AI
AG logo
Govern Global · Agentic AI Architecture Series

BuildingIntelligent Enterprises

The GovernGlobal Story — An Architect's Field Guide
🇦🇺 Part of the Australia AI Guidance Navigator: this story shows how the Navigator, architecture principles and agentic AI controls become practical delivery decisions inside a Microsoft-oriented enterprise AI program.
8
Architecture Chapters
9
Expert Perspectives
14
Countries · 1 Mission
∞
Real-world Patterns
📘All product features, frameworks, tools, and technical concepts described in this chapter are the intellectual property of Microsoft Corporation. This is independent educational commentary. Verify all details against current official documentation at learn.microsoft.com before applying to any real project.

CAST The GovernGlobal Team

GovernGlobal is an enterprise professional services and healthcare consulting firm operating across 14 countries. They advise governments, hospitals, and large enterprises on digital transformation. Patient health records, legal contracts, procurement workflows, and citizen services flow through their systems daily — making compliance, security, and governance non-negotiable from day one. These nine people will design every agent, debate every architecture, and own every outcome.

B

Bingo

Head of AI
"Why are we building this — and can we prove its value to the board in 90 days?"
A

Anir

Platform & Governance Architect
"Which platform, which pattern, and what breaks at scale if we choose wrong?"
Ab

Abi

Engagement Manager / PM
"What order do we build in? Who signs off? Which decision is blocking which?"
S

Supra

Identity & Compliance Admin
"Who is the agent acting as? What data is it allowed to see? HIPAA says no."
N

Leon

Vibe Coder
"How do I ship this — and what does it actually cost to run at 10,000 sessions?"
Su

Sud

Developer
"How do I build this exactly, and what are the deployment rules I must follow?"
NJ

NanaJee

API Management & Integration
"What is the agent allowed to call? Is every endpoint registered in my gateway?"
K

Deeno

Power Platform Architect
"Can we solve this in Copilot Studio — and exactly when does it hit the ceiling?"
G

Gogol

ITSM Specialist
"Where is the ITSM record? Who owns it? AI without a ticket is invisible to governance."
✦
How to read this book. Each chapter follows a real design challenge that GovernGlobal faces. The team debates, the architect decides, the principle is explained. Every concept is anchored to a failure mode: what breaks if you get this wrong in production. Architecture is not about choosing the best option — it is about understanding what each option costs when it fails.
00
Overview
The GovernGlobal Story
Fourteen countries. Nine experts. One mandate: build an AI program that works in production — not just in demos. This is what that looks like from day one.
📘All product features, frameworks, tools, and technical concepts described in this chapter are the intellectual property of Microsoft Corporation. This is independent educational commentary. Verify all details against current official documentation at learn.microsoft.com before applying to any real project.

0.1 The Problem Worth Solving

GovernGlobal's daily operations span 14 countries. Their engagement managers manually summarize client meetings into CRM records. Their accounts payable teams match bank statements row by row in spreadsheets. Their procurement analysts review hundreds of contracts each month without AI assistance. Their field engineers arrive on-site at hospitals with no real-time access to the AI-powered guidance that could help them diagnose equipment faults faster.

None of this is unusual. It is the operational condition of most large professional services firms. What makes GovernGlobal different is that they have decided to change it — systematically, with architecture, not experiments.

The question is not whether AI will transform how we work. The question is whether we build the foundation properly before we build the agents on top of it.

— Anir, Platform & Governance Architect

0.2 What an AI Agent Actually Is

The word "agent" is used imprecisely in most organizations. Before GovernGlobal builds anything, Anir draws a definition on the whiteboard that the whole team agrees to use.

What an AI Agent Is (and Is Not)
User natural language Reason Understand intent Plan a response Choose actions Retrieve Query knowledge sources & systems within boundaries Act or Hand Off Take action in connected systems or escalate to human ✗ Not a search engine ✗ Not a chatbot ✗ Not autonomous AI (without design) Every capability operates within designed, governed boundaries — defined by the architect

0.3 The Five Capability Domains

GovernGlobal's AI transformation is organized into five capability domains. Each one is a structural layer — and like floors in a building, you cannot safely occupy the upper floors without the lower ones being architecturally sound. These five domains form the structure of this book.

05 · ALM, Governance & Responsible AI Supra · Anir · Gogol — the architecture nobody sees until it fails 04 · Dynamics 365 AI Integration Abi · Deeno · NanaJee — the right Copilot for the right role 03 · Agent Platform & Design Patterns Deeno · Leon · Sud — designing agents that don't surprise you 02 · Data & Knowledge Architecture Anir · Supra · NanaJee — what the agent knows and who controls it 01 · AI Adoption Strategy Bingo · Abi — building a program that survives contact with reality FOUNDATION
🗺
The building analogy: You cannot skip the foundation and build the roof. Every GovernGlobal AI agent depends on a governed data layer, which depends on a coherent adoption strategy, which depends on clear business case construction. This is not a linear process — but the dependencies are real, and they run downward.

0.4 The Central Tension of the Whole Program

Before the first line of agent configuration is written, GovernGlobal's team has a tension that every AI program has. It is productive. It is never fully resolved. The book is about how the team manages it.

⚡ Leon wants to ship

  • Prototype in three days, deployed in a week
  • Learns from production — iteration beats planning
  • Technical debt is a future problem
  • Real value requires real users
"Give me a weekend and I'll have something running."

🏗 Anir wants it right

  • Architecture decisions compound — wrong now costs 10x later
  • GovernGlobal handles patient data — mistakes have legal consequences
  • Technical debt in AI isn't just code debt — it's compliance debt
  • The second agent is always easier if the first was done properly
"Give me a week to design it, and I'll save three months of rework."

Both are right. The book shows how GovernGlobal resolves this tension — not by choosing one or the other, but by building the governance structure that lets Leon ship fast inside boundaries that Anir trusts.

01
Module One · Bingo & Abi
AI Adoption Strategy:
Building a Program That Survives Reality
Bingo has 90 days to show the board something real. Abi has to sequence 11 proposals into a deliverable plan. Anir is worried they will skip the foundations. This chapter is about why Anir is right.
📘All product features, frameworks, tools, and technical concepts described in this chapter are the intellectual property of Microsoft Corporation. This is independent educational commentary. Verify all details against current official documentation at learn.microsoft.com before applying to any real project.

1.1 The Cloud Adoption Framework — A Sequencing Model

The Cloud Adoption Framework for Azure is Microsoft's prescribed methodology1 for moving workloads and building AI capabilities in the cloud. It is not a checklist — it is a sequencing model. Each phase produces outputs that are the inputs of the next. Skipping phases doesn't eliminate the work; it moves it to production where it costs ten times more to fix.

CAF for AI — GovernGlobal's Four-Phase Journey
Strategy Align leadership Define objectives Bingo owns this Output: AI vision, exec sponsorship Plan Prioritize use cases Score feasibility Abi sequences this Output: prioritized backlog, CoE charter Ready Govern foundation Sensitivity labels Supra MUST sign off Output: governed envs, labeled data, policies Adopt Ship first agent Measure & iterate Leon + Sud build Output: deployed agent, measured ROI
⚡
The cost of skipping Ready: Leon ships an agent in week two that connects to GovernGlobal's SharePoint knowledge base. Three weeks later, Supra discovers the SharePoint library contains unclassified documents including patient health records from a past project. The agent has been summarizing protected health information to users who weren't authorized to see it. This is not a Leon failure — it is an architecture failure. The Ready phase exists to prevent it.

