Data Resilience and Metadata Management: Critical Components in the AI Era

Council Post: Why Data Resilience Matters Even More In The AI Age

A friend in the healthcare industry recently told me the sort of story that gives CISOs nightmares. The hospital she works for was hit by a sophisticated cyberattack that cut off access to critical datasets. To restore operations, IT had to recover old data from backup systems, a process that took more than a month. The hospital's IT team naturally feared exposing new data to follow-up attacks, so as the recovery process began, the hospital's staff was forced to revert to an old form of data collection. In the age of AI, they had to go back to paper and pen. Today's cyber intrusions are becoming more difficult to detect and more precise in their strikes. Cybercriminals are using large language models (LLMs) to craft phishing emails, minimizing the sort of grammatical errors that typically raise red flags for discerning readers. Yet enterprises are constantly challenged by having to bring their cybersecurity measures up to the level of their attackers, especially with regard to their AI systems. A recent report from IBM and AWS revealed that less than a quarter of generative AI projects are being secured. As large organizations integrate advanced machine learning and AI models and make them an essential part of operational workflows, safeguarding these systems will be critical. But there is another AI-related dimension to cybersecurity the enterprise continues to overlook: enhancing the protection of both the data these tools need to work effectively and the new data they generate. Today, in the age of AI, securing your organization's data may be more important than ever. Data resilience has long been a core tenet of enterprise IT. The difference now is that companies are using data to train and feed their AI and ML solutions. LLMs are so capable because they were trained on a vast corpus of data. After training, when AI models are deployed in the wild, data is equally essential. These advanced systems will not produce valuable insights if they do not have access to secure, cleansed, organized, relevant data. And, importantly, automated workflows that deliver significant business improvements will crater if the underlying data on which those models operate is compromised. Let's say you have a manufacturing business and you are operating a number of plants that produce complex widgets. At each step of the production and assembly process, you have scanners capturing high-resolution, perhaps even three-dimensional images of the components. Manufacturing is already a highly automated industry. AI can be deployed as a driver of automation that helps firms find ways to increase yield and quality while producing more efficiently at a lower cost. To improve quality and failure analysis, for example, you could take those images captured as a component moves through the production line and use AI to pick out potential defects or flaws. Both historical and newly generated data will be essential here. You will need to store a compilation of images to build and train the model. Then this trained model will need the new data to actually do its work and identify potential problems in components as they are produced. This data needs to be fresh, relevant and usable. If there is a cyberattack-related interruption of service or data is encrypted with a strain of ransomware, then the plant's workflow will be completely disrupted. The newly automated failure analysis process will be suspended, and you will have to decide whether to shut down or risk producing flawed components. Neither choice is a good one. Another industry that has led the way in terms of the adoption of AI tools is media and marketing. My company works closely with a number of leading global firms, and one of our customers, a global advertising giant, introduced a custom LLM for in-house creative work and client use. Additionally, these firms have long relied on data from focus groups and studies of consumer behavior, and AI is emerging as a way to generate novel analyses and insights. Again, though, the key variable here is data. If secure, quality data is not available to these models, then they will not produce valuable or reliable results. The good news is that this is a solvable problem. Although cyber threats are becoming more sophisticated and the potential for disruption is greater as organizations rely more on AI solutions, there are built-in security and data resilience tools designed to protect against these threats and, more importantly, accelerate data recovery should such attacks prove successful. The problem is that they are not necessarily an executive-level priority right now. This is a mistake. Organizations need to prioritize planning around data resilience and rapid recovery strategies now more than ever. Protecting the data that feeds AI models must be a core part of an organization's AI playbook -- otherwise, you risk reverting back to paper and pen. Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?

