May 2026 Summaries

Hex introduces a new capability allowing its agent to integrate with Notion, Linear, and other tools that expose an MCP server, thus enabling a more comprehensive understanding and application of data within organizations. By connecting Hex to existing tools, teams can merge raw data from their warehouse with contextual information such as strategy documents, meeting notes, and support tickets, enhancing the analytical process beyond mere numbers. This integration respects existing permissions and offers an annotation layer that enriches data interpretation, providing deeper insights and actionable intelligence. The Hex agent can cross-reference data with contextual knowledge to address queries, such as identifying the causes of changes in metrics or aligning performance with business goals, effectively transforming data into a driver of work rather than just an informational asset. By working across connected tools, Hex ensures that responses are grounded in validated data models and business context, fostering a collaborative environment where projects serve as lasting records of analysis and decisions.

Hex has developed a unique evaluation infrastructure called "The Shoebox" to enhance the performance and assessment of data agents, particularly within the complex domain of data analytics. This system allows for ad-hoc and scheduled evaluations, focusing on pairwise experiments to compare a "candidate" against a "baseline" run, thus facilitating detailed and consistent assessments. To address the scarcity of suitable analytic benchmarks, Hex created a hypothetical business, Shorelane Commerce, which provides a realistic data set reflecting the challenges faced by data agents in real-world scenarios. The Shoebox integrates deeply with Hex's application, allowing for seamless updates and evaluations, though it requires significant maintenance and coordination to ensure consistency across local and production environments. Despite these challenges, the infrastructure supports the development of innovative features and provides a flexible platform for engineers to test new ideas and configurations, making it a valuable asset for Hex's ongoing advancements in data analytics.

Behavioral segmentation for data teams involves using SQL and Python to define user cohorts based on their actions, such as purchase patterns, login frequency, and feature use, rather than relying on dropdown menus for audience traits like marketers do. This process requires transforming raw data from warehouses into reliable, queryable segments through patterns like RFM scoring, cohort retention analysis, engagement scoring, and churn risk detection. These segments are managed by dbt models and involve creating common table expressions (CTEs) with window functions, scoring users on activity patterns, and maintaining consistent metric definitions across various teams. The approach emphasizes the importance of governance, advocating for version-controlled, centrally governed segment definitions to prevent inconsistency and maintain trust. The article also highlights the significance of making segmentation actionable for non-data teams via interactive tools and conversational analytics, ensuring stakeholders can explore and adjust segments without needing additional data team intervention.

Data visualization can be significantly distorted by unclean data, which often manifests in charts with misleading representations, such as duplicate entries inflating totals or inconsistent units causing false anomalies. To address these issues, data teams can follow six essential cleaning steps: handle nulls and missing values deliberately, standardize formats and deduplicate data, triage outliers with a defensible rationale, utilize charts during the cleaning process, maintain reproducibility of cleaning procedures, and automate preventative measures. These practices are crucial for ensuring reliable and trustworthy visualizations, especially as data quality remains a primary concern for data teams and a barrier to AI adoption. Implementing these steps in an iterative cleaning process, supported by environments that integrate SQL, Python, and visualization tools, can enhance the accuracy and trustworthiness of data-driven insights.

Ad hoc queries are a pivotal yet often challenging aspect of data analytics, serving as one-off, user-defined requests to address specific questions outside of existing dashboards and reports, typically supporting high-stakes decisions such as strategic pivots or market entries. These queries are crucial because they provide insights into unexpected questions that pre-built dashboards cannot answer, but they also introduce several challenges, including maintaining data quality, metric consistency, and preserving valuable analysis that often gets lost in ephemeral formats like Slack messages or individual notebooks. The default reliance on ad hoc queries is due to their ability to quickly address urgent questions, yet this approach can lead to a cycle of repeated queries, inconsistent metrics, and ultimately, a loss of good analysis. Implementing governance in the ad hoc workflow, such as defining metrics in a shared semantic layer and creating a query library, can mitigate these challenges, ensuring that ad hoc queries are accurate and reusable while enabling teams to focus on strategic initiatives. By embedding governance into the workflow and facilitating self-service analytics, organizations can transform ad hoc queries from isolated efforts into structured, durable artifacts that enhance decision-making and reduce reliance on overburdened data teams.

