01 / Orientation

AI Starter

A practical introduction to AI — and the architecture that makes it useful at scale.

◇

Intelligence

Models reason and generate.

Claude · GPT

▤

Context

Relevant knowledge makes responses useful.

Instructions · memory · documents

↻

Execution

Agents, tools and workflows turn reasoning into work.

Copilot · Claude Code · Power Automate

AI provides intelligence. Agents put it to work. Shared meaning allows it to scale.

02 / Fundamentals

Five building blocks

These are related, but they are not interchangeable.

◇

Model

The trained intelligence that produces responses.

Claude Sonnet / Opus · GPT

▣

Harness

The runtime that assembles context and connects tools.

ChatGPT · Claude · Claude Code

◎

Agent

A goal-directed system that chooses and carries out steps.

Coding agent · investigation agent

⇄

Workflow

A defined sequence, trigger, and decision path.

Power Automate · Logic Apps

⌘

Tools & skills

Tools perform actions; skills describe reusable procedures.

MCP · APIs · SKILL.md

A model thinks. A harness enables it. An agent pursues goals. A workflow coordinates work.

03 / Context

Six categories of context

Context is the information made available to the model for a particular task.

☷

Instructions

How to behave; rules and priorities.

System prompts · AGENTS.md

▤

Documentation

Reference material about systems and work.

README.md · schemas

◉

Memory

Useful information retained and recalled across sessions.

Preferences · decisions

☰

History

What has been said and done in a conversation.

Messages · compaction

⌕

Retrieved information

Relevant content fetched when needed.

RAG · search

⌘

Tool results

Outputs of executed actions.

SQL results · API responses

Concrete example: instructions in AGENTS.md; reference docs in docs/; decisions in decisions/; chat history in the platform; retrieved facts and tool results in the current run. Git holds curated, durable artifacts.

04 / Architecture

Scopes and capabilities

Three scopes organize context. Agents, skills, tools, and workflows cut across them.

♧ Organization / PersonShared preferences · standards · institutional knowledge

▣ ProjectShared objectives · project instructions · architecture · decisions

☰ ThreadsConversation history · working state

◎ AgentsGoals · plans · execution state

SkillsReusable procedures

ToolsAPIs · databases · code

WorkflowsTriggers · sequencing · approvals

Scopes define where knowledge applies. Capabilities define how work gets done. This is a conceptual model, not a rigid vendor hierarchy.

05 / Context mapping

Where context lives

The same six categories can appear at different scopes, with different responsibilities.

Category	Organization / Person	Project	Thread / Agent
Instructions	Global guidance	Project rules	Task-specific directions
Documentation	Shared standards	Architecture and definitions	Relevant excerpts
Memory	Long-term preferences	Durable decisions	Working state / recalled facts
History	Cross-project summaries if available	Selected past work if available	Actual interaction history
Retrieval	Shared knowledge sources	Project knowledge sources	Retrieved content in context
Tool results	Shared logs if saved	Saved artifacts	Immediate execution results

Not every platform automatically shares context across these boundaries. Permissions and retrieval determine what is available.

06 / Practical context

How you actually store context

One shared Git repository can hold reviewed project knowledge; each assistant loads or links the pieces it needs.

Context type	Git: shared source of truth	Claude Project / Claude Code	ChatGPT Project / custom GPT	Microsoft Copilot
Instructions	AGENTS.md, project rules	Project instructions; CLAUDE.md for Claude Code	Project instructions; GPT instructions	Agent instructions in Copilot Studio; configured guidance
Documentation	README.md, docs/, schemas/	Project knowledge uploads; repo files in Claude Code	Project files; GPT knowledge files	SharePoint / OneDrive sources; Copilot Studio knowledge
Memory / decisions	decisions/, ADRs, reviewed notes	Platform memory where supported; durable decisions in files	Project context/memory where supported; durable decisions in files	Platform-specific memory where supported; authoritative decisions in sources
Conversation history	Only curated summaries, not raw chats by default	Claude chats / coding sessions	Project chats / GPT conversations	Copilot conversations (subject to product and retention)
Retrieved information	Indexed docs / links / source manifests	Search, connectors, MCP, repo retrieval	Project knowledge, search, connected sources	Microsoft Graph, connectors, grounded knowledge
Tool results	Versioned reports, tests, evidence when useful	Tool calls and outputs in session	Tool/action outputs in conversation	Flow/action results, logs and connected data

Git is the authoritative home for versioned, reviewed artifacts—not a live synchronization mechanism for every assistant's chat history, memory or uploaded knowledge. Platform ingestion and permissions must be configured separately.

06 / Memory

Why memory matters

A trained model does not ordinarily update its weights during your conversation. The surrounding system creates continuity.

▣

Store

Preserve useful information outside the model.

Documents · records · saved memory

⌕

Retrieve

Select relevant knowledge for the current task.

Search · recall

▤

Load into context

Bring selected information into the model’s current request.

Context window

Documents preserve knowledge. Memory systems increasingly automate capture, consolidation, recall, and updating — with authority and review still important.

07 / Ecosystem

Where the product names fit

Think of products as components in the architecture — not as competing synonyms.

