With the growing use of AI, new risks are emerging that require mitigation. These risks relate to bias and discrimination, trust and transparency, privacy and ethics, safety, security, finance, and corporate reputation. More risks and more laws and policies to regulate the use of AI have led organisations to start adopting guidelines, practices, roles and tooling for AI governance, including “responsible AI” to limit their AI risk.

Why Trending:

• Although many people are excited about AI’s potential, there are also deep concerns about its growing power and proper usage. Many examples in public media about unethical or malfunctioning AI have greatly raised awareness among consumers, policymakers, politicians and companies about AI risk.
  
• Organisations are increasingly aware that scaling AI from local use and prototyping to companywide production systems means that AI is ever more business-critical. Thus, scaling the use of AI introduces significant risks that require monitoring and mitigation.
  
• Governments and corporations are introducing more policies, guidelines and legislation to regulate the use of AI, balancing its innovative power with the need to protect privacy and avoid unfairness in more automated and opaque AI-powered decision-making.
  
  

Implications:

• A growing number of industry organisations, open-source initiatives, commercial tech firms and others offer guidelines, tools and techniques for AI governance and responsible AI that companies can use to start supporting their own AI risk management initiatives.
  
• Risks include but are not limited to ethical risks such as bias, discrimination or privacy violations. Attention toward risks (e.g., the security of AI itself, the risk of model theft, the poisoning of training data or the circumvention of fraud detection) is also growing. Other risks include financial damage or customer churn in the use of chatbots or automated decision-making.
  
• Managing AI risks is not only about being compliant with regulations. Effective AI governance and responsible AI practices are also of key importance to building trust among stakeholders and to catalysing the use and adoption of AI.
  
  

Actions:

• One option is the Gartner TRISM framework as a structure to implement and execute risk management for AI.
  
• To balance the managing of risks with the required flexibility for experimentation and innovation with AI, follow an adaptive approach to AI governance.
  
• AI governance requires a combined approach of guidelines, processes, roles and techniques. Adopt these as part of the broader emerging discipline of AI engineering while leveraging the growing support for fairness, model monitoring, security and other aspects of AI platforms.
  
• Many AI risks, in particular ethical risks, pose challenges and dilemmas that cannot be fully regulated through guidelines. Instead, they require the awareness of practitioners and the involvement of stakeholders, preferably through an AI governance board.

Observability in Managing Your AI Risk

Description:

Observability is a characteristic that allows the system’s behaviour to be understood based on its outputs and enables questions about their behaviour to be answered. Observability in the D&A system allows the practitioners to monitor and comprehend its behaviour across multiple components within the ecosystem by observing external outputs such as its activities, measurements, requests and dependencies. It enables D&A personnel to control the system’s usage, management and inference to make timely, cost-effective business decisions using reliable and accurate data.

Observability can be applied in a wide range of areas within D&A, including but not limited to data quality, data drift, data pipelines, data orchestration, management operations, analytics, ML model deployment, systems performance, and infrastructure with AI-enabled optimisations and recommendations.

Why Trending:

• With modern D&A systems increasing in complexity and distribution while generating a large volume of data, a significant effort is required to parse, understand and facilitate changes so that they become more adaptable to deliver new capabilities.
  
• D&A platforms that are distributed combine multiple architectural patterns and service tiers. These include data storage, data fabric, metadata management, data integration, data analytics and DSML deployments, and they will require a comprehensive and integrative approach to observe, inform and deliver recommendations with better governance to achieve business goals.
  
• Enterprises adapting to agile development practices and accelerated data-driven transformation using xOps are seeing an increasing demand to gain comprehensive visibility into the entire D&A life cycle. This includes processing, orchestration and execution of data, ML models and pipelines supported by machine learning and predictions to improve D&A system performance, uptime, trustworthiness and total cost of ownership.
  
• Traditional methods for managing data pipelines are no longer sufficient in preventing downtime. Disparate interfaces for monitoring events, the inability to correlate multiple faults in a pipeline to rapidly identify root causes, and statically defined rules to monitor data cannot accommodate the demands of modern data pipelines to ensure reliability and consistency.
  
  

Implications:

• More organisations are seeing an increasing need to monitor, track and record all activities, including actions and movements. The hope is to help with the impact and root cause analysis for issues and offer recommended changes for better management and deployment of their D&A applications.
  
• Agile development practices and accelerated data transformation using dataOps, microservices and serverless functions contribute to an increasing demand for accelerated delivery, greater details and visibility to mobilize data-driven decisions.
  
• Every organization looks at observability differently depending on what it wants to track and trace, as well as on resources and capabilities, to identify issues and address them proactively before the failure happens.
  
• Effective cost management on cloud analytics deployments using FinOps requires the intelligence of data observability in the capacity planning processes to track resource utilization to align with business growth targets and to improve performance.
  
  

Actions:

• Develop practices to track every aspect of your D&A systems and their components. Start with measuring data usage and quality to demonstrate the business value of data observability supported by machine learning and predictions.
  
• Establish a robust data collection process that includes activities, measurements, requests and dependencies across multiple systems, services and components to develop better insights.
• Understand your D&A system’s current state based on the movement, access and utilisation data it generates. In this way, you can proactively learn, understand and visualise through applied intelligence to determine predicted outcomes and be proactive in operating D&A systems efficiently and at reduced costs.
  
• Manage your D&A systems to not only observe and improve infrastructure, data collection and AIOps processes but also gain insight into how the consumers are interfacing and interacting to ensure that they are getting the best experience.