1.2 The Center of Excellence — Three Shapes, Three Stages

A Center of Excellence is not a team. It is a function that changes shape as organizational AI maturity grows. The danger is keeping the same shape past its useful life. A CoE that doesn't evolve becomes the bottleneck it was created to prevent.

CoE Operating Model Evolution — GovernGlobal's Journey
Stage 1 · Centralized Months 1–12 CoE owns all ✓ Standards built ✓ Governance set ⚠ Bottleneck risk ⚠ Shadow AI risk → Stage 2 · Advisory Months 12–24 CoE Standards BU delivers BU delivers CoE consults · BUs build → Stage 3 · Federated Months 24+ Gov Board BU Auto BU Auto BU Auto CoE governs · BUs own fully

GovernGlobal's legal division built a shadow AI tool using a free-tier subscription because the CoE queue had a seven-week wait time. The tool accessed client contract data with no sensitivity labels, no audit trail, and no governance. That is what happens when a centralized CoE stays Stage 1 too long.

— Anir, recounting the incident that forced CoE evolution

1.3 Use Case Prioritization — The Four-Dimension Filter

GovernGlobal receives 11 AI project proposals in the first planning cycle. Bingo wants to approve all of them. Abi draws a 2×2 grid on the whiteboard and asks each team to score their proposal on four dimensions. By the end of the session, three proposals are in the first delivery wave. Eight are deferred — not rejected, deferred.

Use Case Scoring Matrix — GovernGlobal Sprint 1 Selection
Business Value Technical Feasibility Risk Level Org Readiness IT Helpdesk Agent Contract Review AI Multi-Jurisdiction AI HIGH HIGH LOW READY ✓ SPRINT 1 ✓ HIGH MED MED Partially Ready SPRINT 2 HIGH LOW HIGH Not Ready ✗ DEFERRED

1.4 ROAI — Building the Financial Case Properly

Traditional ROI asks: how much did we spend and how much did we save? ROAI asks a harder question: how much did we invest across the full lifecycle — including the costs we haven't paid yet — and what is the realistic range of outcomes given our actual adoption conditions?

ROAI Construction — GovernGlobal Procurement Agent Business Case
① Baseline Measure current state 340 contracts/month 2.4 hrs avg per contract 816 hrs total/month baseline NEVER use benchmarks — measure yours ② Total Costs ALL costs, not just license Development Integration Licensing Training Monitoring Maintenance Future upgrades TCO Calculator → Azure costs3 ③ Three-Band Projection Conservative 25% adoption · Max friction · 204 hrs saved/month Expected 55% adoption · Realistic curve · 449 hrs saved/month Optimistic 80% adoption · Process redesign · 653 hrs saved/month Business Value Toolkit maps each band to business outcomes2

1.5 Who Measures What — The Adoption Measurement Architecture

RoleWhat They MeasureTool / MechanismKey Metric
Bingo Head of AIProgram credibility; executive-level adoption proofCopilot Studio usage estimator; Business Value ToolkitResolved vs abandoned sessions; adoption rate month-over-month
Abi Engagement MgrDelivery sequence; stakeholder sign-off cycleProject tracking; CAF milestone validationDelivery velocity; blocker resolution time
Supra ComplianceSensitive data access events per agent runMicrosoft Purview audit logsSensitive data volume per user per run; anomaly flags
Gogol ITSMAI-attributable incidents; SLA adherenceITSM platform; automated AI incident taggingIncident rate per 1,000 sessions; mean time to resolve
💡
The two cost drivers nobody explains clearly: In consumption-based AI billing, what moves your quarterly cost is not how many users you have or how long sessions last — it is (1) which language model processes each interaction (larger models = significantly more credits per query) and (2) how complex your agent's orchestration is (more chains of model invocations = more credits per user turn). Leon discovers this when GovernGlobal's first agent costs 4x the estimate in month two. The orchestration was more complex than planned.
02
Module Two · Anir, Supra & NanaJee
Data & Knowledge Architecture:
What the Agent Knows and Who Controls It
Deeno wants to connect the agent to everything. Supra has HIPAA obligations and 14 data residency regimes. NanaJee is counting endpoints. Anir is at the whiteboard with the question that decides everything: what does the agent know, and how do we control it?
📘All product features, frameworks, tools, and technical concepts described in this chapter are the intellectual property of Microsoft Corporation. This is independent educational commentary. Verify all details against current official documentation at learn.microsoft.com before applying to any real project.

2.1 Grounding — The Mechanism That Makes AI Trustworthy

Without grounding, an AI agent speaks from its training data — which is vast, general, and impossible to audit. With grounding, the agent speaks from your organization's specific data — which is controlled, attributable, and governable. The grounding architecture is the most consequential design decision in any enterprise AI deployment.

Grounded vs Ungrounded AI — GovernGlobal Compliance Agent
✗ Ungrounded Agent Speaks from training data 🌐 General knowledge from training ⚠ Cannot cite internal policy ⚠ Cannot be audited ⚠ May hallucinate GovernGlobal rules ⚠ Blends general + internal silently ✓ Grounded Agent Speaks from your data Policy Docs Contracts DB Regulations Agent ✓ Every claim is attributable to source ✓ Audit trail exists ✓ Source control = response control ✓ Governance surface is bounded

2.2 Microsoft Dataverse — The Enterprise AI Fabric

Dataverse is not a database with a connector attached4 to it4. It is Microsoft's unified data platform designed from the ground up to serve AI agents, Dynamics 365 apps, Power Apps, and external models simultaneously from a single, governed source of truth. The distinction matters because it changes which design decisions you make.

NeedDataverseAzure Data LakeCosmos DBAzure SQL
Native AI agent connectivityFirst-classNeeds extrasCustom devVia connector
Built-in classification✓ Native AIPRequires Purview add-onCustomExternal tooling
Row-level securityNative — data layerFile-level onlyComplexYes (schema design)
Schema enforcementTyped, validatedSchema-on-readSchemalessFull relational
Best for AI groundingCentral SSOT, multiple AI consumersRaw lake storage at scaleHigh-throughput transactionalRelational transactional

Supra's rule: no AI agent goes near patient data without sensitivity labels applied in Dataverse first. Labels are not metadata — they are enforcement mechanisms. The agent can only retrieve what the label permits the calling user to see. Without them, access is unbounded.

— Supra, after GovernGlobal's first data governance review

2.3 Azure AI Search — When Knowledge Is Distributed

GovernGlobal's knowledge base is not in one system. Policy documents live in SharePoint. Contract data lives in an SQL database. Case management records come from an API. Archived project files are PDFs in blob storage. Azure AI Search solves the heterogeneous data problem by ingesting all source types through configurable indexers and creating one unified semantic index that agents can query.

Azure AI Search — GovernGlobal's Knowledge Unification Architecture
SharePoint Policy Documents SQL Database Contract Records Case Mgmt API Live Case Data Blob Storage Archived PDFs Regulatory Sites Jurisdiction Rules Azure AI Search Configurable Indexers Semantic Ranking Vector Search Incremental Indexing — triggers on source change, not on schedule5 Copilot Studio Agent — unified natural language queries
⚠
Incremental vs scheduled indexing: Standard indexing runs on a schedule — every 6, 12, or 24 hours. Incremental indexing triggers when source data changes. For GovernGlobal's procurement agents, a contract status that changes at 9am must be reflected in agent responses at 9:05am — not at 9pm when the nightly index runs. NanaJee insists every data source feeding the search index has incremental indexing enabled. "Stale data is a compliance risk," he says.

2.4 Data Residency — Governing 14 Countries

GovernGlobal's EU practice stores patient data in Azure Germany.7 Their Southeast Asia practice stores contracts in Azure Singapore. Their AI services run on Azure models that may be deployed in different regions. Cross-region data movement — sending data from one Azure region to another for AI processing — is not assumed. It must be explicitly governed.

