ULRIKA JÄGARE is the MSc. Director of Technology and Architecture at Ericsson AB. She has over 10 years of experience in data, analytics, and machine learning/artificial intelligence and over 20 years’ experience in telecommunications.

• Foreword xiiIntroduction xvChapter 1 Balancing the AI Investment 1Defining AI and Related Concepts 3Operational Readiness and Why It Matters 8Applying an Operational Mind- set from the Start 12The Operational Challenge 15Strategy, People, and Technology Considerations 19Strategic Success Factors in Operating AI 20People and Mind- sets 23The Technology Perspective 28Chapter 2 Data Engineering Focused on AI 31Know Your Data 32Know the Data Structure 32Know the Data Records 34Know the Business Data Oddities 35Know the Data Origin 36Know the Data Collection Scope 37The Data Pipeline 38Types of Data Pipeline Solutions 41Data Quality in Data Pipelines 44The Data Quality Approach in AI/ML 45Scaling Data for AI 49Key Capabilities for Scaling Data 51Introducing a Data Mesh 53When You Have No Data 55The Role of a Data Fabric 56Why a Data Fabric Matters in AI/ML 58Key Competences and Skillsets in Data Engineering 60Chapter 3 Embracing MLOps 71MLOps as a Concept 72From ML Models to ML Pipelines 76The ML Pipeline 78Adopt a Continuous Learning Approach 84The Maturity of Your AI/ML Capability 86Level 0— Model Focus and No MLOps 88Level 1— Pipelines Rather than Models 89Level 2— Leveraging Continuous Learning 90The Model Training Environment 91Enabling ML Experimentation 92Using a Simulator for Model Training 94Environmental Impact of Training AI Models 96Considering the AI/ML Functional Technology Stack 97Key Competences and Toolsets in MLOps 103Clarifying Similarities and Differences 106MLOps Toolsets 107Chapter 4 Deployment with AI Operations in Mind 115Model Serving in Practice 117Feature Stores 118Deploying, Serving, and Inferencing Models at Scale 121The ML Inference Pipeline 123Model Serving Architecture Components 125Considerations Regarding Toolsets for Model Serving 129The Industrialization of AI 129The Importance of a Cultural Shift 139Chapter 5 Operating AI Is Different from Operating Software 143Model Monitoring 144Ensuring Efficient ML Model Monitoring 145Model Scoring in Production 146Retraining in Production Using Continuous Training 151Data Aspects Related to Model Retraining 155Understanding Different Retraining Techniques 156Deployment after Retraining 159Disadvantages of Retraining Models Frequently 159Diagnosing and Managing Model Performance Issues in Operations 161Issues with Data Processing 162Issues with Data Schema Change 163Data Loss at the Source 165Models Are Broken Upstream 166Monitoring Data Quality and Integrity 167Monitoring the Model Calls 167Monitoring the Data Schema 168Detecting Any Missing Data 168Validating the Feature Values 169Monitor the Feature Processing 170Model Monitoring for Stakeholders 171Ensuring Stakeholder Collaboration for Model Success 173Toolsets for Model Monitoring in Production 175Chapter 6 AI Is All About Trust 181Anonymizing Data 182Data Anonymization Techniques 185Pros and Cons of Data Anonymization 187Explainable AI 189Complex AI Models Are Harder to Understand 190What Is Interpretability? 191The Need for Interpretability in Different Phases 192Reducing Bias in Practice 194Rights to the Data and AI Models 199Data Ownership 200Who Owns What in a Trained AI Model? 202Balancing the IP Approach for AI Models 205The Role of AI Model Training 206Addressing IP Ownership in AI Results 207Legal Aspects of AI Techniques 208Operational Governance of Data and AI 210Chapter 7 Achieving Business Value from AI 215The Challenge of Leveraging Value from AI 216Productivity 216Reliability 217Risk 218People 219Top Management and AI Business Realization 219Measuring AI Business Value 223Measuring AI Value in Nonrevenue Terms 227Operating Different AI Business Models 229Operating Artificial Intelligence as a Service 230Operating Embedded AI Solutions 236Operating a Hybrid AI Business Model 239Index 241

ULRIKA JÄGARE is the MSc. Director of Technology and Architecture at Ericsson AB. She has over 10 years of experience in data, analytics, and machine learning/artificial intelligence and over 20 years’ experience in telecommunications.

