Applied AI/ML is a top priority for organizations, according to Tecton‘s first ‘State of Applied Machine Learning’ survey. The survey identified the challenges and opportunities in the applied ML space and pinpointed common trends across a diverse set of ML initiatives.
The survey found that companies in many industries are increasingly adopting applied ML for a wide range of use cases, including customer analytics, personalized recommendations and fraud detection. At the same time, many are also facing multiple challenges on their journey to implementing applied ML, such as generating accurate training data, building production data pipelines and demonstrating business ROI.
Because of the difficulty inherent in data, ML is still difficult to get right for most organizations. However, despite the challenges, survey respondents indicate that their companies are committed to improving their applied ML capabilities, with a growing focus on improving model deployment time, adopting real-time ML and implementing central ML platforms to improve cross-team collaboration and organizational scalability.
“The survey findings confirmed the trends we found working with customers and the ML community,” said Mike Del Balso, co-founder and CEO of Tecton. “Namely, that an increasing number of companies are moving to real-time ML and investing in their MLOps stack to overcome challenges so teams can deploy models to production faster.”
Key Survey Findings:
1) Applied AI/ML Is a Top Priority for Organizations
- Applied ML is declared as the number 1 company initiative for 23.8% of respondents and a top 3 initiative for 60.1%
- Roughly 50% of survey participants indicated their organizations have 6 or more models in production, and companies are planning on deploying more models quickly. Based on survey responses, the projected median for the number of models in production in the next 12 months will increase from 6 to 10
- Companies are leveraging applied ML for use cases that directly impact revenue: The top 3 use cases are recommender systems (45.1% of respondents), customer analytics (43.3%) and personalization (36.4%)
2) Companies Are Investing in Their MLOps Stack to Address Challenges
- The top 3 challenges encountered when deploying new models to production are generating accurate training data (41.1%), building production data pipelines (37.6%) and demonstrating business ROI (34.3%)
- Deploying a new model to production is a long process (more than 1 month for 65.0% of respondents and more than 3 months for 31.7%). However, 54.5% of respondents shared that their organizations aim to reduce deployment time by at least 25% over the next 12 months
- Only 9.5% of respondents say their organization has a full MLOps stack today consisting of 5 components: model serving; model registry and versioning; feature store / feature platform; model monitoring and observability; and data monitoring. However, 59.1% of respondents indicate that their companies are planning to have all 5 components within 12 months
- The Feature Store / Feature Platform and Monitoring & Observability components will see the largest increases (~43 percentage points increase for both) in adoption in the next 12 months
3) Real-Time ML is Picking Up Real Traction
- 68.3% of ML respondents surveyed indicated their teams already have at least one real-time (e.g., sub-100 milliseconds) ML model in production, while 14.7% have more than 10
- 34.1% of respondents say their ML teams use exclusively batch data to power their real-time ML models today, but this number is expected to drop to 11.5% within 12 months as they plan to adopt streaming or real-time data