Cross-Region Architecture — GovernGlobal EU Healthcare Practice
🇪🇺 EU Environment (Germany) Copilot Studio Agent Patient data — GDPR + HIPAA labeled Dataverse (Germany) Sensitivity labels applied by Supra ✗ Blocked without cross-region enabled 🇺🇸 North America (AI Service) Azure OpenAI Service Language processing & generation Supra's Fix: Enable cross-region data movement for EU environment in PP admin center

2.5 Power BI Semantic Models as AI Sources

GovernGlobal's finance team has spent three years building a certified Power BI semantic model representing client revenue, project margins, and resource allocation — validated by audit, trusted by CFOs across 14 countries. This model should serve as the grounding layer for Bingo's executive briefing agent. But there is a right way and three wrong ways to expose it.

✓

Endorse the Semantic Model6

Registers the model in the AI-ready catalog. Preserves live DirectQuery connectivity. Inherits existing row-level security. The executive agent sees current data with the CFO's access controls intact.

✗

Export to CSV

Creates a static snapshot. Loses live connectivity. Loses row-level security. Executive queries yesterday's data as if it were today's. Compliance risk.

✗

Copy to Dataverse

Creates a raw copy with no DAX measures, no dimensional hierarchies, no business logic. The AI gets numbers without meaning.

✗

Share the Workspace

Makes the workspace visible to additional users. Does not register the model as an AI grounding source. Agent cannot query it.

03
Module Three · Deeno, Leon & Sud
Agent Platform Architecture:
Designing Agents That Don't Surprise You
Deeno's architecture review is the most important meeting in GovernGlobal's AI program. Every bad agent decision made here will cost ten times more to fix in production. This chapter is that meeting, made educational.
📘All product features, frameworks, tools, and technical concepts described in this chapter are the intellectual property of Microsoft Corporation. This is independent educational commentary. Verify all details against current official documentation at learn.microsoft.com before applying to any real project.

3.1 The Copilot Studio Ceiling — Know It Before You Design8

Deeno's first question at every architecture review: "Can Studio solve this?" His second question, which most teams never ask: "What happens when it can't?" Understanding the platform ceiling is not pessimism — it is the design discipline that prevents six-week rework when a requirement turns out to exceed the platform's capability.

Platform Capability Map — What Studio Solves vs Where Custom Is Needed
Copilot Studio Low-code · Visual designer · Governed connectors Conversational Agents Knowledge Retrieval Topic Authoring Live Agent Escalation Multi-Channel Deploy Power Automate Flows Computer Use MCP Servers Voice ✓ GovernGlobal's Helpdesk, Procurement, Patient Intake agents → Studio Custom / Azure AI Foundry Code-first · Full pipeline control · AI engineering Custom ML Pipelines Formal Evaluations Custom Vector Stores Semantic Kernel SDK Multi-jurisdiction Legal AI Fine-tuned Models ✓ GovernGlobal's multi-jurisdiction contract legality analyzer → Foundry

3.2 Language Models — Three Contracts, Not Three Quality Tiers

Selecting a language model is not about picking the "most capable" option. It is about matching the interaction contract to the business requirement. Each model type provides a different guarantee — and that guarantee is the design decision, not a performance consideration.

NLU Model Selection — GovernGlobal's Three Agents, Three Contracts
NLU Standard FULLY DETERMINISTIC Same input → same authored response GovernGlobal use: Patient eligibility agent Regulatory compliance Q&A → Audit-compliant ✓ NLU+ Enhanced VOICE CHANNEL READY Enhanced context reasoning Multi-step conversations GovernGlobal use: Healthcare voice appointment agent D365 Contact Center integration → Voice channel ✓ Generative NLU LLM-powered NATURAL VARIED PHRASING LLM generates from grounding Handles any phrasing naturally GovernGlobal use: Internal policy Q&A (staff) Knowledge base exploration → Internal only per Supra ✓

A generative NLU model at temperature zero is not deterministic. It is "approximately consistent." That is not the same thing. When a HIPAA auditor asks us to reproduce a specific patient eligibility response for a compliance review, "approximately consistent" is not an acceptable answer. We use NLU with authored dialog nodes for eligibility. Full stop.9

— Supra, closing the debate on language model selection for the patient agent

3.3 Orchestration Models — Two Routing Philosophies

Classic Orchestration

"Like a fixed airport gate assignment"
  • Trigger phrases route to specific topics — hardcoded matching
  • Same input always routes to same topic — fully predictable
  • New topics require new trigger phrase sets to be defined
  • Cannot dynamically discover or route to other agents
  • Debug easily: trace which phrase triggered which topic
GovernGlobal use: deterministic compliance response agent where every topic has a fixed authored response

Generative Orchestration

25
"Like a smart routing system that reads your destination"
  • Topics activated by description of their purpose — no triggers needed
  • New agents auto-discovered based on capability description
  • Routes to other agents dynamically — enables multi-agent composition
  • Handles varied phrasing naturally without exhaustive trigger lists
  • Required when: "topics selected based on description, not trigger phrase"
GovernGlobal use: internal procurement agent routing to specialist sub-agents for HR, Finance, Legal, and Contracts

3.4 Topic Architecture — The Three Safety Exits

GovernGlobal Patient Intake Agent — Topic Flow with Safety Architecture
Patient calls in Conversation Start Topic Greeting · Context gathering · Auth NLU+ Intent Recognition Voice channel · D365 Contact Center Fallback Topic Intent unrecognized → Graceful default · Offer alternatives Regular Topics Appointments · Test results · Billing → Handle → Resolve Escalation Topic "I need to speak to a nurse" → Full transcript + context transfer → Live clinical queue Customer Engagement Hub routes to on-call team

3.5 Computer Use10 & MCP — Two Different Keys to Locked Rooms

🖱️

Computer Use in Copilot Studio10

The problem it solves: GovernGlobal's government practice submits documents to a legacy benefits portal with no API. The only interface is a web form.

How it works: The AI model evaluates the page layout at runtime — not at design time. When the government portal updates its UI (which it does quarterly without notice), Computer Use re-identifies elements by context and purpose. The automation survives the update.

Use when: no API exists or UI changes frequently

🔌

Model Context Protocol (MCP)

The problem it solves: GovernGlobal's legal team updates their jurisdiction compliance rules weekly. The AI agent needs these updates — but requiring the AI team to update the agent each time is unsustainable.

How it works: An MCP server exposes the legal team's tools through a standardized protocol.11 At runtime, the agent queries the MCP server to discover what tools are currently available, then invokes them. New tools added by the legal team are immediately available to the agent — no redeployment.

Use when: tool catalog changes or is owned by another team

3.6 Multi-Agent Architecture — Earn the Complexity

Abi proposes a seven-agent architecture for GovernGlobal's government practice. Anir draws it on the whiteboard, then asks: "Which of these agents cannot be a topic in a single agent?" The answer is two. The seven-agent proposal becomes a two-agent architecture. Both requirements are met. The debugging surface is 70% smaller.