Drill-down analysis is a data exploration technique that enables analysts to navigate from aggregate summary metrics to more granular details within a predefined data hierarchy, helping to diagnose underlying causes of anomalies in metrics like revenue. Unlike filtering, which is ad hoc and reduces datasets based on chosen criteria, drill-down is sequential and maintains nested context, moving step-by-step through a hierarchy such as Year → Quarter → Month, or Country → State → City. It is crucial in diagnostic analytics for transforming vague metric alerts into specific findings by systematically isolating factors contributing to deviations, such as identifying that revenue decline is due to a specific product issue in a certain region. While drill-down is highly effective when paths are anticipated in advance, its limitations emerge when the required investigative paths were not predefined, necessitating flexibility to switch to ad hoc queries or other analytical methods. Success in drill-down analysis relies on well-structured data models, as poor hierarchy design can lead to aggregation errors and investigative dead-ends, underscoring the importance of collaboration between data engineers, analysts, and business stakeholders in designing hierarchies that reflect real investigative workflows.

The blog post explores the distinction between data literacy and data fluency, emphasizing their importance in the context of AI's increasing role in analytics. Data literacy is defined as the ability to read and interpret data, such as understanding charts and basic statistics, while data fluency extends this by incorporating judgment, enabling individuals to evaluate the soundness of analyses and effectively communicate findings. The text argues that while literacy provides foundational skills, fluency involves deeper critical engagement and is necessary for making informed decisions. As AI automates technical tasks, the need for critical evaluation skills grows, highlighting the importance of building organizational fluency to prevent misinterpretations and ensure trustworthy decision-making. The blog stresses that improving data literacy is an essential first step, but achieving fluency requires real-world practice, feedback, and robust governance, particularly as AI changes the landscape of data work.

Dashboard adoption often falters because visualizations fail to answer the specific questions stakeholders have, leading to underutilized tools. Stephen Few's framework suggests focusing on the message and question before selecting a chart type, which can guide effective visualization decisions. The document outlines various chart types, such as bar charts for comparisons and line charts for time-series data, emphasizing the importance of choosing the right chart for the question at hand. It also highlights common pitfalls, such as metric inconsistency and fragmented toolchains, which hinder effective data communication. To address these issues, the text suggests implementing governed metrics and interactive outputs to enhance decision-making and trust. Hex's collaborative platform is presented as a solution, integrating SQL and Python workflows with visualization and metric consistency to maintain a connected analytical process.

ETL pipelines, which stand for Extract, Transform, Load, are automated systems that streamline data integration by extracting data from various sources, transforming it into a consistent format, and loading it into a destination like a data warehouse for analysis. These pipelines help address the challenge of scattered and inconsistent data by ensuring data quality and enabling teams to query and analyze it effectively. While ETL focuses on transformations before data enters the warehouse, ELT (Extract, Load, Transform) allows for transformations within the warehouse, offering flexibility and efficiency, particularly in modern cloud environments. The choice between ETL and ELT often hinges on industry requirements and the desired level of data quality enforcement. Common tools for building ETL pipelines include managed connectors like Fivetran for data extraction, dbt for transformations, and orchestration tools like Airflow. The effectiveness of these pipelines is bolstered by automated tests and freshness monitors, which help maintain data integrity and trust. As AI becomes more integrated into data workflows, pipelines are evolving to include AI-driven anomaly detection and natural language querying, enhancing the utility and accessibility of data infrastructure.

Underfitting in machine learning is a common issue where a model fails to capture the underlying patterns in the training data, resulting in high errors on both training and validation sets, characterized by high bias and low variance. This occurs when the model is too simple, excessively regularized, or has insufficient training, and is often misinterpreted as a data volume problem rather than an issue with model capacity or feature representation. The article emphasizes the importance of diagnosing underfitting through the bias-variance tradeoff and using techniques such as increasing model complexity, reducing regularization, and engineering better features. It also highlights the deceptive nature of underfitting, especially when AI systems generate confident but incorrect interpretations based on flawed models, leading to a lack of visible failure signals. Effective management involves a continuous, iterative process of balancing bias and variance, using diagnostics like learning curves and cross-validation, and maintaining a collaborative, reproducible workflow to trace improvements accurately.

