Data & AI Governance Unification for Data Executives with Mario Cantin

Executive Summary

This webinar points to the major intersection of AI governance, anomaly detection, and data management within the insurance sector and financial institutions. Mario Cantin emphasises the importance of unity in AI governance, exploring its foundational role in organisational effectiveness and risk management. Howard Diesel focuses on integrating anomaly detection into insurance claim processing, ensuring data quality, and customising best practices for data management. The webinar addresses the necessity of pre-training safety and cybersecurity measures, the significance of fairness in data analysis, and the impact of data management on ROI calculations.

Webinar Details

Title: Data & AI Governance Unification for Data Executives with Mario Cantin
Date: 2025-09-04
Presenter: Howard Diesel & Mario Cantin
Meetup Group: African Data Management Community
Write-up Author: Howard Diesel

AI Governance and Anomaly Detection in Insurance with Mario Cantin

Howard Diesel opened the webinar and shared that in preparation for an upcoming training, Mario Cantin and Howard had been diligently working on a use case related to AI governance, focusing on LLMS and chatbots while addressing sensitive areas within that sphere. Mario then mentioned that his presentation will cover the transition to advanced applications of AI algorithms in anomaly detection for insurance claims, commonly known as “reasonableness checks.”

This method, which was previously used at the central bank to identify significant variances in data submissions, demonstrates the effectiveness of AI in detecting outliers. Additionally, Howard shares that the session will explore how chatbots can support the First Notice of Loss (FNoL) in the insurance process.

The Importance of Unity in AI Governance

The complexity of governance in AI systems necessitates a tailored approach that considers the specific use case of the organisation. For instance, the governance requirements for developing a chatbot differ significantly from those needed for predictive modelling or automated Optical Character Recognition (OCR). A deep understanding of the project’s nature, such as a First Notice of Loss (FNoL) process, is necessary to identify the relevant governance elements. Unified governance can be achieved by recognising these distinct needs and addressing them accordingly.

The application of AI controls extends well beyond chatbots, revealing their significance across diverse agentic AI systems. Effective governance in AI necessitates a complex assessment of various factors, including the specific AI techniques used, the level of risk tolerance, and the compliance requirements that must be met. By adopting this multifaceted approach, stakeholders can better prioritise the elements that matter most in AI development and use, to ensure responsible and effective use of technology.

Exploring AI Governance and Its Importance in the Organisation

AI governance relies on six key pillars: AI frameworks, data management, data operations, risk management, cybersecurity, and ethics and privacy systems. These elements are often managed separately within different sectors of an organisation, creating governance silos. However, for successful AI implementation, a unified approach that considers governance requirements throughout the entire AI lifecycle is necessary. Key governance questions to address include AI eligibility, procurement, adoption facilitation, design, deployment, and operational management, ensuring that all facets of AI governance are integrated and aligned.

Establishing an effective AI governance model is necessary for working through the complexities that can impede progress in AI initiatives. Key considerations include identifying relevant subsets within a structured framework and recognising the balance between external regulations and internal organisational goals, as suggested by the OECD. Importantly, this governance model emphasises two main aspects: the establishment of strong policies, frameworks, and strategies for AI management, along with a focus on stakeholder accountability and capability building. Together, these components pave the way for the successful development and implementation of AI systems.

Effective AI governance involves clearly defining how AI systems should be built and operated, ensuring that governance requirements are transparent and actionable. This process requires the establishment of a prescription mechanism to guide projects in their execution and evidence submission for review. While the governance process is straightforward, focused on ensuring tasks are completed as intended, it can become complex due to the various factors involved, including data management, risk management, and cybersecurity.

There are two key aspects to consider: first, building foundational capabilities to ensure readiness, and second, enabling the success of AI initiatives by creating the necessary evidence to encourage trust in the final products. With recent findings indicating that 95% of organisations see no return on their generative AI investments, it is important to strategically work through these governance processes to avoid falling into this category.

When developing a product that uses your technology, it’s important to consider all necessary facets to ensure a comprehensive framework, particularly in the context of an FNoL chatbot and LLM (Large Language Model) governance. This involves a detailed examination of the structure, personnel alignment, and specific elements related to LLM technology and governance.

Ongoing initiatives in Large Language Models (LLMs) are prompting a deeper investigation into best practices and governance. By using the Prodago solution and the accompanying Playbook as necessary guides, stakeholders are encouraged to focus on gaining a comprehensive understanding of operational details, rather than fixating on the technology alone. This approach aims to enhance the effectiveness and accountability of LLM deployments.