SignalSingle AgentMulti-Agent
Security boundaryAll tasks use data the agent can uniformly accessTasks cross security boundaries — patient data vs operational data requires separate permission scopes
Specialization valueGeneral agent handles 80%+ of scenarios at acceptable qualitySpecialized agents demonstrably outperform a generalist for their specific domain
Team maturityTeam is new to AI agent developmentTeam has successfully operated single-agent systems in production
Anir's ruleStart with the minimum agent count that delivers core value. Add agents only when a single agent demonstrably fails to meet a specific requirement.
04
Module Four · Abi, Deeno & NanaJee
Dynamics 365 AI Integration:
The Right Copilot for the Right Role
Every GovernGlobal practice is asking Deeno for an AI agent. Most of what they want already exists — as prebuilt Copilot capabilities in the products they already use. Abi's rule: configure before you build. This chapter maps every major D365 Copilot capability to the GovernGlobal use case it addresses.
📘All product features, frameworks, tools, and technical concepts described in this chapter are the intellectual property of Microsoft Corporation. This is independent educational commentary. Verify all details against current official documentation at learn.microsoft.com before applying to any real project.

4.1 The First Question: Does It Already Exist?

Deeno is designing a Copilot Studio agent with Exchange connectors for email assistance when Abi stops him. "Does this already exist in D365 Sales?" It does — Copilot email assist ships with the product12. Deeno was about to spend three weeks building something Microsoft already built. Abi's rule applies before any build decision is made.

🔑
The architect's first question for every D365 requirement: Is there a prebuilt Copilot capability in this product for this scenario? If yes: configure it and assign the right security role. If no: build a custom agent in Copilot Studio. Building what is already prebuilt adds development time, maintenance burden, and version management overhead — all for a feature Microsoft updates automatically when you use the native version.

4.2 13Copilot for Sales — Three Surfaces, Three Use Cases

GovernGlobal Sales Practice — Copilot Capability Map
Copilot Email Assist LIVES IN: D365 Sales Hub Drafts personalized email replies using open opportunity context + client history GovernGlobal Use Engagement managers draft follow-ups after client meetings M365 Copilot for Sales LIVES IN: Outlook & Teams Email thread summaries Suggested reply drafts Meeting prep with CRM data from D365 Sales surfaced GovernGlobal Use Account leads prep for government client review calls Opportunity Summary LIVES IN: D365 Sales Hub AI-generated summary of open opportunity from configured fields + activity Customization mechanism: Add fields to summary config NOT Power Automate flows GovernGlobal: added "Regulatory Jurisdiction" field

4.3 Field Service — Two Users, Two Copilots

GovernGlobal's healthcare equipment maintenance practice has dispatchers who coordinate field engineers from an office (Outlook + Teams), and field engineers who visit hospitals to service medical equipment on site (Field Service mobile app)14. Same organization. Two different users. Two completely different Copilot integrations.

📧 Copilot in Outlook — For Dispatchers

  • Drafts service assignment communications
  • Summarizes correspondence threads from engineers and clients
  • Prioritizes work orders by urgency and location
  • User: Back-office dispatcher with full screen access

📱 Copilot in Field Service — For Engineers

  • Work-order summaries on mobile app before arriving on site
  • Suggested resolution steps based on fault code
  • Relevant knowledge articles for the specific equipment model
  • User: Field engineer in a hospital — often hands-full, no desk access
⚠
The anti-pattern Deeno almost made: He configured Copilot in Customer Service for the field engineers. Copilot in Customer Service supports back-office case agents working tickets in the Customer Service workspace — it does not extend to mobile field technicians. The deployment worked. The adoption failed. Engineers had no AI assistance on site. Gogol received three complaints in week one.

4.4 Finance Copilot — Three Distinct Capabilities

CapabilityWhat It DoesGovernGlobal Use CasePrerequisite
Account Reconciliation AgentAutomates matching of bank statements to vendor ledger and invoice records. Surfaces overdue invoices and outstanding balances.AP team processes vendor invoices across 14 countries — 3 hours/day of manual statement matching eliminatedFinance and Operations AI security role15 assigned
Collections Coordinator SummaryGenerative AI aggregates payment histories, aging reports, and communication records into a concise narrative per accountCollections managers review 40 accounts/day — summary reduces per-account review from 25 minutes to 4Finance and Operations AI security role assigned
Finance AI Security RoleGrants permissions to access AP data, invoke AI models, write back results for prebuilt Finance Copilot capabilitiesSupra assigns this role in Power Platform admin center before any Finance Copilot user can access the capabilitiesThis IS the prerequisite — must exist first

4.5 Supply Chain Management Copilot

📊

AI Summaries with Copilot

Analyzes customer order data and generates natural language summaries of trends, purchasing patterns, and anomalies for operations managers. GovernGlobal's procurement directors receive a daily briefing paragraph instead of a Power BI dashboard to manually interpret.

🔮

Generative Insights for Demand Planning

AI-powered demand forecasts based on historical sales patterns, seasonal signals, and current order velocity. GovernGlobal's planners build 90-day procurement forecasts with AI augmentation — reducing planning cycle time and improving forecast accuracy.

🗂️

Agent Management16

GovernGlobal's supply chain CTO wants to evaluate which prebuilt AI agents are available before committing to deployment. Agent Management in D365 SCM is the discovery catalog — descriptions, use cases, estimated value, and capability previews. This is where you go before Copilot Studio.

4.6 Contact Center — The Live Agent Hand-Off Architecture

GovernGlobal's citizen services practice runs a contact center for a government client. Citizens call with benefits questions. Most are routine. Some are distressed. The agent must be able to help the first group and safely hand off the second. The hand-off architecture is not optional — it is the safety mechanism.

1

Citizen calls — NLU+ handles voice channel

GovernGlobal's Contact Center agent uses NLU+ — required for voice channel integration with D365 Contact Center.24 Enhanced context awareness ensures multi-turn benefits conversations are handled correctly.

2

Agent handles routine queries from indexed knowledge

Eligibility questions, document requirements, appointment scheduling — all handled by authored topics with deterministic responses. Supra has approved every authored response.

3

Escalation trigger: "I want to speak to someone"

The Escalation Topic detects the intent. It ends the bot session. It invokes the built-in hand-off action — passing the full conversation transcript, citizen case number, query category, and session ID to the routing layer.

4

Customer Engagement Hub routes to correct queue17

The Hub receives the escalation with full context. It routes to the appropriate agent queue based on query category and citizen language preference. The live agent sees the complete transcript before saying hello.

5

Gogol's ITSM record is created automatically

Every escalation generates an ITSM ticket: escalation reason, session ID, citizen identifier (masked), timestamp, queue route. If the bot's escalation rate spikes, Gogol sees it in his dashboard before anyone else knows. That is the operational governance loop closed.

05
Module Five · Supra, Anir & Gogol
ALM, Governance & Responsible AI:
The Architecture Nobody Sees Until It Fails
GovernGlobal is 90 days in. Three agents are live.22 Then Gogol raises a flag: a developer made a production change with no change record. Supra discovers it touched clinical data. This chapter is how GovernGlobal builds the architecture that prevents it permanently.
📘All product features, frameworks, tools, and technical concepts described in this chapter are the intellectual property of Microsoft Corporation. This is independent educational commentary. Verify all details against current official documentation at learn.microsoft.com before applying to any real project.

5.1 Managed vs Unmanaged — The Deployment Contract18

The incident that starts this chapter: a developer found a bug in a live agent, fixed it directly in the production environment because it was deployed as an unmanaged solution18, and did not create a change record. Three days later, Supra's Purview alert fires — the modified component had accessed a Dataverse table containing patient health records. There is no audit trail for the modification. This is what unmanaged solutions in production environments enable.

GovernGlobal ALM Pipeline — Managed Solution Deployment
Development UNMANAGED Sud & Leon build and iterate freely Export as managed solution for promotion Managed export Test / Staging MANAGED IMPORT Anir + Abi validate Cannot be edited here Approval required to proceed Approved managed Production MANAGED — IMMUTABLE Supra + Gogol own it Zero direct edits permitted Every change: back to Dev → Test → here Gogol gets ITSM change record for every deployment

5.2 Microsoft Purview — GovernGlobal's AI Oversight Command Center20

Purview is not one tool for one requirement. For GovernGlobal's AI program, it serves three distinct functions — each owned by a different role, each solving a different governance problem.

