Polling almost 750 business leaders across all kinds of companies, the 2020 State of Enterprise Machine Learning report arrives at seven key findings and a general conclusion that predicts significant growth. “We will likely see a boom in the number of ML (machine learning) companies providing (deployment and maintenance) services.”
Although the terms are sometimes used interchangeably, machine learning is a tool or application in the larger AI system. A good definition from the Expert System company explains the relationship and essential, unique talents of ML: “Machine learning is an application of artificial intelligence that provides systems the ability to automatically learn and improve from experience without being explicitly programmed.” A measure of the independence of ML applications is seen in the two general types of “learning”—there’s both supervised and unsupervised learning.
SEVEN TRENDS IN ENTERPRISE ML
The seven key findings on the state of ML models in business are summarized here.
- The number of company data scientists is growing rapidly across all industries. In 2018, 18% of the companies surveyed reported 11 or more data scientists on their teams. In 2019, that number jumped to 39%. The task of hiring data scientists is becoming more difficult for small and midsized companies due to the high demand, a talent shortage, and competition from larger companies like Amazon, Apple, Facebook, and Google.
- Business use cases for machine learning are becoming more varied. In the survey, the top three use cases for all the companies, no matter the size, were: (1) reducing company costs; (2) generating customer insights and intelligence; and (3) improving customer experience. In general, Algorithmia found that large companies use ML primarily for internal applications to reduce costs, and smaller companies focused on customer-centric issues.
- The survey tried to map where the responding companies were on the machine learning road map, and the conclusion wasn’t surprising. In the 2018 survey, almost 40% said they were just beginning to develop ML. In 2019, 22% described themselves as early-stage adopters, 15% said they were at midstage level (with models in production for two to four years), and 8% were sophisticated (with models in production for more than five years). As a general marker, 55% of the companies haven’t yet deployed a machine-learning model.
- The fourth key finding identified an unreasonably long road to deployment. Almost half the companies report they spend between eight and 90 days to deploy one ML model. A significant 18% are taking longer than 90 days, with some taking more than a year.
- The problem is scaling ML projects up to full potential. In 2018, 30% of the companies reported difficulties with scaling up models. In 2019, those struggling went up to 43%. When you look at the numbers for companies of more than 10,000 employees, 58% say scaling up was their top challenge.
- Budgets for most ML programs are growing by 25%, with the largest growth seen in banking, manufacturing, and IT industries. Overall, the numbers for this year’s survey reflect growth at companies of all sizes. Algorithmia points to MarketWatch figures that compare compound annual growth rate of $23.94 billion in 2018 ranging upward to an impressive $208.49 billion by 2025. They caution that the general upward budgetary trend shows a doubling down on tech investing of those at midlevel (with models built and deployed for two to five years), which will require those at very early deployment stages to triple their efforts to be competitive. For those in the banking and financial services group, the recent survey shows that budget increases of 40% increased 1%-25%, 27% increased 26%-50%, 33% showed no change, and none of those surveyed decreased spending.
- How do you measure machine learning success? Finally, the survey asked this question, and the two metrics that tied for first place were: “business metrics, such as guaranteed ROI, and a more technical evaluation of ML model performance.”
The survey and the 29 pages detailing the results are worth a look. You can download the full report in PDF format from Algorithmia (bit.ly/3akD0VI).