• Foreword xiiIntroduction xvChapter 1 Balancing the AI Investment 1Defining AI and Related Concepts 3Operational Readiness and Why It Matters 8Applying an Operational Mind- set from the Start 12The Operational Challenge 15Strategy, People, and Technology Considerations 19Strategic Success Factors in Operating AI 20People and Mind- sets 23The Technology Perspective 28Chapter 2 Data Engineering Focused on AI 31Know Your Data 32Know the Data Structure 32Know the Data Records 34Know the Business Data Oddities 35Know the Data Origin 36Know the Data Collection Scope 37The Data Pipeline 38Types of Data Pipeline Solutions 41Data Quality in Data Pipelines 44The Data Quality Approach in AI/ML 45Scaling Data for AI 49Key Capabilities for Scaling Data 51Introducing a Data Mesh 53When You Have No Data 55The Role of a Data Fabric 56Why a Data Fabric Matters in AI/ML 58Key Competences and Skillsets in Data Engineering 60Chapter 3 Embracing MLOps 71MLOps as a Concept 72From ML Models to ML Pipelines 76The ML Pipeline 78Adopt a Continuous Learning Approach 84The Maturity of Your AI/ML Capability 86Level 0— Model Focus and No MLOps 88Level 1— Pipelines Rather than Models 89Level 2— Leveraging Continuous Learning 90The Model Training Environment 91Enabling ML Experimentation 92Using a Simulator for Model Training 94Environmental Impact of Training AI Models 96Considering the AI/ML Functional Technology Stack 97Key Competences and Toolsets in MLOps 103Clarifying Similarities and Differences 106MLOps Toolsets 107Chapter 4 Deployment with AI Operations in Mind 115Model Serving in Practice 117Feature Stores 118Deploying, Serving, and Inferencing Models at Scale 121The ML Inference Pipeline 123Model Serving Architecture Components 125Considerations Regarding Toolsets for Model Serving 129The Industrialization of AI 129The Importance of a Cultural Shift 139Chapter 5 Operating AI Is Different from Operating Software 143Model Monitoring 144Ensuring Efficient ML Model Monitoring 145Model Scoring in Production 146Retraining in Production Using Continuous Training 151Data Aspects Related to Model Retraining 155Understanding Different Retraining Techniques 156Deployment after Retraining 159Disadvantages of Retraining Models Frequently 159Diagnosing and Managing Model Performance Issues in Operations 161Issues with Data Processing 162Issues with Data Schema Change 163Data Loss at the Source 165Models Are Broken Upstream 166Monitoring Data Quality and Integrity 167Monitoring the Model Calls 167Monitoring the Data Schema 168Detecting Any Missing Data 168Validating the Feature Values 169Monitor the Feature Processing 170Model Monitoring for Stakeholders 171Ensuring Stakeholder Collaboration for Model Success 173Toolsets for Model Monitoring in Production 175Chapter 6 AI Is All About Trust 181Anonymizing Data 182Data Anonymization Techniques 185Pros and Cons of Data Anonymization 187Explainable AI 189Complex AI Models Are Harder to Understand 190What Is Interpretability? 191The Need for Interpretability in Different Phases 192Reducing Bias in Practice 194Rights to the Data and AI Models 199Data Ownership 200Who Owns What in a Trained AI Model? 202Balancing the IP Approach for AI Models 205The Role of AI Model Training 206Addressing IP Ownership in AI Results 207Legal Aspects of AI Techniques 208Operational Governance of Data and AI 210Chapter 7 Achieving Business Value from AI 215The Challenge of Leveraging Value from AI 216Productivity 216Reliability 217Risk 218People 219Top Management and AI Business Realization 219Measuring AI Business Value 223Measuring AI Value in Nonrevenue Terms 227Operating Different AI Business Models 229Operating Artificial Intelligence as a Service 230Operating Embedded AI Solutions 236Operating a Hybrid AI Business Model 239Index 241

ULRIKA JÄGARE is the MSc. Director of Technology and Architecture at Ericsson AB. She has over 10 years of experience in data, analytics, and machine learning/artificial intelligence and over 20 years’ experience in telecommunications.

