Reference

The data, AI and Microsoft terms that matter.

A working glossary of the technologies, regulations and patterns we ship every day · written for executives, engineers and everyone in between.

Agentic AI: AI systems that plan, call tools, and take actions to achieve goals · with optional human-in-the-loop on critical decisions. The next layer beyond chatbots and copilots.
AI Act: EU regulation establishing harmonised rules on artificial intelligence. In force from August 2024 with phased enforcement: prohibited practices Feb 2025, GPAI obligations Aug 2025, high-risk system obligations Aug 2026, full enforcement Aug 2027.
Anthropic: AI safety company behind the Claude family of large language models · a foundation-model provider used for enterprise assistants, coding and agentic workloads. Claude models are available through Azure AI Foundry and AWS Bedrock.
Apache Kafka: Distributed event-streaming platform where producers publish events and consumers read them · moves millions of messages per second between systems and underpins real-time pipelines and event-driven architectures.
Apache Spark: Open-source engine for large-scale data processing · handles batch, streaming, SQL and machine learning in memory, far faster than legacy Hadoop MapReduce. The compute engine behind Microsoft Fabric and Databricks.
Article 9 (AI Act): Risk management system obligation for high-risk AI · the spine of the entire compliance framework. Requires identifying, analysing and mitigating risks throughout the AI system lifecycle.
Azure AI Foundry: Microsoft's unified platform for building, deploying and managing AI applications · including OpenAI models, Azure ML, fine-tuning, evaluation harnesses and Responsible AI guardrails.
Azure Synapse: Microsoft's previous-generation analytics service (SQL pools, Spark, pipelines). Microsoft Fabric is its SaaS successor; most new mid-market builds start on Fabric.
Bronze / Silver / Gold (medallion): Layered data architecture pattern in lakehouse: Bronze = raw events, Silver = cleansed and conformed, Gold = curated business-ready. Standard pattern in Microsoft Fabric and Databricks.
Capacity (Microsoft Fabric): Compute unit in Microsoft Fabric, sized in F-units (F2, F4, F8 ... F2048). Most mid-market starts with F8 or F16 for the first three months · F32+ is rarely needed for organisations under 500 employees.
Centre of Excellence (CoE): Governance structure for citizen-developer programs · provides environments, DLP policies, ALM pipelines, naming standards, and a certified template gallery. Essential for safely scaling Power Platform adoption.
Change Data Capture (CDC): Integration technique that captures only the row-level changes (inserts, updates, deletes) in a source system and streams them downstream · keeps lakehouses and analytics in near real time without full reloads.
ChatGPT: OpenAI's conversational assistant built on its GPT and o-series models · the most widely used generative-AI product and, for many organisations, the first exposure to LLMs at work.
Citizen developer: Business user who builds applications, automations or reports using low-code tools (typically Power Platform) · with IT-defined guardrails. Absorbs the 80/20 of low-risk work that traditional engineering teams cannot ship fast enough.
Copilot Studio: Microsoft platform for building domain-specific AI agents grounded on enterprise data (SharePoint, Dataverse, Microsoft Fabric, custom APIs). Successor to Power Virtual Agents.
Data catalogue: Searchable inventory of an organisation's data assets with descriptions, ownership, classification and lineage · helps people find, trust and correctly use data. Delivered in the Microsoft stack through Purview.
Data citizenship: Culture in which every employee handles data responsibly, ethically and competently · the behavioural counterpart to data governance, turning policy into everyday practice. Pairs with data literacy.
Data custodian: Technical role responsible for the safe transport, storage and operation of data on behalf of the data owner · typically IT or platform engineers who implement access controls, backups and security.
Data governance: Framework of policies, roles and processes that defines how data is managed, secured and used across an organisation · sets the rules that data owners approve and data stewards enforce.
Data lake: Low-cost store for raw, structured and unstructured data in open formats · flexible and scalable, but needs governance and structure (see lakehouse) to become reliable for analytics.
Data lineage: The traceable path of data from source system to dashboard cell. Makes impact analysis, auditing and trust possible; automated in Microsoft Purview.
Data maturity: Measure of how effectively an organisation turns data into value · spanning strategy, platform, governance, skills and culture. Sparkle's Data Maturity Scan benchmarks where you stand and what to fix first.
Data mesh: Decentralised approach that treats data as a product owned by domain teams, on shared self-service infrastructure with federated governance · an operating-model shift as much as an architecture.
Data owner: Senior business leader accountable for a data domain · approves policy, sets access rules and owns the budget. Accountable for the data, not hands-on with it day to day.