Data orchestration is a crucial process that automates and coordinates data workflows across systems, ensuring tasks are executed in the correct order and dependencies are managed effectively, often using Directed Acyclic Graphs (DAGs) to maintain order and reliability. Despite its efficiency in managing data collection, transformation, and activation, orchestration alone cannot solve the "last mile" problem, where technically available data fails to drive informed decisions due to issues like inconsistent metrics, lack of trust in dashboards, and the proliferation of ad hoc requests. Extending governance into the analytics layer becomes essential to bridge this gap, emphasizing the importance of a semantic layer that defines consistent metrics, access controls that apply across all tools, and lineage beyond pipeline execution to ensure reliable and trusted data delivery. This approach facilitates self-service analytics, where business users can confidently access and utilize data without overwhelming analysts with repetitive requests. As orchestration and governance integrate, it enables more strategic work for analysts and more reliable, context-grounded insights for business users, thereby improving overall data trust and utilization.

Feature engineering is a critical aspect of data science that involves transforming raw data into a format that machine learning models can utilize effectively, often proving more crucial than the choice of algorithm. It includes selecting significant variables, modifying existing ones through scaling or encoding, and creating new features to convey domain-specific insights. Techniques vary by data type, such as standardization for numerical features or one-hot encoding for categorical data. Automated tools like Featuretools can generate many candidate features, but practitioner judgment is essential to select those that truly add value. The process is time-consuming, often constituting the bulk of project work, but is vital for producing models that perform well in real-world applications. Ensuring reproducibility with pipelines and avoiding pitfalls like data leakage are key best practices. While deep learning can automate feature extraction in unstructured data, traditional feature engineering remains significant for structured data, underscoring the importance of domain knowledge in shaping effective models.

Hex is introducing a credit model for its Hex Agents, offering monthly credit grants based on user roles and enabling organizations to monitor and manage their usage effectively. This model supports different tiers, such as Professional, Team, and Enterprise Editor seats, each with varying credit allocations, and allows for the purchase of add-on credits with options to set monthly spending limits. The effort-based consumption of credits reflects the complexity and resources required for tasks, with a focus on cost efficiency by directing tasks to models that offer the best balance of cost and accuracy. Hex provides comprehensive visibility into credit consumption through its Context Studio, which allows users and admins to analyze usage patterns, optimize context curation, and enhance data accuracy. As more organizations integrate AI, Hex aims to help leaders understand and manage costs while ensuring efficient use of credits.

Katie Bauer, Hex's Head of Data, shares strategies for optimizing AI credit usage within the Hex platform, focusing on maximizing efficiency rather than minimizing costs. The effort-based AI pricing model in Hex implies that the cost is determined by the amount of work an AI agent must perform, influenced by factors like context gathering, question complexity, and back-and-forth interactions. Bauer emphasizes the importance of building robust context to minimize AI effort, suggesting tools like Context Studio for proactive management and user education to improve prompt efficiency. Moreover, she highlights using admin controls to manage and predict AI credit usage, framing AI expenditure as an investment in data accessibility and organizational efficiency. Bauer suggests aligning financial expectations and discussing the return on investment by considering specific use cases where streamlined data access could significantly benefit business operations.

Account 360, initially heralded as a groundbreaking customer health dashboard at Hex, was ultimately replaced by Threads due to its limitations in speed and adaptability. While Account 360 aggregated scattered data and improved over previous fragmented workflows, it became cumbersome as the data team faced constant requests to add new insights, making the platform slow and difficult to navigate. Its inability to provide real-time answers or analyze multiple accounts simultaneously led to its decline in use. Threads emerged as a more dynamic solution, enabling team members to directly query data and receive immediate, reliable insights, thus transforming customer interactions by offering deeper narratives and understanding. This transition underscores the evolving standards for customer intelligence, emphasizing speed, accessibility, and the need for systems that facilitate exploratory conversations rather than static dashboards. Despite Account 360's obsolescence, its foundational data continues to support the advanced functionalities of Threads, illustrating the necessity of continually raising the bar in customer intelligence tools. The shift reflects Hex's commitment to creating impactful, interactive data products and highlights the ongoing evolution in how customer teams access and utilize information.