Figure 1 Forces – AI Governance Framework

Figure 2 AI Governance Model

Figure 3 Ensuring Readiness

Figure 4 Enable Success

Requirements for Data Management and AI Governance in LLM Initiatives

The platform, Prodago, features an AI technique Playbook tailored for Large Language Models (LLMs), outlining specific governance requirements pertinent to LLMs. It points to six key pillars: AI Framework, Data Operation, Risk, Ethics, Privacy, and Cybersecurity, each encompassing various associated requirements. The focus will shift from examining these pillars individually to exploring the overall needs for effective AI management and the implications for governance and operation.

Effective governance and operational requirements are important for the successful development of AI systems, especially in the context of LLM technology. This necessitates the establishment of clear AI management structures and specific requirements that facilitate scalability at the enterprise level. Aligning with regulatory frameworks like the EU AI Act and encouraging data and AI literacy among team members are necessary to ensure compliance and effective implementation. Grasping these foundational elements is key to driving successful AI initiatives within organisations.

When developing AI technologies, it is necessary to consider foundational requirements that specific laws and regulations may supplement. For instance, the EUA Act and the DPDPA in India introduce additional compliance and ethical considerations, particularly regarding privacy and the treatment of individuals.

These laws may impose new obligations that must be addressed, emphasising the importance of starting with a clear understanding of the technology’s nature and relevant compliance requirements. The EU AI Act, in particular, points to the necessity of incorporating specific guidelines related to personnel and ethical standards into AI development.

Figure 5 List of Product Playbook

Figure 6 AI Technique Playbook – LLM

Figure 7 Operating Practices

A Deep Dive into AI Initiative and Governance

In developing an AI initiative, it’s necessary to ensure business alignment by incorporating mechanisms that align the system with enterprise priorities and assess risk-benefit trade-offs. This project-based activity requires careful planning and the implementation of elements that facilitate alignment. Team training is important, focusing on responsible AI practices, ethical data management, and broader AI considerations. While many components are involved in the process, not all are major to success, pointing to the importance of strategic prioritisation in project execution.

In exploring AI management, it is necessary to adopt a “crawl, walk, run” framework to identify priorities that align with the organisation’s maturity regarding AI governance. This involves assessing all necessary tasks and selecting a relevant subset tailored for a low-maturity organisation. Key components include establishing a governance responsibility matrix, designating accountable individuals for AI system outcomes, and ensuring compliance with applicable laws while integrating stakeholder management. Training elements are important for effective implementation, alongside the development of structured frameworks to support these initiatives.

An AI project will require a focus on business alignment, team training, and accountability while ensuring stakeholder legal compliance, specifically concerning data operations in the context of Application Lifecycle Management (ALM). Key tasks include clarifying data accountabilities, identifying data sources at both system and business term levels, and thoroughly documenting metadata for training datasets. It is necessary to address these elements comprehensively throughout the project and to develop a data retention plan to meet retention requirements. Each of these aspects constitutes a major activity that must be managed within the initiative.

Figure 8 Operating Practices pt.2

Figure 9 Operating Practices pt.3

Data Operations, Risk Management, and Cybersecurity in AI Development

Data operations play a important role in effective records management, particularly in the context of chatbots and LLMs. This involves not only the management of the data itself but also the content that these technologies use, necessitating careful identification of major data and adherence to open data requirements, especially within the Government of Canada.

A strong governance framework must incorporate various elements such as accountability, metadata retention, data requirements, and risk management strategies. To mitigate potential risks associated with project development, it may be necessary to conduct different types of assessments, including AI risk and AI safety impact assessments. Prioritising these aspects will ensure a comprehensive and responsible approach to data operations.

The implementation of LLMs requires careful consideration of risk management and cybersecurity to ensure safety and security throughout the initiative. This includes conducting red team exercises and assessing business continuity, particularly if the systems are major for the organisation.

Pre-training safety matters a great deal and involves defining filtering criteria for datasets to eliminate undesirable content. Techniques such as domain block lists, model-based filtering, keyword matching, and black box filtering can enhance the safety of the training process. Cybersecurity plays a key role in the training and establishment of LLMs, necessitating thorough planning and risk mitigation activities.

The involvement in defining appropriate data for language model training emphasises the importance of governance concerning data quality and suitability for intended purposes. This includes considerations for data appropriateness across various stages: training, testing, and production.