Role	Illustrative products	Why you would use them
Model / reasoning	Anthropic Claude models · OpenAI GPT · Azure-hosted models	Reasoning and generation
Assistant / coding harness	Claude.ai · Claude Code · Microsoft 365 Copilot · GitHub Copilot	Interactive assistance and coding
Agent building	Claude Agent SDK · Copilot Studio · Microsoft Foundry	Develop agents and tool-use loops
Workflow automation	Power Automate · Azure Logic Apps · custom orchestration with Claude API/Agent SDK	Triggers, routing, approvals, orchestration
Data & analytics	Microsoft Fabric · OneLake · Power BI · Claude via governed data tools	Data, metrics, and analysis
Integration & security	MCP · Claude connectors · Azure Functions · Entra ID	Tool access, execution, identity, controls

Anthropic supplies models, Claude assistants, coding and agent-development capabilities; Microsoft supplies complementary productivity, data, workflow and cloud services. Enterprise licensing and permissions vary.

08 / Execution

Three execution patterns

Not everything needs an agent — and not every AI workflow needs an open-ended loop.

→

1. Deterministic automation

Trigger → inputs → fixed logic → output.

Example: scheduled validation pipeline

◇

2. AI-enhanced workflow

A defined workflow includes an AI interpretation or classification step.

Example: categorize exceptions

↻

3. Agentic execution

A goal-directed agent chooses next steps based on observed results.

Example: investigate unknown root cause

Use deterministic logic where rules are known; use AI judgment where ambiguity or changing conditions warrant it.

09 / Agent loop

Inside the agent loop

The harness or application must provide the execution loop and its boundaries.

◎ GOAL Outcome + success criteria

1 · PlanChoose next steps

→

2 · ActCall tools / build

→

3 · ObserveInspect / test

4 · Assess: goal achieved?

YES → Finish & reportNO → Replan ↺BLOCKED → Escalate

For coding: design → build → test → review → deploy → verify. Deployment and consequential changes may require human approval.

10 / Adoption

Four levels of AI adoption

A teaching framework for increasing integration and autonomy — not a universal industry standard.

011 · AssistantA person asks AI for help.

022 · AI-enhanced workflowAI performs a defined step in a repeatable process.

033 · Agentic investigationAI chooses and iterates through investigative steps.

044 · Governed autonomyAI monitors and acts within explicit authority, with review where needed.

Higher maturity does not mean replacing deterministic systems with AI. It means using intelligence appropriately within reliable processes.

11 / Example A

Example: data-quality exceptions

The same problem across four adoption levels.

Level	What changes	Possible technology
1 · Assistant	Analyst asks why records failed.	Claude · Copilot
2 · Workflow	Schedule checks; AI classifies exceptions.	Fabric · Power Automate · model API
3 · Agent	Agent traces evidence, diagnoses patterns, proposes fixes.	Claude Agent SDK · Foundry
4 · Governed	Monitor → group proposals → approve → execute → verify.	Agent + queue + deterministic services

The semantic question underneath: do all systems agree on what an “exception,” “record,” or “valid relationship” means?

12 / Example B

Example: monthly performance report

A familiar business process makes the same progression visible.

Level	What changes	Possible technology
1 · Assistant	Summarize spreadsheet results.	Excel Copilot · Claude
2 · Workflow	Produce scheduled metric narratives.	Power BI · Power Automate
3 · Agent	Investigate unusual changes and prepare findings.	Fabric · Copilot Studio · Claude
4 · Governed	Monitor, assign follow-ups, track resolution and escalation.	Agents + workflows + approvals

A shared metric definition prevents different agents from telling different stories about the same business performance.

13 / Governance

The queue as a governance layer

Agent suggestions become reviewable proposals — not immediate changes.

01DetectIdentify candidate issues or actions.

02ClassifyGroup by common rule, evidence, and risk.

03QueueStore proposed changes with provenance and impact.

04AuthorizeApprove, reject, or escalate — individually or in batches.

05Execute & verifyApply through deterministic services and check outcomes.

For batch approval, bind authorization to an immutable proposed change set; recheck if underlying records have changed.

14 / The scaling problem

What happens when many agents work?

Individual intelligence is useful. Shared interpretation is what makes coordinated intelligence reliable.

≠

Different definitions

Two agents calculate “active customer” differently.

Conflicting results

⇄

Different systems

A “provider” in one application means something else in another.

Ambiguous relationships

◎

More autonomy

Inconsistent interpretations can become inconsistent actions.

Greater operational risk

More agents multiply the value of shared meaning — and the cost of inconsistent meaning.

15 / The missing layer

A shared semantic foundation

A common representation helps people, analytics, and agents interpret business reality consistently.

People · Assistants · Agents · Workflows

↕

◎ Semantic Layer / OntologyObjects · relationships · definitions · calculations · rules · provenanceShared, governed meaning

↕

Enterprise Data

Applications

External Systems

An ontology formally models entities and relationships; a semantic layer makes business meaning usable. They overlap but are not identical.

16 / The conclusion

From AI tools to trusted intelligence

Models are increasingly capable. The enduring challenge is helping many systems understand and act on the same reality.

AI provides intelligence.

Agents put it to work.

Shared meaning lets it scale.

A governed semantic foundation can amplify reuse, consistency, and safe automation — potentially dramatically, but not automatically or by a guaranteed factor.

17 / Glossary

Quick-reference glossary

A few terms you will hear repeatedly.

Term	Meaning	Term	Meaning
LLM	Large language model	Inference	Running a trained model
Harness	Runtime surrounding a model	Agent	Goal-directed execution
Context window	Information in one model request	Memory	Persistence + recall across requests
RAG	Retrieval-augmented generation	MCP	Standard for connecting tools/resources
Skill	Reusable agent procedure	Workflow	Defined coordination of steps
Ontology	Formal entities and relationships	Semantic layer	Shared business meaning
Compaction	Summarizing older context	Guardrails	Constraints and checks

Remember: model → context → tools → action; shared semantics help keep the action meaningful.