment or prediction is required, and where a small improvement in accuracy would have a strong benefit to the business, are candidates for ML automation. The humans surrounding the effort to get AI to work are critical. They need to decide the use case and make sure the data is of high enough quality to be useful, before giving the algorithm a task, and then training it.   

“The human mind is by far the best pattern-matching machine in the universe,” Jepsen stated. “The average two-year-old can probably correctly identify a cat after it’s seen two or three, while an ML algorithm might need to see 2,000 before it can be sure. But, once trained, ML excels at dealing with huge volumes of data and processing it very quickly, while never getting bored performing repetitive, tedious tasks day in, day out.” 

Machine Learning Catching on in Africa  

This insight of machine learning extending automation beyond what software development has so far achieved, extends to Africa, where machine learning is making gains. IDC analysts have projected that spending on AI in the Middle East and Africa is expected to maintain its strong growth trajectory as businesses continue to invest in projects that use AI software and platforms, according to an account in Intelligent CIO Africa.   

AN IDC survey of IT leaders found that ML improved customer and employee experience and led to accelerated rates of innovation in the organization.  

ML Platform Buy or Build Considerations  

Companies that commit to pursuing machine learning for AI software development face a decision on whether to buy or build the needed ML platform. 

“Building a solution takes years and headcount,” states Charna Parkey, data science lead at Kaskada of Seattle, in a recent account in builtin. Kaskada is building a machine learning platform aimed at enabling collaboration on feature engineering and repeatable success in production.  

Airbnb for example took three months to decide what to build in their ML platform and four years to build it; they call it Bighead. Its developers used a range of open source technologies, working to “fix the gaps in the path to production” with their own services and user interface. This meant they had to support multiple frameworks, feature management and model and data transformation. In a similar experience, Uber has been working for five years on its platform, called Michaelangelo.da. 

Finding the needed talent is always a challenge. The basic decision is whether to hire a classically trained data scientist, or hire a domain expert and upskill. “I chose to upskill,” Kaskada stated, and she is not alone. Some 46% of organizations surveyed by PwC in 2020 reported they were rolling out AI upskilling to handle the shift to more AI, and 38% were implementing credentialing programs.   

Buying a pre-built ML platform saves the initial costs to build, the integration costs for “custom, brittle workflows,” and it comes with dedicated external support, she stated. It also reduces the time it takes to onboard new employees to proprietary software. The costs of moving to a pre-built platform including having to adopt new workflows instead of building to those the company has in place, and perhaps telling developers their favorite tools are no longer in vogue. 

“Not all platforms will support the entirety of your ML operations or your company’s unique needs,” Kaskada suggested. “Evaluate carefully.” 

New Book: Real World AI: A Practical Guide for Responsible Machine Learning  

In the real world of applied ML applications, the challenges are just beginning to be understood, suggest the authors of a new book, Real World AI: A Practical Guide for Responsible Machine Learning, by Alyssa Simpson Rochwerger and Wilson Pang, two experienced practitioners of applied machine learning. Rochwerger is a former director of product at IBM Watson, and Pang is the CTO of Appen, a company focused on improving the quality of data for ML applications, based in Chatswood, Australia. 

“Only 20% of AI in pilot stages at major companies make it to production, and many fail to serve their customers as well as they could,” Rochwerger and Pang write in Real World AI, according to an account of the book recently published in TechTalks. “In some cases, it’s because they’re trying to solve the wrong problem. In others, it’s because they fail to account for all the variables—or latent biases—that are crucial to a model’s success or failure.”  

The real world clashes with the academic roots of AI when it comes to data.  

“When creating AI in the real world, the data used to train the model is far more important than the model itself,” Rochwerger and Pang write in Real World AI. “This is a reversal of the typical paradigm represented by academia, where data science PhDs spend most of their focus and effort on creating new models. But the data used to train models in academia are only meant to prove the functionality of the model, not solve real problems. Out in the real world, high-quality and accurate data that can be used to train a working model is incredibly tricky to collect.”