Best practices from global frameworks inform these efforts, pointing to the need for a comprehensive approach to data governance. Insights were drawn from a detailed survey conducted by Peking University titled “LLM Safety Holistic Survey,” which covers necessary elements related to effective data management in language model development.

Figure 10 Operating Practices pt.4

Figure 11 Operating Practices pt.5

Figure 12 Operating Practices pt.6

Pre-training Safety and Cybersecurity Requirements

Pre-training safety is important for ensuring the integrity of LLMs by prioritising data filtering to eliminate undesirable content. This process is guided by a comprehensive framework that emphasises key-based matching for identifying and removing inappropriate material, while also exploring various techniques and technologies that enhance data quality.

Addressing cybersecurity requirements within this framework is necessary, as it dictates the level of detail and control measures that can be applied to bolster both data integrity and safety in LLMs. A strong focus on pre-training safety lays the foundation for creating more reliable and responsible AI systems.

The deployment of LLMs involves major safety mechanisms that require defining the acceptable distance between fine-tuned and aligned models. This process necessitates governance initiatives that involve stewardship and business expertise, emphasising the importance of techniques such as unlearning and editing.

Strategies to mitigate risks, such as prompt injection attacks, must be considered, alongside the governance frameworks that encompass cybersecurity and privacy rights. Key elements include privacy notifications, consent management, the right to opt out of chatbots, and data minimisation practices, all of which are necessary to ensure the responsible and safe implementation of LLM technologies.

Figure 13 Large Language Model Safety

Figure 14 Large Language Model Safety pt.2

Figure 15 Large Language Model Safety pt.3

Customising Standard Practices, Templates, and Data Quality Assessments

The increasing demand for customizable practice templates has led to the development of out-of-the-box solutions that organisations can easily adapt. These templates can be duplicated and customised while preserving the integrity of the original designs, addressing the longstanding challenge of integrating various frameworks.

To support this initiative, a harmonised semantics list has been created, allowing organisations to implement their own specifications without modifying the necessary components. Alongside this, the creation of metadata templates and a comprehensive meta model detailing the attributes of each template, as well as customizable scorecards, ensures that users have the flexibility to tailor their implementations to meet specific needs. These advancements promise to enhance organisational adaptability and effectiveness in practice.

Effective data governance and quality management are important for enhancing organisational processes. Key concerns include the need for efficient scoring of these processes and verification of their proper implementation. To simplify this effort, it is recommended that the existing 310 templates be consolidated into a single document for easier review, moving away from the cumbersome practice of accessing each program individually.

These templates are primarily structured around the DMBoK framework, supplemented with necessary elements for data governance deliverables and definitions of data quality. Mario is also willing to collaborate with Adam to provide further insights, encouraging a more strong approach to data governance.

The incorporation of a data quality factor rating matrix is necessary for ensuring effective onboarding and training with the Prodago product. This matrix includes various elements, each assigned a metadata tag and a corresponding list of appropriateness and sample data, covering major data elements, quality expectations, and assessments. It also points to any errors present in the file, providing a comprehensive information package that users can use to establish their business processes with relevant artefacts and templates. Users have the opportunity to review this information prior to training, ensuring they are well-prepared for the sessions.

Figure 16 Operating Practices pt.7

Figure 17 Data Quality Expectations

Software Comparative Analysis and AI Governance Requirements

A thorough comparative analysis of competing software is necessary for identifying unique advantages that differentiate products in the market. While Mario has played a important role in developing the meta model and associated metadata, there is still uncertainty about having a comprehensive comparison of features. Understanding what sets a product apart and the distinctive features that make it more appealing than its competitors is key for informed product development. Addressing these considerations will enhance the effectiveness of our strategic approach in the competitive landscape.

The orchestration layer presents significant challenges in data integration due to the lack of existing software specifically designed for this purpose. This gap points to the recognition of certain technologies, such as Prodago, which has been named a two-time Gartner Cool Vendor and identified as a top technology trend for 2025.

Traditional alternatives often involve collaboration with major integrators, but these solutions typically do not provide the necessary harmonised semantics to unify various frameworks effectively. While companies such as Calibra and Informatica offer data governance programs, they often fall short in delivering comprehensive out-of-the-box guidance, requiring organisations to develop many elements themselves. Effective orchestration hinges on creating a cohesive structure that integrates different projects and data governance initiatives.

