MLOps with impact
Drive business outcomes with enterprise adoption of AI
Fundamentally world-changing AI technology
As artificial intelligence (AI) gains importance, it’s creating amazing results across many industries. From retail sales forecasting to supply chain issue resolution to potential disease prediction to customer service automation, there are endless opportunities.
Every department in every company wants some aspect of AI to drive business value. The technology is fundamentally world-changing. Its invention can be equated to that of the lightbulb.
Moving AI/ML models from development to production
Enabling the machine learning (ML) models that drive AI is no easy feat. And deploying AI solutions in production is challenging. If business stakeholders and technologists struggle to collaborate effectively, resulting investments in AI can fail to address the business need.
Too often, the focus of data science teams can lie more on designing and deploying highly accurate AI/ML models than working with business and product teams to ensure end-to-end orchestration with business workflow solutions. Machine learning operations (MLOps) can become costly. By redefining the framing of MLOps, organizations can better meet the needs of the business and drive value.
Value arrives in a calculated, ongoing process
MLOps can tie models to business value. However, AI is not a typical technology deployment. ML models need to be observed with feedback loops to ensure optimal capabilities. It’s not a “once and done” scenario. It’s a calculated, ongoing process—and a mindset—that gives data science teams a structured way to rapidly develop, deploy, monitor, and maintain AI/ML solutions that make a real impact on the business. It is not a single tool or technology.
MLOps is an end-to-end AI/ML life cycle management approach necessary for governance and agility. And, it needs to have guardrails in place.
The four key phases of MLOps
And how one company used it to power its sales efforts
Here’s a taste of how one life sciences company successfully implemented MLOps. For more details on this machine learning case study and to learn how a fast-food restaurant company used the same approach for a totally different use case, download the report.
Phase 1: Collaborating with business
The Envision phase helps the entire team understand what they are building and why.1 The vision for this life sciences company was to drive more targeted, personalized engagement with its customers. The sales team needed better intelligence to drive their efforts and in less time. Additionally, they wanted to ensure that the models were built thoughtfully, fairly, and accurately.
Envision: define an AI solution hypothesis that includes business value to be achieved and risks to be addressed.
1. Deloitte AI Institute, ”Take a new view on MLOps,” 2022. https://www2.deloitte.com/us/en/pages/consulting/articles/mlops-for-business.html
Phase 2: Models are deployed
Each model that was built ran through this iterative process:
• Review requirements
• Collect data
• Build models
• Set controls
• Show insights
Build: collect and process data; build, validate, and test models; and utilize controls to manage versions, audits, and reusable components.
Phase 3: Game on! Change management is essential
The effort should be measured. This company took the deployment seriously by looking at sales professionals’ adoption rates by brand and region. A/B testing was used to measure receptiveness and adoption by the sales teams—and it was done by teams in local markets to reflect their unique needs. Finally, the models were fine-tuned and rolled out nationally and then globally.
Deploy: adopt the model into the product or business processes and establish mechanisms to monitor the model's performance.
Phase 4: Watch, learn, listen
Once deployed, the models were measured for criteria such as drift and shop values. A key component was creating a feedback loop with the model developers and the sales professionals, so development teams knew what was meaningful to them or what additional training they needed to do on the models.
Monitor: create feedback loops and oversee a continuous learning environment.