Microsoft Purview — Three Functions, Three Owners at GovernGlobal
AI Resource Inventory 🔍 Supra's tool Discovers & catalogs all AI models and agents across the tenant Monthly compliance audit starts here Sensitive Data Tracking 📋 Supra's alert system Per-run, per-user record of which sensitivity-labeled data was accessed & when Anomaly → Gogol incident auto-created Sensitivity Label Admin 🏷️ CIO / Supra's domain Applies AIP labels to Dataverse + SharePoint Labels = enforcement, not just classification Required before agent deployment

5.3 Responsible AI — Six Principles as Architectural Requirements

Microsoft's Responsible AI framework defines six principles21. GovernGlobal's mistake in month three is treating them as a values statement to be completed at the deployment gate. Anir's correction: these are technical requirements with specific architectural implementations. Discovering them post-development creates expensive rework. The Microsoft Cloud Adoption Framework's responsible innovation guidance26 reinforces this design-time approach.

GovernGlobal RAI Assessment — Patient Eligibility Agent Findings
⚖ Fairness FINDING: Violation Non-English interactions: statistically lower quality outcomes vs English. 14-country org must serve all. Fix: Expand multilingual model coverage before scaling to affected regions. 🔍 Transparency FINDING: Violation Eligibility determinations made with no explanation to patient. Outcome given. Basis not given. Fix: Add explanation to every determination: "Assessed based on coverage + procedure code." 👤 Accountability FINDING: Violation Human reviewer approves flagged cases without seeing the AI's flagging rationale. Rubber-stamp. Fix: Present AI reasoning with each flag. Require explicit approval with acknowledgment. 🛡 Reliability & Safety ✓ PASSED Authored dialog nodes for eligibility provide deterministic responses. Guardrails tested. Fallback designed. Anir's NLU architecture decision made this possible. 🔒 Privacy & Security ✓ PASSED Sensitivity labels applied by Supra. Row-level security in Dataverse. Cross-region movement governed. Supra's work in Module 2 protected this pillar. 🌍 Inclusiveness IN PROGRESS Accessibility standards met. Multi-language support in roadmap (tied to Fairness finding above). Multilingual expansion prerequisite to full pass. The Co-Violation: Transparency + Accountability always fail together Any autonomous decision system where the affected person cannot understand the decision (Transparency) and no human can meaningfully review it because they lack the AI's reasoning (Accountability) violates both principles simultaneously. This is the most common failure pattern in enterprise AI.

5.4 Prompt Engineering as Governed Practice

ProblemWhat Goes Wrong in ProductionCorrect Design
Vague instruction
"Help the customer as best you can"
High-variance outputs. Some interactions work. Some don't. Variance is invisible until a high-stakes interaction fails.Clear task-specific instruction: the task, constraints, scope, and format. Explicitly.
No source constraints
General knowledge enabled
Agent blends validated internal policy with training-data general knowledge. Answers sound authoritative. Cannot be attributed. Compliance incident waiting.Restrict to indexed reference sources only. Disable general model knowledge for compliance-sensitive agents.
Identical-phrasing test inputsAgent passes testing but fails when real users phrase questions the way real users do — colloquial, incomplete, ambiguous.Test with varied phrasing: synonyms, informal expressions, partial sentences, different languages. Cover the real input distribution.
Wrong quality metrics
Generation time, word count
Team optimizes for speed and brevity. Whether the AI actually helped is never measured.Primary: response relevance + accuracy. Secondary: consistency. Generation time is performance. Word count is formatting. Neither measures value.

5.5 The Full GovernGlobal Architecture — What Was Built

At the end of this program, GovernGlobal has not just deployed agents. They have built a compounding system — one where the governance infrastructure strengthens every agent added to it, rather than one that accumulates technical and compliance debt with each deployment.

Governance
Purview AI Hub
Sensitivity Labels
RAI Quarterly Assessment
Gogol's ITSM Loop
ALM Pipeline
Dev Environment
Managed Export
Test Approval Gate
Production Immutable
D365 Copilots
Sales Email Assist
Field Service
Finance AI
SCM Summaries
Contact Center
Agent Platform
Copilot Studio
NLU / NLU+ / Gen
Computer Use
MCP Protocol
Escalation Topics
Data Layer
Dataverse SSOT
Azure AI Search
Power BI Endorsed
Cross-Region Governed
Foundation
AI Adoption Strategy
CoE Operating Model
ROAI Framework
Use Case Governance

We did not build a technology project. We built an architecture. Technology projects end when the last agent is deployed. Architectures compound — every agent we add now is safer, faster to build, and easier to govern than the one before it. That is what the foundation work buys.

— Bingo, presenting to GovernGlobal's board at Month 6
✦
What every team member now has: Bingo has proven ROAI with measured baselines. Anir has an architecture that holds under audit. Abi delivered in sequence with every stakeholder signed off. Supra has labeled data, governed access, and a Purview dashboard. Leon ships inside the pipeline instead of around it. Sud has deployment rules and a prompt testing framework. NanaJee has a registered endpoint for every data flow. Deeno knows when Studio solves it and when to escalate. Gogol has an ITSM record for every change and every incident. The architecture is not finished. But it is sound.
06
Applied Tooling · Deeno, Anir & NanaJee
Applied Tooling:
What GovernGlobal Uses, Where, and Why
The first five modules explain adoption, data, agents, Dynamics 365 AI, ALM and Responsible AI. This module turns those principles into scenario-level implementation choices using documented Microsoft capabilities only.
📘
This module maps the same GovernGlobal characters and scenarios to practical implementation patterns using Microsoft-documented capabilities: Copilot Studio, supported knowledge sources, Dataverse, Azure AI Search, Dynamics 365 Copilot experiences, Power Platform DLP, Microsoft Purview, Entra ID and Azure AI Foundry. Validate product availability, licensing and regional support against current official documentation before implementation.

6.1 Anir's Rule: Tool Choice Follows the Scenario

GovernGlobal stops asking "Which AI tool should we use?" Anir changes the question: "What decision does this scenario influence, what data does it touch, what action can it take, and what evidence must exist before production?" The platform decision follows from risk, data, integration and operating requirements — not from excitement about a capability.

Deeno writes six decision rules on the whiteboard. These become GovernGlobal's standing pattern for every new AI request.

1

Check native capability first — Deeno

If Dynamics 365 or Microsoft 365 already provides the required Copilot capability, configure the native feature before building a custom agent. Reduces maintenance, security and ALM overhead.

2

Use Copilot Studio for bounded conversational agents — Sud and Leon

Use Copilot Studio when the requirement is a low-code agent with instructions, topics, supported knowledge sources, tools, triggers, channels and testable conversation flows.

3

Use Dataverse for governed business records — Supra

Use Dataverse where the agent needs structured records, business tables, security roles, auditable data changes and integration with Power Platform or Dynamics 365.

4

Use Azure AI Search for distributed enterprise content — Anir

Use Azure AI Search where the solution needs custom indexing, hybrid or vector retrieval, semantic ranking, enrichment or retrieval pipelines outside Copilot Studio's built-in knowledge-source limits.

5

Use Power Automate, connectors and APIs for actions — NanaJee

Use supported connectors or Power Automate flows for controlled actions. Use API Management or custom APIs where enterprise APIs need authentication, throttling, logging and policy control. Every endpoint registered before any agent can call it.