The assessment of artefacts related to the probability and severity of issues in AI systems is major for effective governance. As illustrated by Howard’s example, this process points to the necessity of incorporating governance requirements when developing systems, especially those that use Large Language Model techniques.

Key considerations such as privacy, consent management, data minimisation, and ensuring human oversight are necessary to encouraging responsible AI use. Addressing these attributes is key to establishing a strong governance framework that supports the successful implementation of AI technologies.

To effectively initiate AI governance, it is important to determine a starting point and identify the necessary components. A structured approach, categorised into crawl, walk, and run phases, helps narrow down the necessary elements to consider. This involves focusing on specific areas such as governance frameworks, alignment, and quality in both AI management and project execution.

The objective is to move from conceptual understanding to tangible actions by identifying relevant subsets on which to act. The governance requirements will vary depending on the project, such as predictive modelling for automating visa decision requests, which underlines the necessity of tailoring your AI governance strategy accordingly.

Figure 18 Operating Practices pt.8

Figure 19 Operating Practices pt.9

Figure 20 Operating Practices pt.10

Understanding the Insurance Claim Processing and Anomaly Detection

The insurance claim processing scenario begins with the major stage of the First Notice of Loss (FNOL), where a chatbot plays a central role. This chatbot assists both claimants and brokers by facilitating the submission of necessary details, verifying the accuracy of the information, and integrating data from various sources, including incident reports, police records, and image recognition technologies.

By simplifying the collection of necessary data, the chatbot not only enhances efficiency but also lays the foundation for effective anomaly detection in the claims process. This ensures that all key information is accurately gathered before proceeding to the next steps in the claims process.

Anomaly detection in data analysis is important for identifying unreasonable data points, often linked to issues such as fraudulent transactions. By using AI techniques, we can develop patterns within the data, making it easier to flag anything that deviates from these patterns as an anomaly.

This intersection between anomaly detection and data quality points to the significance of recognising reasonable failures and reinforces the importance of ensuring data consistency. The potential to use AI in enhancing data quality through anomaly detection presents an exciting opportunity for further exploration.

Figure 21 Anomaly Detection Recipe

Anomaly Detection and Its Integration with Data Quality

Anomaly detection is a process that involves defining thresholds to identify unusual patterns in data, often without the need for supervised learning. An interesting aspect of this process is its relationship with the data quality dimension of reasonableness validation. By using AI techniques for anomaly detection, one can effectively assess the reasonableness of data, such as ensuring that claim amounts do not exceed a predefined limit, for example, 1,000,000. This integration enhances the ability to validate data quality and accuracy.

The implementation of business rules for claim detection requires adherence to an active policy timeframe, with a focus on anomaly detection using dynamic thresholds that adapt based on historical data. As more data is integrated, these thresholds are updated to reflect current trends. The process uses unsupervised learning to identify patterns within historical samples, reducing the reliance on subject matter experts for oversight.

Complementing rule-based checks, the analysis considers multidimensional outliers through multivariate approaches, establishing a feedback loop that identifies anomalies and informs the model of potential challenges. In Power BI, a dataset has been created using three techniques: one-class SVM, isolation forest, and PCA (autoencoder), to classify risk levels into distinct categories based on the data analysis results.

The analysis of 10,000 insurance claims points to the major importance of identifying outliers by incident type, including accidents, fires, medical claims, and thefts. Through the use of synthetic data generation, the findings reveal that accidents present the most significant challenges, enabling a deeper investigation into specific outliers within this category.

By focusing on key features such as claim amount, duration in days, claimant age, policy duration, and previous claims, the analysis provides a thorough understanding of the data elements and their interrelations. This detailed examination of claims encourages the detection of anomalies, enhancing the overall integrity of the claims dataset.

Figure 22 Data Quality Reasonableness Dimension

Figure 23 Insurance Claim Processing Application

Figure 24 Anomaly Review

Figure 25 Anomaly Review pt.2

Transparency and Supervision in Unsupervised Learning

The transparency of unsupervised learning methods, particularly the isolation forest technique, is important for stakeholders to understand and evaluate the system’s outputs. This approach enables the display of scores assigned to individual claims, which helps clarify the rationale behind these scores, such as a score of 0.61 surpassing the threshold of 0.60 set by experts. Effective communication of how these scores are calculated and the contributing factors for flagging claims enhances stakeholder trust.

Using multiple scoring methods, including one-class SVM, not only strengthens the evaluation process but also aids in distinguishing legitimate high scores from potential false positives. Promoting transparency in these scoring mechanisms encourages informed decision-making and builds confidence in the system.