Hex introduces a novel approach for data teams by allowing them to connect their Git repositories to its platform, thereby enabling a more comprehensive understanding of data beyond mere content. This integration allows Hex's agent to analyze, parse, and synthesize information from dbt and application repositories, providing insights into the logic behind data, which traditionally only engineers could access. By doing so, it empowers self-service users to tackle complex queries and understand data lineage with confidence, as they can now access nuanced upstream logic and product functionality without solely relying on the data team. This capability enhances the ability to answer intricate questions about data filtering, feature implementation, and event analysis, while ensuring that users have an intelligent agent that synthesizes context from multiple sources, leading to more impactful data products.

Hex has introduced Generative Data Apps in beta, enabling users to harness AI-generated code to build customizable data applications within its secure and governed platform. These apps allow for highly flexible designs, such as interactive charts and tailored visualizations, while ensuring data governance, traceability, and security. Hex's platform integrates context like semantic models and existing dashboards, allowing data teams to maintain a clear line of analysis and logic. Each app is a fully traceable Hex project, facilitating easy updates and maintenance without separate deployment steps. The platform's security measures ensure that data remains protected within a controlled environment, preventing unauthorized access and data leaks. Currently available to Editor+ roles, Generative Data Apps will eventually transition to a usage-based AI credit pricing model.

AI analytics adoption presents several pitfalls that organizations should be aware of, primarily rooted in governance issues and data quality rather than the AI models themselves. Despite the enthusiasm for AI's potential, only a small percentage of organizations prioritize its implementation, resulting in fragmented data insights and inconsistent metric definitions. Ungoverned AI usage can lead to security and compliance risks, as unauthorized tools create inconsistent business logic. The success of AI analytics hinges on integrating them into a unified workspace with trusted context and endorsed data sets, rather than relying on disconnected tools and raw data. Effective AI deployment requires a balance between governance and progress, ensuring that AI systems are grounded in reliable, curated business knowledge. Organizations are encouraged to incrementally develop their semantic layers and observability practices to improve AI answer quality, while also recognizing that data quality issues often underlie performance problems attributed to AI models.

Hex has developed a sophisticated topic discovery system to enhance its Context Agent, allowing users to ask questions in natural language and receive accurate responses. Initially, the team attempted a machine learning approach with traditional clustering algorithms like HDBSCAN and UMAP, but these proved inefficient due to high computational demands. Instead, they pivoted to a simpler, more effective method using large language models (LLMs) that improved topic quality by separating topic discovery and thread classification into distinct stages. This system operates by analyzing thread summaries to propose topics and then classifying new threads accordingly, ensuring ongoing topic refinement. Despite the limitations of a flat topic structure, customer feedback has been positive, and Hex plans to explore further enhancements, including topic-specific guides and evaluation suites, to better understand and address user needs. This approach underscores the value of leveraging LLMs to simplify complex processes, reducing the gap between concept and implementation, and reflects Hex's commitment to creating impactful data products.

In an era where traditional engineering constraints are disappearing, destination engineers, who prioritize user experience over code elegance, are emerging as key players in the tech industry. This shift is driven by the rapid advancement of technology and AI, which has blurred the lines between product, design, and engineering, allowing engineers to engage more directly with users and make decisions on what is viable to ship. As routine tasks become automated, engineers now spend less time on boilerplate work and more on critical decision-making, effectively elevating their roles and responsibilities. While this transformation challenges the conventional role of project managers, it also expands their scope, requiring them to focus on strategic decision-making and path-clearing rather than micromanaging specifications. The evolving landscape has made product development more dynamic and engaging, emphasizing the importance of solving the right problems for the right users, which positions destination engineers at the forefront of innovation.