• Foreword xiiIntroduction xvChapter 1 Balancing the AI Investment 1Defining AI and Related Concepts 3Operational Readiness and Why It Matters 8Applying an Operational Mind- set from the Start 12The Operational Challenge 15Strategy, People, and Technology Considerations 19Strategic Success Factors in Operating AI 20People and Mind- sets 23The Technology Perspective 28Chapter 2 Data Engineering Focused on AI 31Know Your Data 32Know the Data Structure 32Know the Data Records 34Know the Business Data Oddities 35Know the Data Origin 36Know the Data Collection Scope 37The Data Pipeline 38Types of Data Pipeline Solutions 41Data Quality in Data Pipelines 44The Data Quality Approach in AI/ML 45Scaling Data for AI 49Key Capabilities for Scaling Data 51Introducing a Data Mesh 53When You Have No Data 55The Role of a Data Fabric 56Why a Data Fabric Matters in AI/ML 58Key Competences and Skillsets in Data Engineering 60Chapter 3 Embracing MLOps 71MLOps as a Concept 72From ML Models to ML Pipelines 76The ML Pipeline 78Adopt a Continuous Learning Approach 84The Maturity of Your AI/ML Capability 86Level 0— Model Focus and No MLOps 88Level 1— Pipelines Rather than Models 89Level 2— Leveraging Continuous Learning 90The Model Training Environment 91Enabling ML Experimentation 92Using a Simulator for Model Training 94Environmental Impact of Training AI Models 96Considering the AI/ML Functional Technology Stack 97Key Competences and Toolsets in MLOps 103Clarifying Similarities and Differences 106MLOps Toolsets 107Chapter 4 Deployment with AI Operations in Mind 115Model Serving in Practice 117Feature Stores 118Deploying, Serving, and Inferencing Models at Scale 121The ML Inference Pipeline 123Model Serving Architecture Components 125Considerations Regarding Toolsets for Model Serving 129The Industrialization of AI 129The Importance of a Cultural Shift 139Chapter 5 Operating AI Is Different from Operating Software 143Model Monitoring 144Ensuring Efficient ML Model Monitoring 145Model Scoring in Production 146Retraining in Production Using Continuous Training 151Data Aspects Related to Model Retraining 155Understanding Different Retraining Techniques 156Deployment after Retraining 159Disadvantages of Retraining Models Frequently 159Diagnosing and Managing Model Performance Issues in Operations 161Issues with Data Processing 162Issues with Data Schema Change 163Data Loss at the Source 165Models Are Broken Upstream 166Monitoring Data Quality and Integrity 167Monitoring the Model Calls 167Monitoring the Data Schema 168Detecting Any Missing Data 168Validating the Feature Values 169Monitor the Feature Processing 170Model Monitoring for Stakeholders 171Ensuring Stakeholder Collaboration for Model Success 173Toolsets for Model Monitoring in Production 175Chapter 6 AI Is All About Trust 181Anonymizing Data 182Data Anonymization Techniques 185Pros and Cons of Data Anonymization 187Explainable AI 189Complex AI Models Are Harder to Understand 190What Is Interpretability? 191The Need for Interpretability in Different Phases 192Reducing Bias in Practice 194Rights to the Data and AI Models 199Data Ownership 200Who Owns What in a Trained AI Model? 202Balancing the IP Approach for AI Models 205The Role of AI Model Training 206Addressing IP Ownership in AI Results 207Legal Aspects of AI Techniques 208Operational Governance of Data and AI 210Chapter 7 Achieving Business Value from AI 215The Challenge of Leveraging Value from AI 216Productivity 216Reliability 217Risk 218People 219Top Management and AI Business Realization 219Measuring AI Business Value 223Measuring AI Value in Nonrevenue Terms 227Operating Different AI Business Models 229Operating Artificial Intelligence as a Service 230Operating Embedded AI Solutions 236Operating a Hybrid AI Business Model 239Index 241