Recent News in Analytics and AI: October 2025 Edition

October’s roundup of recent news in analytics and AI highlights a diverse set of advancements: new capabilities in Power BI, major security and performance updates in Microsoft Fabric, significant enhancements to Microsoft 365 Copilot and Copilot Studio, and key developments across Azure and the open-source analytics and AI ecosystem. This edition also covers industry-shaping moves such as the dbt Labs and Fivetran merger, the evolution of agentic AI, and practical innovations in data modeling and risk simulation.

Read on and get up to speed.

Power BI

• A recent article on the Power BI blog deep dives into how Direct Lake and import tables can be combined in the same model. This capability offers greater flexibility for complex analytics scenarios, enabling richer relationships and improved performance without sacrificing speed. By mixing OneLake data with other sources, users can create advanced models that support calculated columns, hierarchies, and more. The blog explains how this works in both Power BI Desktop and the service, making it easier to design scalable solutions for enterprise reporting. Learn more.

• Previously limited to Power BI Desktop, the ‘Prep Your Data for AI’ feature is now available in the Power BI service, allowing users to configure Copilot schemas, verified answers, and AI instructions directly in the browser. This shift reduces friction by enabling Copilot readiness without switching tools and supports additional scenarios like Direct Lake models. The goal is to make semantic models AI-ready faster and more consistently, unlocking richer insights and improving organisational confidence in Copilot-driven experiences. Learn more.

• The ‘Report Copilot’ experience has been upgraded to deliver smarter, more intuitive report creation and editing. Copilot now offers better visual recommendations, a broader visual library, and improved context awareness for nuanced prompts. Beyond generating new pages, users can collaborate with Copilot to refine existing reports; adding, replacing, or removing visuals with full undo/redo support. These enhancements make building and iterating on reports faster and more flexible, with the updated experience available in the Power BI service and coming soon to Desktop. Learn more.

• A new ‘App-scoped Copilot’ feature brings AI-driven insights into curated app experiences. Users can ask questions or request summaries based on reports within an app, ensuring responses stay relevant to the app’s context. App authors can enable verified answers for trusted, human-approved responses, while users benefit from streamlined navigation and faster access to insights. This complements report-level Copilot, creating a layered AI experience across Power BI apps. Learn more.

Microsoft Fabric

• Microsoft Fabric now supports Row and Column Level Security in Spark through a dual-environment architecture that separates user code from secure data preparation. When a Spark job accesses protected tables, policies are applied in a secure cluster before data reaches the user code, ensuring compliance without manual configuration. Direct file-level access is blocked; data must be queried via namespace references like lakehouse.schema.table. To activate secure clusters, users must pin a schema-enabled lakehouse as default. The feature is now in public preview and scales dynamically with query demand. Learn more.

• Microsoft Fabric announced the full public preview of OneLake security, a fine-grained access control model that enforces consistent data protection across engines like Power BI, Spark, and Copilot. Users can define roles with folder-, row-, and column-level permissions, ensuring sensitive data like PII is automatically restricted. OneLake security integrates seamlessly with shortcuts, enabling secure data sharing without duplication. Recent updates include faster RLS/CLS queries, improved Spark notebook startup, and enhanced SQL Endpoint diagnostics. All workspaces using data access roles have been migrated automatically. Learn more.

• Microsoft Fabric Spark introduces Adaptive Target File Size, a new feature that automates file size optimization for Delta tables. Enabled via a single session setting, it dynamically adjusts target file sizes based on table growth, from megabytes to terabytes, eliminating manual tuning. This ensures consistent performance across operations like OPTIMIZE, Auto Compaction, and Optimized Writes. Users can also override defaults with a unified table-level property. Benchmarks show up to 30% faster ELT cycles and improved query speed. The feature is available now in Fabric Spark Runtime 1.3. Learn more.

• Fabric Data Agent now supports CI/CD, ALM flow, and Git integration, enabling structured, version-controlled development of data agent configurations. Users can manage schema selections, data source setups, and instructions across Lakehouse, Warehouse, Power BI models, and KQL databases. Git integration ensures every change is tracked, reviewable, and reversible, supporting collaborative workflows with branching and pull requests. Deployment pipelines allow updates to move through development, test, and production workspaces, reducing risk and improving reliability. This approach aligns data agent management with modern software engineering practices. Learn more.

Microsoft 365 Copilot and Copilot Studio

• Microsoft 365 Copilot introduces ‘Researcher with Computer Use’, a major upgrade that enables autonomous AI to act on users’ behalf via a secure virtual machine. This enhancement allows Researcher to navigate gated and interactive web content, access premium sources, and generate rich outputs like presentations and code, all while maintaining user control and enterprise-grade security. Powered by Windows 365, the sandboxed environment includes a browser, terminal, and safety classifiers to ensure secure, task-relevant execution. Benchmarks show significant performance gains in complex research tasks, with rollout beginning via the Frontier program. Learn more.

• Microsoft 365 Copilot Web Search delivers real-time answers with enterprise-grade protection through four layers of control: Admin Controls, User Protections, Query Safeguards, and Contractual Commitments. Admins can scope access, audit usage, and enforce DLP policies, while users can toggle web search and inspect query keywords. Copilot strips identifiers from queries and securely transmits only essential terms to Bing. Microsoft commits not to use query data for ads or model training, treating it as Customer Confidential Information. This layered approach ensures secure, transparent, and policy-driven web grounding. Learn more.