Ground truth in machine learning is the verified, labeled data that acts as the benchmark for training, validating, and evaluating models, but its implementation is fraught with challenges such as high costs, human error, and degradation over time. In supervised learning, ground truth represents the correct labels from which a model learns and evaluates its predictions against. The concept, borrowed from fields like remote sensing, highlights that ground truth can often be subjective, with labels derived from human judgments or proxy measurements. As AI technologies advance into production analytics, the accuracy of ground truth becomes critical for business decision-making, necessitating robust data quality monitoring and context layers to ensure AI agents operate from a reliable foundation. Maintaining ground truth involves addressing issues like label leakage, annotation drift, and distribution mismatch, which can all lead to model failures if not managed properly. By treating ground truth as a dynamic dataset and incorporating best practices such as defining clear labeling schemas, measuring inter-annotator agreement, and establishing feedback loops, organizations can improve model reliability and trust in AI analytics.

Business logic refers to the rules and calculations that translate raw data into meaningful metrics for organizational decision-making, and its mismanagement can lead to discrepancies in reported figures, as illustrated by differing revenue reports from finance, marketing, and data warehouses. The problem arises from decentralized implementations of business logic across various tools and platforms, causing metric drift that erodes trust in data reliability. To address these issues, the article suggests centralizing business logic through documentation, version-controlled transformations, and formalized semantic models, which ensures consistent and accurate data interpretation across the organization. This centralization not only mitigates metric inconsistencies but also enhances self-serve analytics and AI adoption by providing a stable reference point for data queries. The piece emphasizes the importance of clear ownership and collaboration between business stakeholders and analytics engineers to maintain and update metric definitions, as demonstrated by successful case studies like Calendly’s Standardized Metric Library. By establishing shared definitions and explicit ownership, organizations can transform data disputes into strategic discussions about leveraging insights for decision-making.

Feature leakage in machine learning refers to the inadvertent inclusion of information during model training that would not be available during actual predictions, leading to models that perform well during validation but fail in production. This issue can arise from several sources, including target leakage, train-test contamination, and temporal leakage, where future data is incorrectly used to predict past events. Such leakage often remains undetected because it inflates both training and test metrics simultaneously, making models appear more reliable than they are. Effective prevention strategies include conducting exploratory data analysis (EDA) on properly split datasets to identify suspiciously high correlations and dominant features early on, and ensuring preprocessing steps are fit only on training data. Collaborative, reproducible workflows are emphasized as essential, allowing for thorough peer review and auditing of the data pipeline to catch leaks that automated checks might miss. AI-generated code, while speeding up development, can introduce new leakage risks, underscoring the importance of human oversight. By maintaining transparent and versioned analysis environments, teams can minimize the risk of feature leakage and build more trustworthy models.

Data analytics and AI have the potential to significantly impact business outcomes, but many organizations struggle to connect their analytics efforts to tangible results. Traditional analytics workflows often create a disconnect between insights and decision-making due to slow response times and inconsistent data definitions. With the integration of AI, the process can become more efficient by allowing direct questions and reducing the time from inquiry to insight. However, AI's effectiveness depends heavily on clean data foundations and governance, as inconsistent definitions can lead to faster delivery of incorrect insights. Successful implementation requires teams to prioritize semantic layer development, ensure consistent definitions across platforms, and maintain strong data governance to avoid security and accuracy issues. Case studies from companies like PandaDoc, ClickUp, and Ramp illustrate how integrating AI and analytics into workflow systems can streamline processes and enhance decision-making. Ultimately, the focus should be on creating a culture where analytics inform decisions, supported by trustworthy data governance and AI tools that allow for transparent query verification.

Data-driven change management requires a cohesive approach to measurement and data analytics, emphasizing the need for consistent metric definitions and alignment across teams to track progress effectively. Disparate data sources and inconsistent definitions often lead to mistrust in adoption figures, resulting in stalled change initiatives. Establishing clear metric definitions, mapping data sources, and designing dashboards that endure beyond initial launch phases are crucial steps in ensuring successful change management. The role of data teams is strategic, involving co-creation of definitions, connecting data signals across systems, and enabling self-service analytics for stakeholders. This approach reduces dependency on ad hoc requests and enhances trust in the data being used to make decisions. Hex provides a collaborative workspace that integrates these elements, supporting organizations in making change management truly data-driven by facilitating structured independence, from governed metric definitions to interactive applications.