6

Use Azure AI Foundry for high-control AI engineering — Anir

Use Azure AI Foundry where the scenario needs custom model orchestration, formal evaluation, monitoring, tracing, prompt and model lifecycle management, or deeper engineering control beyond low-code configuration.

✦
Architecture principle: Choose the lowest-complexity platform that satisfies the risk and control requirement. Native before custom. Low-code before custom code. Azure AI Foundry or custom architecture only when evaluation depth, retrieval complexity, integration complexity or risk level requires it.

6.2 Scenario-to-Tool Map

The same GovernGlobal use cases from Modules 1–5 now have implementable platform choices. Each row states the preferred toolset, when to use it, what to avoid, and what evidence must exist before go-live.

ScenarioRecommended ToolingUse WhenAvoid / Important NoteGo-live Evidence
IT Helpdesk Agent
Low risk · Wave 1
Copilot Studio, Teams channel, SharePoint knowledge source, Power Automate → ITSM, Power Platform DLP, Copilot Studio analytics Internal support Q&A, service request triage, KB search, ticket creation, human escalation Do not allow privileged actions — password reset, device isolation, group changes or script execution — without a separate approval path Approved KB source list, fallback topic, escalation topic, ITSM ticket test, DLP policy, usage and unresolved-session metrics
Procurement Contract Review
Medium risk · Wave 2
Copilot Studio, SharePoint or Dataverse knowledge where sufficient, Azure AI Search where custom indexing is required, Power Automate approval, Purview sensitivity labels Contract clause search, policy comparison, supplier metadata lookup, review summary with cited sources Do not present AI output as legal advice. Do not allow ungrounded answers for contractual interpretation. Human review required for high-risk clauses. Data classification, source permission test, citation samples, ungrounded-answer setting reviewed, approval workflow, audit sample
Patient Intake / Eligibility
High risk · Wave 3
Copilot Studio authored topics, validated intake questions, Dataverse records, D365 Contact Center or Customer Service escalation where applicable, Entra ID, Purview audit Administrative intake, appointment routing, collection of non-emergency information, controlled eligibility pre-check where the final decision remains human-owned Microsoft documentation is explicit: Copilot Studio is not a medical device and must not be used for emergencies, medical advice, diagnosis or treatment. Emergency pathway must be designed separately. Clinical or legal review of wording, emergency disclaimer, escalation path, deterministic topic tests, access controls, audit trail, human review path
D365 Sales Assistance
Native first · Wave 2
Copilot in Dynamics 365 Sales and Microsoft 365 Copilot for Sales where licensed and enabled Email assistance, opportunity summaries, CRM-context sales preparation, Outlook and Teams sales workflows Do not rebuild native D365 Sales Copilot in Copilot Studio unless native configuration cannot meet the requirement. Abi's first question: does it already exist? Security-role review, feature availability check, data access validation, adoption metrics, user training
Field Service Guidance
Native first · Wave 2
Dynamics 365 Field Service Copilot capabilities, Field Service mobile app, work-order data, knowledge articles Technician work-order summaries, service history, knowledge articles, resolution assistance Do not solve mobile field-technician workflows with only back-office Copilot experiences. Validate the actual user surface — Copilot in Customer Service does not extend to field mobile workers. Role access, mobile app test, work-order data quality check, knowledge article ownership, field-pilot feedback
Finance Reconciliation
High risk · Wave 3
Dynamics 365 Finance Copilot capabilities where available, Finance and Operations AI security role, Power BI reporting, Purview audit, approval workflow Finance summaries, reconciliation assistance, collections preparation and exception review Do not allow automated financial write-back without segregation of duties and human review. Treat finance AI as decision support unless formally approved otherwise. Security-role assignment, SoD review, audit trail, exception queue, finance owner sign-off, reconciliation test
Contact Centre Escalation
Medium-high risk · Wave 3
Copilot Studio agent, escalation topic, D365 Contact Center or Customer Service hand-off where applicable, transcript and context transfer, service dashboard Routine customer or citizen queries with explicit hand-off to live agents for complex or distressed interactions Do not deploy customer-facing automation without tested escalation, transcript transfer and service-quality monitoring. Escalation tests, hand-off transcript validation, queue routing test, service owner sign-off, abandoned-session metric
Legacy Portal Automation
Medium-high risk
Copilot Studio Computer Use only where an API or connector does not exist, with narrow task scope and human approval for consequential actions UI automation against legacy web or desktop interfaces that cannot be accessed through supported APIs Computer Use is not the first option. Prefer APIs and connectors. UI automation can be brittle and action-impact risk is higher. Strong guardrails required. Allowed screen and action list, test account, rollback path, human approval gate, audit log, exception handling
Multi-Jurisdiction Legal AI
Very high risk · Wave 4
Azure AI Foundry, Azure AI Search, Azure OpenAI where approved, evaluation and monitoring, API Management, legal review workflow Complex retrieval, jurisdiction-specific knowledge, evaluation-driven quality control, monitored reasoning assistance Do not use basic low-code Q&A for legal interpretation. Do not present outputs as final legal advice. Require legal owner review and contestability path. Evaluation results, red-team tests, jurisdiction source register, citations, legal review, risk acceptance, monitoring dashboard

6.3 Who Owns What — Character Responsibilities

The missing link in most AI guidance is ownership. Principles are easy to write and hard to operationalise. GovernGlobal makes ownership explicit: every principle has a character, every character makes a decision, every decision produces evidence.

CharacterTheir DecisionArchitecture Principle They ApplyEvidence They Produce
Bingo — Head of AISelects use cases, owns business value, accepts executive riskAI investment must be outcome-led, measurable and linked to real operational valueROAI baseline, adoption metrics, benefit realisation dashboard, executive risk acceptance
Anir — Platform ArchitectChooses platform pattern, integration model, agent boundary and production architectureEvery AI system must be grounded, identity-aware, bounded, observable and designed to fail safelyArchitecture Decision Record, data-flow diagram, pattern selection rationale
Abi — Engagement ManagerControls sequencing, stakeholder sign-off, delivery dependencies and adoption readinessAI delivery must follow gated adoption, not uncontrolled experimentationIntake register, delivery roadmap, approval log, pilot exit criteria, readiness assessment
Supra — Identity & ComplianceControls identity, sensitivity labels, compliance posture, data access and RAI reviewNo agent may access data unless classification, access control and audit paths are confirmedData classification record, Purview audit evidence, access review, privacy and security assessment
Leon — Vibe CoderPrototypes quickly, validates technical feasibility and proves the user experienceSpeed is acceptable only inside approved guardrailsPrototype scope, sandbox boundary, telemetry sample, cost estimate
Sud — DeveloperBuilds agents, flows, connectors, prompts, topics and deployment assetsPrompting, topic logic and connector actions must be versioned and testableManaged solution export, prompt test cases, deployment notes, rollback plan
NanaJee — API & IntegrationControls APIs, MCP servers, endpoint registration, gateway and action boundariesAgents must use approved integration paths and must not become shadow API clientsEndpoint catalogue, API permission map, MCP tool register, logging evidence
Deeno — Power Platform ArchitectDecides whether Copilot Studio, D365 native Copilot or custom Azure AI is the right patternConfigure before building; low-code before custom code; custom only when requirements exceed platform ceilingPlatform-fit assessment, D365 capability check, Copilot Studio ceiling assessment, exception approval
Gogol — ITSM SpecialistOwns operational visibility, incidents, change records, service health and support readinessEvery production AI system must produce operational signals and support recordsITSM change record, incident taxonomy, support runbook, escalation trend dashboard

6.4 Tool Control Matrix

This matrix is the practical bridge from architecture principle to implementation tool. It uses real control planes available in the Microsoft ecosystem — no invented products.