Data pipeline: Automated sequence that moves and transforms data from source to destination · the plumbing behind every report, model and dashboard. Built and scheduled with tools such as Data Factory, Airflow or dbt.
Data quality: Degree to which data is accurate, complete, consistent, timely and fit for purpose · measured, monitored and improved as a discipline, because decisions are only as good as the data behind them.
Data steward: Business practitioner responsible for day-to-day data quality, definitions and compliance within a domain · the guardian who enforces the rules the data owner approves.
Data warehouse: Central repository of structured, modelled data optimised for reporting and analytics · the traditional home of BI, now increasingly delivered as part of a lakehouse.
Databricks: Lakehouse platform built on Apache Spark and Delta Lake for large-scale data engineering, ML and analytics. On Azure it integrates with OneLake and Microsoft Fabric.
dbt: Data build tool. A SQL-first framework for transforming data in the warehouse with version control, testing, documentation and lineage. Runs on Fabric Warehouse via the dbt-fabric adapter.
Delta Lake: Open storage format that adds ACID transactions, schema and time travel to data-lake files · the reliability layer that makes lakehouses possible. OneLake stores data in Delta Parquet.
Direct Lake: Power BI storage mode that queries data directly from OneLake delta tables without import or DirectQuery. Provides sub-second query performance on tables of billions of rows.
DORA: EU Digital Operational Resilience Act · regulation for financial sector entities on ICT risk management, third-party risk, incident reporting and cyber resilience. In force from January 2025.
Embeddings & vector database: Numeric representations of text or images that capture meaning, stored in a vector database for similarity search. The retrieval layer behind RAG and semantic search.
ETL / ELT: Two patterns for moving data · ETL transforms before loading, ELT loads raw then transforms in the target platform. Modern cloud platforms favour ELT because storage is cheap and compute is elastic.
EU AI Act risk tiers: Four-level classification: unacceptable risk (prohibited), high risk (Articles 9-17 obligations), limited risk (transparency obligations Art. 50), minimal risk (no specific obligations). Most enterprise use cases fall in limited or high.
EventStream (Microsoft Fabric): Real-time data ingestion service in Microsoft Fabric · pipes events from Event Hubs, Kafka, IoT Hub or custom sources into KQL databases or lakehouses for sub-minute analytics.
Fabric (Microsoft Fabric): Microsoft's unified SaaS analytics platform combining data engineering, warehousing, real-time intelligence, data science, BI and AI · all built on OneLake with separated storage and compute. SaaS evolution of Synapse.
Feature store: Centralised repository for reusable ML features · ensures consistency between training and serving, reduces feature-engineering duplication across use cases, supports lineage and access control.
Fine-tuning: Further training of a foundation model on your own examples to specialise its behaviour · an alternative or complement to prompting and RAG when you need consistent, domain-specific output.
GDPR-Health: GDPR Article 9 special category data protections applied to health information · processing prohibited by default, with limited exceptions (consent, public interest, healthcare provision). Stricter than general personal data.
Generative AI (GenAI): AI that creates new content · text, code, images, audio or video · from patterns learned in training data. The category that includes LLMs and the tools built on them.
Google Gemini: Google's family of multimodal foundation models and the assistant built on them · a leading alternative to OpenAI, integrated across Google Workspace and Google Cloud.
GPAI (General-Purpose AI): AI models with broad capabilities (GPT-4, Claude, Gemini, Llama, Mistral). Subject to specific AI Act obligations under Articles 51-55, scaled by computational resources used in training.
Hallucination: When a generative model produces confident but false or fabricated output · the core reason enterprise AI needs grounding (RAG), guardrails and human review.
KQL (Kusto Query Language): Query language for time-series and log analytics in Azure Data Explorer and Microsoft Fabric KQL Database · optimised for sub-second queries on billions of events.
Lakehouse: Architecture combining data lake (low-cost object storage, open formats) with warehouse (ACID transactions, schema, performance). Microsoft Fabric, Databricks Delta Lake, Snowflake Iceberg are lakehouse platforms.
LLM (Large Language Model): Large Language Model. A model trained on vast amounts of text to generate and reason over language (GPT, Claude, Gemini, Llama). The engine behind chat, drafting, summarisation and agentic AI.
Machine learning: Field of AI where systems learn patterns from data to make predictions or decisions, rather than following hand-written rules · underpins forecasting, classification, recommendation and anomaly detection.
Master data management (MDM): Discipline of creating one trusted, consistent version of core business entities · customers, products, suppliers · shared across systems. Ends the "which customer list is right?" problem.
MDR (Medical Device Regulation): EU regulation governing medical devices including software · diagnostic AI often dual-regulated under MDR + AI Act high-risk obligations. Requires CE marking, post-market surveillance, technical documentation.