Multiple detection algorithms have proven effective in surpassing established criteria for claim analysis in California, illustrating their major role in identifying anomalies. The discussion points to the intricacies of multivariate claim duration, considering factors such as duration days, claimant hours, and policy duration over a 17-year period. It emphasises the necessity for data stewards to undertake further investigations when anomalies are detected, reinforcing the importance of implementing reasonableness checks. Although the specific causes of these anomalies are not clearly defined, this ongoing process encourages continuous learning and improvement. These efforts empower stewards to identify fraudulent claims more accurately and enhance the quality of unsupervised learning through comprehensive supervision.

The policy in question has been active for 17 years, with a claimant aged 66, and only one prior claim has been recorded. The current claim amount is significantly larger than the previous one and has been open longer than typically acceptable. Analysis of the scatter plot reveals its outlier position compared to other claims, which are part of a larger dataset totalling around 10,000 claims.

Identifying clustering patterns in data is important for effective analysis and decision-making. The presence of light grey dots signifies areas of concentration, prompting a discussion on how data stewards and subject matter experts (SMEs) can address claims that fall outside these identified clusters. To facilitate this process, an agent is anticipated that will enable stewards and SMEs to engage in meaningful discussions, supported by box and whisker plots that enhance their understanding of the data and help identify outliers. This approach aims to improve the accuracy and reliability of claims assessment, ensuring a more thorough evaluation of the data landscape.

The Importance of Fairness in Data Analysis

The concept of reasonableness in data quality is important when evaluating whether data adheres to the required patterns. Unlike other data quality dimensions that operate on a binary scale of right or wrong, reasonableness involves evaluating data that may fall on the boundary of acceptable ranges. This necessitates collaboration with subject matter experts (SMEs) to investigate why certain data points deviate from expected patterns.

The aim is to simplify this evaluation process, particularly for users like David Stewart. The discussion also points to a shift from unsupervised learning to using labelled data to improve supervision, particularly in identifying anomalies through synthetic data labelling, which can help differentiate between true and false anomalies using historical data.

Figure 26 Anomaly Review pt.3

Data Anomaly Detection and Optimisation in Power BI

The process developed involves using synthetic data to identify patterns and inject anomalies, allowing for the testing and validation of unsupervised learning models in detecting these anomalies. In Power BI, a flag indicates whether a data point is an anomaly, pointing to differences between generated and injected anomalies.

The visual representation shows how injected anomalies alter the boundaries, revealing significant variability, with some claims deviating from the norm by as much as 70 units. A detailed drill-through feature provides insights into specific claims, such as one with a duration anomaly of a year, aiding in the assessment of the effectiveness of anomaly detection techniques.

Creating realistic test data is important for effectively identifying system anomalies. Testers often generate perfect records, which can lead to misleading results that falsely indicate the absence of issues within the system. By intentionally crafting flawed test datasets, we can better detect potential errors and system failures. It is key to exclude these anomalies from standardised datasets, like those used for training, to ensure they do not negatively impact the analysis of results. Embracing imperfection in test data is necessary for achieving accurate and reliable system evaluations.

Data standardisation plays a important role in ensuring accuracy, particularly by emphasising the exclusion of anomalies from the dataset. During the discussion, William pointed to the complexities involved in combining multiple algorithms, mentioning that such an approach can sometimes obscure underlying issues and lead to erroneous interpretations. However, he clarified that his own analysis does not employ a combination of algorithms, suggesting a more straightforward approach to maintain clarity and precision in the results.

The implementation of algorithms for error detection is important, as it allows for independent operation without the need to combine results. This process involves generating and splitting data, normalising it, and configuring anomaly detection to ensure accuracy and reliability. A significant advantage of this approach is its integration with Power BI, enabling business users to interactively adjust model parameters and set scoring boundaries. By giving users control over the analysis, this method enhances data insights and empowers decision-making beyond the traditional reliance on data analysts.

Figure 27 Anomaly Detection Recipe pt.2

Figure 28 Anomaly Review pt.4

Understanding the Process of Data Normalisation and Anomaly Detection

When normalising data, it is important to address the differences in scale between variables, such as claim amount, claim duration, and claim age. For example, claim amounts can reach up to 10,000, while claim durations average around 365 days and claim ages can extend to 120 days. This disparity can overshadow other important variables, necessitating the use of Z-scores to standardise the data.