ToolWhat GovernGlobal Uses It ForKey Implementation NoteOwner
Copilot StudioLow-code agents, topics, instructions, tools, triggers, channels, testing and analyticsUse for bounded conversational agents. Configure knowledge, fallback, escalation and channel deployment explicitly.Deeno + Sud
Supported knowledge sourcesGrounding agents in enterprise contentUse SharePoint, documents, Dataverse, websites or connector-based sources depending on source type and authentication requirements.Anir + Supra
DataverseStructured records, app data and governed business tablesUse security roles, table design, audit settings and environment governance. Do not treat Dataverse as a dumping ground for unclassified content.Supra + NanaJee
Azure AI SearchCustom retrieval over distributed enterprise dataUse for index design, enrichment, vector and hybrid retrieval, semantic ranking and retrieval patterns needing more control than built-in knowledge sources.Anir
Power Platform DLP23Connector governance across Power Apps, Power Automate and Copilot StudioClassify connectors into appropriate policy groups. Prevent unsafe mixing of business and non-business connectors.Supra
Microsoft PurviewSensitivity, audit, compliance and AI governance evidenceUse sensitivity labels, audit review and data governance controls. Do not rely on agent design alone for data governance.Supra
Entra IDAuthentication, user identity and access boundariesValidate whether the agent uses user identity, app identity, delegated access or connector-specific authentication. Review access for sensitive scenarios.Supra
Power AutomateWorkflow actions, approvals and ticket creationUse for repeatable business workflows, human approval and integration steps. Keep high-impact actions behind approvals.Sud + Gogol
API Management / custom APIsGoverned enterprise API exposureUse where agent actions call enterprise services requiring authentication, throttling, policy, logging and lifecycle ownership.NanaJee
MCP serverTool exposure to agents where the toolset is external or changes over timeUse only when MCP is required and supported for the agent pattern. Maintain a tool register, owner and allow-list. See the MCP Governance Framework for the five-pillar governance model: architecture, controls, operating model, assurance and Australian AI alignment.27NanaJee + Anir
Computer UseUI automation where no API or connector existsUse narrowly, with test accounts, explicit allowed actions and human approval for consequential steps. Prefer APIs first.Anir + Gogol
Azure AI FoundryEvaluation, monitoring, tracing, model and prompt lifecycle and high-control AI engineeringUse for complex or high-risk AI systems requiring evaluation pipelines, observability and engineering control beyond low-code agent configuration.Anir
Power BI / FabricAdoption, quality, cost, incident and service reportingUse as reporting layer over operational data, telemetry exports, ITSM data and adoption metrics.Bingo + Gogol
ITSM platformOperational ownership, incidents, changes and support requestsEvery production agent must have an owner, support group, incident category, change process and escalation path.Gogol

6.5 Visual Pattern — IT Helpdesk Agent

This is the safest first production pattern: internal, bounded, measurable and easy to escalate. The agent answers from approved support content and creates ITSM tickets. It does not directly remediate endpoints.

IT Helpdesk Agent — Implementable Pattern (GovernGlobal Wave 1)
Employee Teams · internal portal Copilot Studio Instructions + Topics Fallback topic Escalation topic SharePoint KB Approved runbooks Known errors · FAQs Power Automate Create ITSM ticket Controls DLP policy · No privileged actions Analytics · ITSM audit trail Service Desk (human) Escalation with context · Gogol monitors

6.6 Visual Pattern — Patient Intake

For healthcare-style scenarios, GovernGlobal applies the strictest interpretation: the agent supports administration and routing. It does not diagnose, treat, replace professional judgment or handle emergencies. High-impact decisions stay human-owned.

Patient Intake — Safety-First Pattern (GovernGlobal Wave 3)
Patient Approved channel · disclaimer shown Authored topics Structured intake only No generative eligibility Deterministic routing Emergency → redirect immediately Dataverse record Structured intake state · security roles Human escalation queue Distressed · uncertain · high-risk → clinical team Full context and transcript passed Governance (Supra + Gogol) Privacy and security review before go-live Purview audit · Entra ID · sensitivity labels Deterministic topic test evidence Clinical or legal wording review Emergency disclaimer + escalation path
07
Production Assurance · Abi, Gogol & Supra
Production Assurance:
From Responsible AI Theory to Release Evidence
This module keeps the Responsible AI and Australia AI governance layer but turns it into concrete gates, evidence and operating controls for the GovernGlobal scenarios. Every agent GovernGlobal ships moves through these gates. There are no exceptions.
📘
"Control room" means an operating view assembled from real systems: Copilot Studio analytics, Power Platform admin controls, Purview, Entra, Azure AI Foundry monitoring where applicable, Power BI reporting and ITSM records. GovernGlobal does not invent a separate AI control product.
How the Three Guidance Pages Become GovernGlobal Delivery Decisions
Learn
Australia AI Guidance Navigator
AI Architecture Principles
MCP / Agentic AI Framework
Apply
Governance intent
Architecture design rules
Agent action controls
Decide
Which scenario is safe?
Which tool fits?
What data can be used?
What actions are allowed?
What evidence is required?
Deliver
GovernGlobal applied scenario
ADR
Responsible AI evidence
Production readiness
Reference pageWhat the architect learns thereHow it is applied in GovernGlobal
Australia AI Guidance Navigator Why the AI solution must be governed: ethics principles, AI6-style practices, accountability, transparency, safety and human oversight. Used in Production Assurance to decide what evidence is needed before an agent goes live: data classification, human escalation, contestability, audit and owner sign-off.
AI Architecture Principles How to design the solution: grounded data, identity-aware access, bounded actions, observability, fallback, ALM and production support. Used in each GovernGlobal scenario to choose the right pattern: Copilot Studio, Dataverse, Azure AI Search, D365 Copilot, Foundry, MCP or controlled automation.
MCP / Agentic AI Guidance Framework How to control agentic behaviour: tool registration, action boundary, approval, monitoring, incident response and runtime governance. Used when the agent can call tools, APIs, workflows, MCP servers or legacy UI automation. It defines what the agent is allowed to do and what must be logged.
✓
Simple architect view: learn the three guidance pages first, then use the GovernGlobal story to see the same ideas applied to real delivery choices. The story is the practical layer: which scenario, which tool, which control, which evidence, and which owner.

7.1 Australia AI and Microsoft Responsible AI — The Mapping

GovernGlobal uses Microsoft implementation tools, but their Australian government client requires alignment to Australia's AI Ethics Principles. The Australia AI Navigator makes this straightforward: the local governance language28 and Microsoft's Responsible AI principles name the same requirements in different words. The Navigator's AI6 Practices and the May 2026 agentic AI guidance (ASD ACSC, CISA, NSA, CCCS, NCSC-NZ, NCSC-UK) add the operational and agentic-specific controls on top.

