paulasilva Paula Silva | Software Global Black Belt
1 / 1
Microsoft AI Agents · Masterclass

Three platforms, one decision framework for enterprise AI agents.

When to use Copilot Studio, GitHub Copilot Agents, and Azure AI Foundry, with real industry use cases and Human-in-the-Loop guardrails.

AuthorPaula Silva
RoleSoftware Global Black Belt
Contactlinkedin.com/in/paulanunes
Date2026-05-06
Agenda

Eight parts. From definition to deployment.

IWhat is an AI Agent, definition and anatomy.
IIThe three platforms, Copilot Studio, GitHub Copilot Agents, Azure AI Foundry.
IIIAgent types, development, productivity, business, data.
IVHuman-in-the-Loop, when and how to review.
VCase study, vehicle leasing and sales industry.
VIVertical examples, finance, healthcare, retail, manufacturing.
VIIArchitecture best practices.
VIIIImplementation roadmap.
Part I
I

What is an AI agent?

Beyond chatbots: agents perceive, decide, act, and learn.

Definition

An AI agent is software that perceives, decides, acts , with autonomy bounded by policy.

Chatbot (yesterday)
Q&A. Stateless. Reactive.
Receives a question, returns a single answer. No memory across turns. No tool use. No goal pursuit.
Agent (today)
Goal-driven. Stateful. Tool-using.
Receives a goal, plans, calls tools, observes results, iterates until the goal is met or escalates to a human.
Anatomy

Five components every agent needs.

01 · BRAIN LLM GPT-4o, Claude,Llama, Phi.Reasons over input,plans next step. Picked per task. 02 · MEMORY Context + state Conversation history,scratchpad, vectorstore for RAG, episodicmemory across runs. Bounded by policy. 03 · TOOLS Functions + APIs Search, query DB,call REST, run code,trigger workflows.MCP, plugins, skills. Catalogued + scoped. 04 · POLICY Guardrails RBAC, content filters,rate limits, HITLcheckpoints, audit. Non-negotiable. 05 · ORCH. Loop runtime Plan, act, observe,repeat. Multi-agentcoordination, retries,timeouts. The platform’s job. Brain reasons. Memory persists. Tools act on the world. Policy bounds. Orchestration drives the loop.
Part II
II

The three platforms.

Copilot Studio, GitHub Copilot Agents, Azure AI Foundry, when to use each.

Microsoft · AI agent platforms

Three platforms, three audiences, one Microsoft AI stack.

01 · Copilot Studio
Low-code business agents
For business users and citizen developers. Drag-and-drop topics, connectors to Microsoft 365, Dataverse, Power Platform. Best for HR, IT helpdesk, customer service bots.
02 · GitHub Copilot Agents
Developer agents in the IDE
For software engineers. Agent Mode, custom agents in repos, MCP tools, extensions. Best for code generation, refactoring, PR automation, tests.
03 · Azure AI Foundry
Pro-code custom agents
For ML engineers and platform teams. Full SDK, model catalog, multi-agent orchestration, fine-tuning, observability. Best for production-grade vertical agents.
Decision matrix

Pick the platform that matches the audience and the control surface.

CRITERIA COPILOT STUDIO COPILOT AGENTS (GH) AI FOUNDRY Audience Business / citizen dev Software engineers ML / platform teams Skill level Low-code, no-code Code-first, IDE-native Pro-code, SDK Best for HR, IT helpdesk, CS DevEx, code, PRs Vertical, multi-agent Connectors M365, Dataverse, PP Repos, MCP, GitHub API Any (Python, REST, MCP) Time-to-pilot Hours to days Days Weeks Customization Topics + connectors Custom agents in repo Full code + fine-tune Governance Power Platform admin GitHub policies Azure RBAC + Content Safety The three are complementary, not competitive. Most enterprises run all three; the question is which use case lands where.
Part III
III

Agent types.

Where each platform shines: development, productivity, business, data.

Four agent types · Where they fit

Match the agent type to the right platform.