• Microsoft 365 Copilot now empowers users to build apps, workflows, and agents using natural language, thanks to new tools like App Builder and Workflows. These features allow employees to create dashboards, automate tasks, and develop conversational agents, all within the Copilot experience. Outputs are secure, governed, and integrated with Microsoft 365 data, ensuring compliance and ease of sharing. App Builder enables rapid app creation without database setup, while Workflows automates tasks across Outlook, Teams, and more. These capabilities are available to Frontier program users, with broader rollout expected soon. Learn more.

• Microsoft Copilot Studio now includes Agent Evaluation in public preview, enabling structured, automated testing of AI agents directly within the development environment. Makers can create evaluation sets using manual inputs, recent interactions, or AI-generated queries, and apply flexible test methods such as exact match, semantic similarity, and intent recognition. Success criteria can be customized to align with business needs, and results are presented with clear pass/fail indicators and detailed scoring. This feature supports continuous improvement and model comparison, streamlining the agent lifecycle from build to deployment. Learn more.

Azure Analytics and AI

• Microsoft introduced the Agent Framework, an open-source SDK and runtime for building, deploying, and managing multi-agent AI systems. It merges the enterprise-grade stability of Semantic Kernel with the experimental orchestration of AutoGen, offering developers a unified foundation for agentic applications. The framework supports both creative, LLM-driven agent orchestration and deterministic workflow orchestration, with built-in observability, compliance, and security. It’s designed for extensibility, open standards, and production readiness, and is already being adopted by enterprises like BMW, KPMG, and Fujitsu for real-world AI solutions. Learn more.

• Azure AI Foundry released a comprehensive guide to smarter fine-tuning, offering developers a streamlined path to customize large language models for domain-specific tasks. The guide outlines best practices for data preparation, model selection, and iterative evaluation, emphasizing techniques like Reinforcement Fine-Tuning (RFT), Direct Preference Optimization (DPO), and distillation. New features include global training support, developer-tier hosting, and enhanced evaluation tools such as Auto-Evals and Python Grader. These updates make fine-tuning more accessible, scalable, and cost-effective, enabling teams to build agents that deliver precise, reliable intelligence. Learn more.

• Databricks highlights the shift from generative to agentic AI, marking a new phase in enterprise intelligence. While 65% of organizations have deployed generative AI, 68% plan to invest in agentic systems, AI that can reason, decide, and act autonomously. Early adopters like 3M are using agentic AI to enhance R&D and operational efficiency. Success hinges on unified data, governance, and AI strategies, with leaders prioritizing transparency, automation, and quality. As enterprises move from experimentation to execution, agentic AI is set to redefine scale, trust, and impact across industries. Learn more.

• Databricks outlined best practices for evaluating large language models (LLMs), emphasizing the importance of robust metrics, datasets, and frameworks to ensure performance, safety, and reliability. The blog explores both reference-based and reference-free metrics, including BLEU, ROUGE, perplexity, and toxicity detection. It highlights techniques like LLM-as-a-Judge and human assessments for nuanced evaluation, and introduces Mosaic AI Agent Evaluation for scoring agentic applications across development and production. As LLMs evolve toward multi-agent and tool-using systems, evaluation methods must adapt to assess reasoning, coordination, and ethical alignment. Learn more.

Open-Source Analytics and AI

• Meta AI shared its approach to practical agent security, detailing architectural safeguards for building robust and trustworthy AI agents. The framework includes modular components for perception, memory, planning, and action, each with built-in safety checks. Techniques like adversarial testing, red teaming, and simulation-based evaluation are used to identify vulnerabilities and improve resilience. Meta emphasizes transparency and reproducibility, releasing open-source tools and benchmarks to support the broader research community. These efforts aim to ensure agentic AI systems behave reliably in dynamic, real-world environments. Learn more.

• dbt Labs and Fivetran announced a merger to form a unified company with over $600 million in ARR and more than 10,000 customers. The partnership brings together ingestion and transformation capabilities to deliver open data infrastructure, an integrated, standards-based approach that works across any cloud or compute engine. Both companies will retain their product identities and continue supporting the dbt Community and open source projects like dbt Core and Fusion. The merger aims to simplify data engineering and expand openness, enabling scalable, interoperable analytics and AI solutions. Learn more.

• DuckDB now supports graph queries through the DuckPGQ extension, enabling native analysis of complex networks using SQL/PGQ syntax from the SQL:2023 standard. This allows users to model data as property graphs and perform pattern matching and path-finding directly within DuckDB, no need for external graph databases. The blog demonstrates how to detect suspicious financial activity, such as smurfing and transaction cycles, using visual graph syntax. DuckPGQ simplifies traditionally complex recursive SQL queries, making graph analytics more accessible and performant for real-world use cases. Learn more.

• La Mobilière, a leading Swiss insurer, rebuilt its risk simulation engine using Polars to better model catastrophic events and meet regulatory capital requirements. The new engine runs 5–10x faster than its pandas-based predecessor, enabling actuaries to simulate millions of years of losses across multiple risk types and business lines. Polars’ intuitive API and multi-core performance allow for granular analysis, including tail value at risk (TVaR), directly on actuaries’ laptops. This transformation improves both speed and clarity in risk modeling, helping La Mobilière uphold its financial commitments, even in black swan years. Learn more.

Stay in the Know

Get notified when we post something new by following us on X and LinkedIn.