Microsoft Copilot: Microsoft's AI assistant embedded across Microsoft 365, Windows and the Power Platform · grounded on your organisation's data through the Microsoft Graph, with enterprise security and compliance.
Mirroring (Microsoft Fabric): Near real-time replication of an operational database (Azure SQL, Cosmos DB, Snowflake and more) into OneLake at no compute cost, often removing an entire ETL project.
MLOps: Practices and tooling to deploy, monitor and retrain machine-learning models reliably: versioning, CI/CD, drift detection and governance across the model lifecycle.
Model Context Protocol (MCP): Open standard for connecting AI assistants to tools, data and systems through a common interface · the emerging "USB-C for AI", making agents easier to integrate with enterprise software.
NIS2: EU directive on cybersecurity for essential and important entities · expanded scope (energy, transport, banking, health, public administration, digital infrastructure, manufacturing). National transposition deadline October 2024.
OneLake: Single logical lake underlying all Microsoft Fabric workloads · open Delta Parquet format, automatic governance, shortcuts to other storage (Synapse, S3, ADLS, Snowflake) without copying data.
OpenAI: AI research company behind the GPT and o-series models and ChatGPT · closely tied to Microsoft, with its models available through Azure OpenAI in Azure AI Foundry.
Orchestration: Coordinating the order, scheduling, dependencies and monitoring of data pipeline steps · handled by tools such as Azure Data Factory, Apache Airflow or Dagster.
Perplexity: AI-powered answer engine that responds to questions with synthesised, cited answers drawn from live web sources · a conversational alternative to traditional search.
Power BI: Microsoft's analytics and data-visualisation platform · builds interactive reports and dashboards on governed semantic models, with Direct Lake connectivity to Fabric. The most widely adopted enterprise BI tool.
Power Platform: Microsoft low-code suite: Power BI (analytics), Power Apps (custom apps), Power Automate (workflow), Copilot Studio (AI agents), AI Builder (ML on tap), Pages (websites). Foundation for citizen-developer programs.
Prompt engineering: Practice of designing the instructions and context given to a generative model to get reliable, useful output · the cheapest lever for quality before fine-tuning or extra tooling.
Purview (Microsoft Purview): Microsoft's unified data governance platform · classification, sensitivity labels, lineage, data discovery, AI risk register. Replaces Information Protection + Compliance Manager + Data Map under one brand.
PySpark: Python interface to Apache Spark · lets data engineers write large-scale, distributed data processing in Python. The default way to work with Spark in Fabric and Databricks notebooks.
Python: The dominant programming language for data engineering, analytics and AI · readable, versatile and backed by a vast ecosystem of data libraries. The common language across the modern data stack.
RAG (Retrieval-Augmented Generation): Retrieval-Augmented Generation. Grounding an LLM's answers in your own documents by retrieving relevant passages at query time, so responses are accurate, current and citable instead of invented.
Real-Time Intelligence (Fabric): Microsoft Fabric workload for streaming analytics · combines EventStream, KQL Database, Activator (alerting). Sub-minute freshness with native integration into Power BI and AI workloads.
Responsible AI: Practice of designing, building and deploying AI with fairness, reliability, safety, privacy, security, inclusiveness, transparency and accountability built in. Operationalised through evaluation harnesses, content filters, human-in-the-loop, model cards.
Row-Level Security (RLS): Filtering rows in a semantic model or database by who is querying, so each user sees only the data they are permitted to. Core to governed self-service BI.
Self-service BI: Model in which business users build their own reports and analyses on governed, certified data · fast for users, safe for IT, because the semantic model holds the shared definitions.
Semantic model: Logical layer over your data with certified definitions, calculated measures, row-level security, calculation groups, hierarchies. The spine of governed self-service BI · the reports become commodity, the model is the asset.
Shortcut (Microsoft Fabric): Pointer in OneLake to data physically stored elsewhere (Synapse, ADLS, S3, Snowflake) · enables querying without copying. Critical for incremental migration and avoiding data duplication.
Solution Partner (Microsoft): Microsoft Partner Network designation requiring scoring on partner capability score (skills, deployments, customer success, performance). Sparkle holds Solution Partner Data & AI · Azure plus Solutions Partner Power Platform · Modern Work.
Tier-1 / Tier-2 / Tier-3 KPI: KPI hierarchy framework: Tier-1 (GOLD) moves strategy · 5-7 metrics for board and exec; Tier-2 (SILVER) moves operations · 10-15 metrics; Tier-3 (BRONZE) supports specialists · 30+. Prevents the "55 KPIs all at executive level" trap.

Term missing?

We add terms as we build with customers. Suggest one and we'll consider it for the next quarterly update.

Suggest a term