01 · Development
Bug-fix, refactor, test gen
Lives in the repo and IDE. GitHub Copilot Agents is the home. Custom agents in .github/agents/, MCP tools for build and ticket integration.
02 · Productivity
Meeting notes, drafts, search
Inside Microsoft 365. Copilot Studio for custom topics; Microsoft 365 Copilot Agents for embedded experiences in Teams, Outlook, SharePoint.
03 · Business + processes
Onboarding, approvals, claims
Cross-system orchestration. Copilot Studio for the conversational layer; AI Foundry when SLAs, multi-agent, or fine-tuning are required.
04 · Data + analytics
NL-to-SQL, insight summaries
AI Foundry with grounding on Fabric, Synapse, or your warehouse. Strict prompt engineering, evaluation harness, and human approval before publish.
Part IV
IV

Human-in-the-Loop.

Autonomy is not the goal. Trustworthy outcomes are.

HITL · Principle

More autonomy requires more guardrails, not fewer.

An agent that can act on regulated data, sign documents, move money, or change customer records must have explicit human checkpoints. The cost of one wrong autonomous action exceeds the savings of one hundred correct ones.

HITL · Decision matrix

Map every action to a quadrant before deploying.

IMPACT → REVERSIBILITY ↑ Q1 · AUTOMATE Low impact, high reversibility Suggest commit messages, draft email replies,summarize meetings, search docs. Run in agent loop. Audit, no checkpoint. Q2 · REVIEW High impact, high reversibility Open PR, schedule meeting, send draft to client,propose code change for review. Human approves before commit / send. Q3 · CONFIRM Low impact, low reversibility Delete temp file, archive ticket,close conversation. Single confirmation step. Q4 · DUAL CONTROL High impact, low reversibility Sign contract, transfer funds, change RBAC,delete production data, file taxes. Two humans approve. Audit + signed evidence.
HITL · Implementation patterns

Three patterns cover most cases.

Pattern 01 · Approve-before
Agent proposes, human approves.
Agent prepares the action and waits at a checkpoint. Approval logged with reviewer identity. Use for PRs, drafts, recommendations.
Pattern 02 · Sample-and-review
Agent acts, human spot-checks.
High-volume, low-individual-impact actions. Random sample audited. KPIs trigger escalation. Use for tagging, classification, routing.
Pattern 03 · Dual-control
Two humans, one action.
Irreversible or regulated. Initiator + approver, never the same person. Cryptographic signature on both. Use for finance, legal, prod data.
Part V · Case study
V

Vehicle leasing & sales.

Two real problems: traffic-fine driver indication, and tax compliance for NF-e.

Problem 01 · Traffic fines

Driver indication has a 15-day deadline. Miss it, the leasing company pays.

Manual today
Mailroom → spreadsheet → chase customer.
Notifications arrive by mail, scanned by hand, transcribed into Excel. Operator emails the customer who leased the car. Some replies, many do not. Deadlines slip; the company pays + penalty.
Agent-led
Ingest → extract → route → track.
Document Intelligence agent extracts plate, date, code, value, deadline. Lookup agent finds the lessee in SAP/Salesforce. Notification agent sends portal link. Tracking agent escalates on day 7.
Architecture · Multi-agent flow

Four agents in choreography on Azure AI Foundry.

SOURCES Mail, email, API Scanned PDFs,SEFAZ feed,DETRAN portal. Ingestion endpoint. AGENT 01 Ingestion DocumentIntelligenceextracts fields. Plate, date, value. AGENT 02 Lookup Cross-referencesSAP / Salesforcefor active lease. Customer + contract. AGENT 03 Notification Sends portal link,SMS, email.HITL on template. Reviewed by ops. AGENT 04 Tracking Day 7 escalation,day 12 final ping,day 14 paralegal. Closes the loop.
Problem 02 · Tax compliance

NF-e errors trigger SEFAZ rejections + ICMS-ST risk.

