Automation, analytics, decisions
AI agents integrate enterprise context with generative AI to do three kinds of work — embedded in the tools people already use (Word, Outlook, Teams, Dynamics 365) via Microsoft Copilot:
- Task automation
- Draft documents/emails, summarize threads & meetings, and trigger multi-step workflows (via Copilot Studio, Power Automate).
- Data analytics
- Turn natural-language questions into insights — summarize datasets, surface trends and outliers, generate visualizations.
- Decision support
- Scenario recommendations, risk identification through pattern recognition, and context summaries backed by enterprise knowledge.
Key concept · Assessing measurable value
Agents earn their place when they reduce time on manual tasks, improve accuracy through consistent automation, accelerate insights from complex data, support faster decisions, and enable non-experts to use advanced capabilities. The rule: start with the business outcome you want to improve, use automation to reduce repetitive work (not replace critical thinking), and measure the result.
Five dimensions of trustworthy data
Grounding connects an agent to trusted, access-controlled organizational data so its answers are accurate and hallucinations are reduced. Microsoft Copilot grounds via semantic indexing over Microsoft Graph, and the Copilot Retrieval API pulls passages from SharePoint/OneDrive while honoring user permissions. Grounding data is evaluated on five dimensions:
Key concept · The five data-quality dimensions
- Accuracy — data reflects verifiable facts (validated by SMEs, matched to authoritative sources)
- Relevance — content aligns with the agent's use case and domain, so retrieval isn't “similar but wrong”
- Timeliness — data is current, with refresh schedules; the semantic index updates as content changes
- Cleanliness — structured, deduplicated, minimal noisy metadata → better retrieval precision & embeddings
- Availability — accessible and indexable to authorized users, with clear access controls
Best practice: evaluate content quality before upload, store authoritative content in SharePoint/OneDrive (so it joins the semantic index), keep formatting consistent, review permissions regularly, and validate accuracy with domain SMEs.
An AI-ready data estate
Well-organized data becomes consumable by Copilot, custom agents, RAG apps, analytics, and automation. Microsoft frames it as a layered Azure data estate:
Operational databases
Store app and business data with structure.
Analytical stores (lakehouse / warehouse)
Curate data for AI/ML — Microsoft Fabric as the unified analytics platform.
Intelligence layer (search + vector index)
Azure AI Search, vector indexes, and semantic indexing enable grounding, retrieval, and semantic search.
AI apps & agents
Copilot, custom RAG apps, and autonomous agents consume the prepared data.
Key concept · RAG pipelines & governance
A production RAG pipeline handles every step that separates a prototype from a trustworthy system: ingestion → cleaning → chunking → embedding → indexing → retrieval → prompt assembly → orchestration → monitoring. Data reaches agents through multiple access paths (APIs, search indexes, RAG pipelines, Graph connectors, SQL endpoints) and Azure databases with vector support (Azure SQL, Cosmos DB, PostgreSQL). Govern it early with Microsoft Purview — access policies, sensitivity labels, lineage tracking, and data-quality rules.
Module 2 quiz
Twelve questions on value assessment, grounding, and data readiness. Answer first, then expand Show answer.
Part A · Multiple choice
What does “grounding” do for an AI agent?
- A Speeds up the model's inference
- B Connects it to trusted, access-controlled org data to improve accuracy and reduce hallucinations
- C Compresses the prompt
- D Replaces the language model
Show answer
B. Grounding ties responses to approved, permission-respecting organizational data — in Copilot, via semantic indexing over Microsoft Graph.
Which is NOT one of the five grounding data-quality dimensions?
- A Accuracy
- B Timeliness
- C Profitability
- D Availability
Show answer
C — Profitability. The five are accuracy, relevance, timeliness, cleanliness, and availability.
An agent retrieves content that is topically similar but wrong for the user's actual scenario. Which dimension is failing?
- A Timeliness
- B Relevance
- C Cleanliness
- D Availability
Show answer
B — Relevance. The data must align with the agent's intended use case and domain, preventing “conceptually similar but contextually wrong” retrieval.
Which service provides semantic ranking and retrieval in the intelligence layer of an AI-ready data estate?
- A Azure AI Search
- B Power Automate
- C Microsoft Purview
- D Outlook
Show answer
A — Azure AI Search. With vector indexes and semantic indexing, it enables grounding, retrieval, and semantic search.
Which tool is recommended for governing AI-ready data (access policies, sensitivity labels, lineage)?
- A Microsoft Fabric
- B Microsoft Purview
- C Copilot Studio
- D Azure Cosmos DB
Show answer
B — Microsoft Purview. Implement governance early: access policies, sensitivity labels, lineage tracking, and data-quality rules.
Part B · True or false
AI agents should be used to replace critical thinking wherever possible.
Show answer
False. Best practice is to use automation to reduce repetitive work, not replace critical thinking.
Storing authoritative content in SharePoint or OneDrive helps it become part of the semantic index Copilot grounds from.
Show answer
True. Authoritative, well-formatted content in these locations is indexed for reliable grounding within permission boundaries.
A production RAG pipeline is just a single retrieval call with no additional steps.
Show answer
False. It handles ingestion, cleaning, chunking, embedding, indexing, retrieval, prompt assembly, orchestration, and monitoring — that's what separates a trustworthy system from a prototype.
Data analytics agents let non-experts get insights by asking questions in natural language.
Show answer
True. They convert natural-language questions into summaries, trend/outlier detection, and visualizations without requiring advanced analytics skills.
The Copilot Retrieval API ignores user permissions when fetching passages.
Show answer
False. It retrieves relevant text from SharePoint, OneDrive, and connected sources while honoring user permissions.
Part C · Short answer
Name the three areas where AI agents add value in this module.
Show answer
Task automation, data analytics, and decision-making support.
List the four layers of the Azure data estate for AI-ready data.
Show answer
Operational databases → analytical stores (lakehouse/warehouse) → intelligence layer (search + vector index) → AI apps & agents.