Governance PrincipleWhat It Means in PracticeTool Evidence GovernGlobal ProducesScenario Where It Applies
Privacy protection and security
AU P4 · MS Privacy & Security
The agent only accesses approved data, with clear identity and connector boundariesEntra access model, Dataverse security role, SharePoint permission test, DLP policy, Purview audit sampleProcurement contract review, patient intake, finance reconciliation
Reliability and safety
AU P5 · MS Reliability & Safety
The agent fails safely, uses tested flows and escalates when uncertainTopic test cases, fallback test, escalation test, Azure AI Foundry evaluation where usedPatient intake, contact centre escalation, eligibility pre-check
Transparency and explainability
AU P6 · MS Transparency
The user knows the source, the limitation and the next step. High-impact outputs need an explanation.Citation samples, disclosure text, explanation template, response-quality reviewContract review, legal research, finance summaries
Contestability
AU P7 · MS Accountability
Affected users can challenge or escalate the AI-assisted outcomeHuman review queue, ITSM or case record, escalation SLA, owner assignmentEligibility pre-check, citizen services, finance exceptions
Fairness and inclusiveness
AU P3 · MS Fairness
The pilot covers realistic users, languages, channels and accessibility needsRepresentative test set, pilot feedback, quality comparison by channel and language where availableContact centre, field service, 14-country internal support
Accountability
AU P8 · MS Accountability
Every production agent has an owner, a support model, a change record and a review cadenceNamed owner in ITSM service catalog, architecture decision record, ITSM change record, monthly service reviewAll production agents

7.2 The Five-Gate Release Process

Abi runs the intake. Supra classifies the data. Anir scores the architecture risk. Deeno recommends the platform. Gogol defines the operational support model. Only then does Leon prototype and Sud build. GovernGlobal stops treating AI ideas as tickets to build.

1

Use case intake — Abi

Business owner, intended users, decision impact, expected value, deployment channel, and whether the agent only informs, recommends or acts. Recorded in ITSM before any build begins.

2

Data and identity review — Supra

Data classification, source permissions, authentication model, DLP policy and sensitivity and audit needs. Supra signs. No agent accesses data until this gate is closed.

3

Architecture pattern decision — Anir

Confirms whether the pattern is native D365 Copilot, Copilot Studio, custom retrieval with Azure AI Search, Computer Use, MCP, or Azure AI Foundry. Records the ADR with reasons for selection and rejection. Deeno validates the platform fit.

4

Build and test evidence — Sud and Leon

Topics, fallback, escalation, prompt and response samples, retrieval and citation tests, workflow action tests, evaluation results where applicable. 200-case prompt pipeline from Module 5 runs here.

5

Operate and review — Gogol

Support group, incident category, change process, service dashboard and review cadence confirmed before go-live. ITSM change record created. Rollback plan documented.

7.3 Risk Tiers and Minimum Evidence

GovernGlobal does not apply the same release gate to every agent. Risk determines evidence depth. A helpdesk Q&A agent does not need the same review as a healthcare or legal assistant — but it still needs a support model, approved sources and a fallback path.

Risk TierExample ScenariosMinimum Evidence RequiredAdditional for This Tier
Low
Wave 1
Internal helpdesk Q&A, internal policy FAQ Approved KB source list, fallback topic, escalation topic, DLP policy, ITSM ticket test, support owner named Monthly analytics review
Medium
Wave 2
Procurement contract review, D365 Sales assistance, Field Service guidance Data classification, citation testing, source permission test, role-based access review, approval workflow, audit sample Legal or procurement owner sign-off for high-risk clause categories; adoption training
High
Wave 3
Patient intake, Finance reconciliation, Contact Centre escalation, citizen service triage Human escalation tested, deterministic authored logic for high-impact decisions, privacy and security review, service quality monitoring dashboard active Fairness and inclusion test set, incident playbook, formal owner risk acceptance, Purview audit evidence
Very high
Wave 4
Multi-jurisdiction legal AI, cross-region clinical guidance, automated contract legality analysis All high-risk controls plus: Azure AI Foundry evaluation, red-team testing, legal review, contestability workflow Executive risk acceptance, post-go-live formal review, jurisdiction source register, legal owner named on every output

7.4 Deployment Waves — Same Story, Now With Gates

GovernGlobal does not start with patient eligibility or multi-jurisdiction legal reasoning. It starts with bounded, measurable, recoverable scenarios and only scales after governance muscle is working. Wave order is not negotiable.

GovernGlobal Deployment Waves — Risk-Gated Progression
Wave 1 · Low Risk IT Helpdesk Agent Internal policy Q&A Gates required: KB source approval Fallback + escalation DLP · ITSM ticketing Gogol confirms support Wave 2 · Medium Procurement review D365 Sales · Field Service Gates required: Data classification Citation tests Role-based access review Supra signs ADR Wave 3 · High Risk Patient Intake Contact Centre Finance reconciliation Gates required: Human escalation tested Deterministic authored paths Fairness test set Privacy + security review Architecture board approval Wave 4 · Very High Multi-jurisdiction Legal AI Gates required: Foundry evaluation Red-team tests Legal review Contestability path Executive sign-off

7.5 Architecture Decision Record

Every agent now has a short Architecture Decision Record. It captures not only what was chosen, but what was rejected and why. Anir's rule: if you cannot explain why you did not use the simpler option, you have not justified the complex one.

ADR Template — GovernGlobal AI Agent Decision

FieldWhat to Record
Use caseName of the agent and the business problem it solves
Business ownerNamed accountable person — not a team mailbox
Users and channelsTeams, web, mobile, D365, contact centre or other approved channel
Data sourcesSharePoint, Dataverse, Dynamics 365, Azure AI Search index, API, file source or other approved source
Authentication modelUser identity, app identity, connector auth, delegated access or service principal
Selected platformNative D365 Copilot, Copilot Studio, Azure AI Search, Azure AI Foundry, Computer Use, MCP, or custom integration
Rejected optionsState why a simpler or native option was not sufficient
Actions allowedInform only, recommend, create ticket, submit approval, update record or call API
Actions blockedExplicitly state what the agent must not do
Responsible AI controlsFairness, reliability/safety, privacy/security, inclusiveness, transparency and accountability evidence
Go-live evidenceTest results, approvals, monitoring setup, support model, rollback path and first review date

7.6 Production Operating View

GovernGlobal's production view is not a new product. It is a combined view built from real control planes. Gogol assembles it. Bingo reads it quarterly. Supra checks it monthly for security signals.

Experience
Copilot Studio analytics
D365 feature usage
Channel feedback
Platform
PP admin centre
DLP policies
Environment strategy
Solutions / ALM
Data
SharePoint permissions
Dataverse security roles
Azure AI Search indexes
Power BI / Fabric models
Security
Entra ID
Purview audit
Sensitivity labels
Operations
ITSM incidents
Change records
Support runbooks
Power BI dashboard

7.7 Production Readiness Checklist

The final gate requires evidence, not optimism. If an item cannot be proven in a system record or test result, the agent is not production-ready. Gogol owns this list. Nothing goes live until every item has an answer.

#Gate ItemWhat It Means
1Owner namedBusiness owner, architecture owner, data owner and support owner are documented in ITSM
2Tool decision recordedADR explains why the chosen platform was selected and why simpler or native options were not sufficient
3Data classifiedSources, sensitivity labels, access model, residency and retention are documented and Supra has signed off
4Grounding testedKnowledge sources, citations, retrieval quality and fallback behaviour are tested with evidence attached
5Actions constrainedAllowed and blocked actions are explicit. High-impact actions have approval gates or mandatory human review.
6DLP appliedConnector policy has been reviewed so business data is not mixed with inappropriate connector categories
7RAI evidence completeFairness, reliability, safety, privacy, transparency, inclusiveness and accountability have test or review evidence attached
8Escalation worksFallback, live hand-off, ITSM ticketing or case routing has been tested end-to-end
9Monitoring existsUsage, failures, escalations, unresolved sessions, cost and quality signals have a named owner and a review cadence
10Change path definedDevelopment, test and production deployment plus rollback follow a managed solution or approved change process

A tool choice is not an architecture. The architecture is the tool choice plus the data boundary, the action boundary, the evidence pack, the operating model and the accountable owner. When all five exist, the agent is ready to go live. Not before.

— Anir, final production readiness review · GovernGlobal AI Program Month 9