Every contract closure, lease termination, and inter-branch transfer must produce a valid NF-e or NFS-e. Tax rules vary by state (RICMS), municipality, and product category (NCM). Manual issuance breaks at scale; one wrong CFOP code blocks invoices for the whole month.

Architecture · NF-e compliance flow

Three agents from ERP event to SEFAZ-validated invoice.

TRIGGER ERP event SAP / Oracle / TOTVS:contract closed,vehicle returned,inter-branch transfer. AGENT 01 Tax classifier (RAG) Azure AI Search +RICMS legislation+ NCM table.Picks CFOP, CST, ICMS. AGENT 02 NF-e issuer Calls SEFAZ NF-e 4.0API. Handles retries,contingency mode,cancellation flow. AGENT 03 Bookkeeping Daily run: SPED,EFD-Contribuições,in/out ledgers,ICMS-ST report. HITL CFO sign-off Monthly close.Dual-control ontax filing.
Part VI
VI

Vertical examples.

Same pattern, different domain.

Vertical · Same pattern, different domain

Where multi-agent flows pay off in four industries.

Finance
KYC + AML triage
Document agent reads onboarding files, risk-scoring agent matches sanctions lists, escalation agent routes to compliance officer with dual-control on approval.
Healthcare
Prior authorization
Clinical-extraction agent + payer-rules RAG + HITL with physician before submission. Cuts authorization time from days to hours, all decisions logged for audit.
Retail
Returns & reverse logistics
Triage agent classifies return reason, refund agent runs policy check, logistics agent books pickup. HITL only on exceptions over the SKU threshold.
Manufacturing
Predictive maintenance
Telemetry agent watches IoT signals, diagnostics agent proposes work orders, scheduling agent slots technicians. Plant manager approves work orders over 4 hours.
Part VII
VII

Architecture best practices.

Six layers that turn pilots into platforms.

Architecture · Six layers

Build agents on top of platform primitives, not around them.

06 · SURFACE Teams, Outlook, IDE, custom apps — where users meet the agent. 05 · ORCHESTRATION Multi-agent runtime, planning, retries, HITL checkpoints, evaluation. 04 · TOOLS · MCP Catalogued, versioned, scoped tools. MCP servers + native plugins + custom skills. 03 · CONTEXT · RAG Vector store, semantic context layer, recipe registry, identity-scoped retrieval. 02 · MODELS Foundry catalog. GPT-4o, Claude, Llama, Phi. Picked per task. Fine-tuned where it pays. 01 · PLATFORM · GOVERNANCE Identity, RBAC, Content Safety, audit, observability, FinOps. Same as the rest of Azure.
Architecture · Five principles

Five rules every Microsoft AI agent program should adopt.

01Start narrow. One use case, one agent, one team. Generalize after value is proven.
02Reuse Azure governance. Identity, RBAC, Content Safety, Purview — same controls as the rest of the stack.
03Catalogue every tool. MCP server registry, scoped tokens, audit per call. No agent talks to a random API.
04Evaluate before deploy. Golden test set per agent, regression on every model swap, latency budget per loop.
05Plan the deprecation. Version every prompt, log every change, keep the rollback — agents drift like models do.
Part VIII
VIII

Implementation roadmap.

From first pilot to production-grade fleet.

Roadmap · Five phases over 12 months

Phase by phase, with measurable exit criteria.

01 Discover Month 1 Use-case mapValue vs effortStakeholder signoff 02 Pilot Months 2-3 One agentOne teamClosed-loop eval 03 Harden Months 4-6 RBAC + auditSLA + cost capHITL checkpoints 04 Scale Months 7-9 Tool registryMulti-agentSelf-service onboarding 05 Operate Months 10-12+ FinOps + driftModel swap drillsQuarterly review Each phase has explicit exit criteria. Skip none. Pilot without harden is a demo; scale without operate is debt.
Closing thought
paulasilva

Three platforms. One Microsoft AI fabric.

Building the future of software development with AI and Agentic DevOps.

Paula Silva | Software Global Black Belt
linkedin.com/in/paulanunes
Agentic DevOps Hub↗
PDF
ENESPT