Council Post: Metadata Management: A Critical Discipline That Can Help Unlock AI

Metadata seems to be talked about a lot more now. Marc Benioff, CEO of Salesforce, mentioned it five times since 2018 on earnings calls. In 2024's call, he called it out 26 times. So what is the sudden interest? Metadata is what allows apps to be easily configurable by every customer for their differing requirements. They need to be configured and extended via nontechnical teams, and this is by metadata. Some examples of these enterprise apps are Salesforce, ServiceNow and Workday. Metadata includes database objects and fields, screen layouts, automation, reporting and access permissions. Obviously, there can be many, many more metadata types. These metadata items can be changed or created, and they make up the overall configuration of the app for the organization. For large organizations, there can easily be more than 100,000 metadata configurations, and there are also dependencies between the metadata items. As configuration grows, metadata management is important. Without it, the ability to make changes quickly evaporates. Forrester coined the phrase "Salesforce@scale dilemma," referring to the challenges and complexities organizations face when trying to scale their use of Salesforce. Although they call out Salesforce by name, the issue is true for any other metadata-driven enterprise app: "Every app change risks breaking one of hundreds of data and process customizations, integration links, and third-party add-ons. The result: Every change requires long and expensive impact-analysis and regression testing projects -- killing the responsiveness that made Salesforce attractive at the start." The Change Intelligence Report Series from my company, Elements.cloud, put some hard numbers around Forrester's observations: 50% of custom databases are never used and 41% of custom-created fields are never populated. This is a huge waste of effort and an increase in tech debt. But to drive down tech debt requires better control of metadata. The heart of any metadata-driven app is a metadata dictionary. AI makes that heart beat faster. A metadata dictionary is a listing of all metadata, grouped by type. For every metadata item, it has a listing with additional attributes that are downloaded or calculated for the item, such as version, usage and dependencies. Any metadata item should also be able to have third-party content attached, such as business analysis (requirements, process maps, ERD, user stories), design (specification, wireframes, changelog) and end-user training materials. Without a metadata dictionary, managing applications can become inefficient and risky. The common alternative, using spreadsheets, is difficult to maintain and does not support impact analysis. Enterprise application vendors should include a metadata dictionary as part of their platform. However, most vendors do not provide this feature, although they often offer APIs that enable third parties to create one. Establishing a metadata dictionary involves several actionable steps that can streamline data management and enhance productivity. Firstly, engaging management and building a compelling business case is crucial. Highlight the time saved in analysis and the risks mitigated when implementing changes. This approach not only secures buy-in but also underscores the practical benefits of a structured metadata framework. Consider how implementing a metadata dictionary will impact your support and development processes. Will it streamline operations or improve data consistency across applications? If there's already a similar tool in place, such as a data dictionary or master data management system managed by IT, evaluate its alignment with your specific needs. Ensure your applications support APIs for automatic population and maintenance of the metadata dictionary. Relying on manual updates via spreadsheets can lead to inaccuracies and diminish team confidence and adoption rates over time. When evaluating existing metadata dictionaries, consider key criteria such as user interface and navigability, ability to handle large volumes of metadata, real-time or nightly automatic updates, and features like metadata insights and dependencies visualization. Additionally, assess customization options such as attaching additional content or custom fields to each metadata item, and the availability of dashboards and AI-driven insights. A well-structured and maintained metadata framework can significantly enhance AI capabilities. It provides crucial configuration information for data management and supports various productivity improvements: * Change impact analysis becomes more efficient, allowing queries into unstructured data to prioritize fields based on their importance and impact, estimating analysis times accordingly. * It aids in reducing technical debt by identifying unused metadata items and those without dependencies, streamlining resource allocation. * Documentation quality can be evaluated more effectively, ensuring urgent documentation needs for critical metadata items are prioritized, enhancing overall data governance. A robust metadata dictionary not only optimizes data management but also serves as a foundation for leveraging AI capabilities effectively. By focusing on structured implementation and continuous updates, organizations can unlock substantial productivity gains across their operations. As there are security and hallucination concerns about AI, getting buy-in can be a challenge. However, this use case sidesteps these issues. This involves metadata, not customer data. The risk of hallucination is reduced because you are not asking AI to create content. Instead, you are providing AI with a dataset of metadata and asking it to perform analysis. Nevertheless, you may still face hurdles. Persevere, because the productivity gains are compelling. As enterprise apps become more strategic for customers and the implementations are increasingly complex, sound metadata management disciplines are critical. A cornerstone of that is a metadata dictionary that supports impact analysis and documentation. Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?