By transforming these features to have a mean of 0 and a standard deviation of 1, we can effectively bring them closer to the same scale, allowing for better analysis. The process involves subtracting the mean from each feature value and dividing by the standard deviation, thereby facilitating a more accurate comparison of claims.

Figure 29 Normalising Labelled Synthetic Data

Figure 30 Standardisation (Normalisation) Procedure

Figure 31 Maula Outlier Review

Figure 32 Business Glossary

Anomaly Detection and Data Management in AI Implementation

The analysis focuses on assessing Z-scores, particularly when their absolute value exceeds 3 or 4, indicating significant deviations from the norm. The evaluation process involves transitioning from unscaled data to scaled and standardised data, which aids in validating each row in terms of data quality. This systematic approach allows for the comparison of multiple variables by normalising the data. It emphasises the importance of splitting the data and labelling anomalies, thereby using effective techniques for identifying and addressing outliers in the dataset.

In the anomaly detection process, a total dataset of 10,000 entries is augmented with 25 injected anomalies, resulting in a training split of 80%, specifically 7,900 entries, alongside separate test and production datasets. The methodology involves using various features to establish thresholds, followed by implementing flags for different detection algorithms, such as one-class SVM, isolation forest, and autoencoder; a score exceeding the threshold results in a flag of one. The findings can then be visualised and inspected using Power BI, allowing subject matter experts to analyse specific areas, pose questions, and validate the results. This analysis has been published on Power BI.com, enabling wider access and interaction with the data for a broader audience.

To create a trusted product, it is necessary to identify a relevant subset of fundamental elements rather than attempting to meet all requirements. This process entails understanding and aligning business objectives with governance requirements, emphasising the importance of agility in connecting these aspects. When collaborating with entities such as the Canadian Department of Defence, the focus should not be on merely building capabilities but rather on establishing mechanisms that facilitate the identification of major governance requirements aligned with project goals. This strategic mindset is important for effectively supporting the overall process.

Many organisations struggle to achieve a positive return on investment (ROI) from Generative AI, with research from MIT indicating that 95% experience little to no return. This low ROI often stems from a lack of capability in aligning AI model development with specific business objectives. To avoid falling into the majority, organisations should focus on identifying the minimal requirements necessary to connect their AI initiatives to their goals, thereby enhancing the direct impact on ROI. Success in this area does not depend on trying to cover every aspect but rather focusing on what’s necessary.

Figure 33 Split Validation

The Role of Data Management in Return on Investment (ROI) Calculation

The evaluation of ROI in AI initiatives necessitates a clear distinction between direct and indirect revenue contributions. While AI products may generate direct financial returns, effective data management plays a major role in enhancing overall ROI through indirect benefits. Organisations must acknowledge that the success of AI implementation relies not only on the technology itself, referred to as the “engine”, but also on strong data management, which acts as the “car” that drives these initiatives forward. Understanding the percentage contribution of data management to the overall benefits of an AI project is necessary for grasping its true value.

Data management plays a important role in generating a return on investment (ROI) for business initiatives, with its contribution quantifiable as a percentage, often considered an indirect cost. When implementing a large language model (LLM) in a project-based approach, it is necessary to recognise that effective data management should not be seen as a one-time effort but rather as an ongoing responsibility. This means that even if an LLM is successfully integrated into a project, there needs to be a sustained focus on data management to ensure long-term success and to accurately assess its impact on the overall ROI.

In evaluating a business initiative, it’s necessary to consider the return on investment (ROI), which includes both direct and indirect costs associated with data management. Effective data management prevents the need for additional expenditures on filling in gaps due to poor initial data handling. Data management should be integrated into the project from the start, acknowledging its ongoing costs rather than treating it as a one-time expense. The ROI from data management should be assessed over several years, typically spanning three to six years, with regular recalibrations each year to ensure accurate valuation of its contribution to the initiative.

Data Management and AI in Financial Institutions

In the context of financial institutions, data management groups often face challenges regarding their contributions to the bottom line. One effective approach to demonstrate their value is by identifying necessary data management elements that are important for the success of AI initiatives, even if these elements are not directly linked to business objectives.

This shift from compliance-driven to value-driven data management can be achieved by aligning data management strategies with business goals, particularly through the lens of AI objectives. Many organisations are recognising this need, prompting Chief Data Officers (CDOs) to strategically integrate their initiatives with broader business objectives, thereby enhancing overall value.