Explore Responsible AI

Principles, Risks, and Best Practices for Ethical AI Implementation

What Is Responsible AI?

The Definition
AI systems designed and deployed with ethical principles, accountability, and
human well-being at the core

Why It Matters
Technology impacts billions—responsible development ensures fairness,
safety, and trust in AI systems

Misinformation Algorithmic Bias Privacy Threats

Risk
Systems perpetuate Unauthorized data
AI-generated false discrimination in hiring, collection and
content spreads rapidly lending, and healthcare surveillance concerns
online, eroding public decisions threaten individual
trust rights
Six Core Principles of Responsible AI
Fairness Reliability & Safety
AI systems treat all people equitably without Systems perform consistently and safely across diverse
discrimination based on protected characteristics scenarios and edge cases

Privacy & Security Inclusiveness

Data protection and user privacy maintained AI benefits everyone, designed for diverse needs and
throughout the AI lifecycle abilities across populations

Transparency Accountability
Clear communication about how AI systems work and Humans remain responsible for AI system outcomes
make decisions and impacts
Common Risks & Ethical Concerns

Deepfakes & Synthetic Media

Realistic fake videos and audio manipulate public perception, erode trust in authentic
content

Copyright & Misinformation

AI-generated content raises ownership questions, spreads false information at
unprecedented scale

Harmful Content Generation

Systems can produce offensive, dangerous, or inappropriate material without proper
safeguards

Training Data Bias

Historical prejudices embedded in datasets lead to discriminatory AI decisions and
outcomes
Best Practices for Responsible AI
01 02 03

Verify AI Outputs Implement Human Oversight Protect User Privacy

Always fact-check generated content—AI Maintain human review for critical Minimize data collection, anonymize
can hallucinate or produce inaccurate decisions—AI assists but shouldn't information, ensure secure storage and
information replace judgment transmission

04 05

Evaluate for Bias Document AI Usage

Test datasets and outputs across demographics—identify and Keep clear records of AI applications, decisions made, and
correct discriminatory patterns regular system performance reviews
Human-AI Interaction & Global Impact

Privacy Concerns

• Surveillance capabilities expanding worldwide

• Data consent often unclear or misleading
• Cross-border data flows raise jurisdictional questions

Cultural Fairness

• Western bias in training data sets

• Language representation gaps persist
• Cultural context misunderstandings common
Real-World Examples

AI Misuse: Biased Hiring Tool ✓ Responsible AI Success: Healthcare Diagnostics

Case: Major tech company's recruiting AI systematically downgraded Case: Medical imaging AI developed with diverse patient data, human
female candidates—trained on historical data reflecting male-dominated radiologist oversight, transparent decision explanations
workforce
Impact: Improved diagnosis accuracy across all demographics,
Impact: Perpetuated gender discrimination, required complete system maintained doctor-patient trust
overhaul

Reflection Questions

1. What safeguards could have prevented the hiring tool bias?

2. Why is human oversight critical even when AI improves accuracy?
Key Takeaways
AI must be designed with ethical principles from the
start
Fairness, transparency, and accountability aren't optional features

Human oversight remains essential for responsible

deployment
AI assists decision-making but cannot replace human judgment

Bias and privacy risks require constant vigilance

Regular testing, diverse datasets, and clear consent processes

Global impact demands cultural awareness and

regulatory compliance
What works in one region may harm another—design inclusively
Your Responsible AI Checklist

Before Using AI During AI Use

• Understand system limitations • Maintain human review

• Verify training data sources • Document decisions
• Check for known biases • Monitor for errors
• Review privacy policies • Protect sensitive data
Building Trust Through
Responsibility
Responsible AI isn't just about preventing harm—it's about
creating technology that empowers everyone fairly, safely,
and transparently.

Remember Act
Every AI interaction is an Apply these principles in your
opportunity to demonstrate projects and organizations
ethical leadership today

Advocate
Champion responsible AI practices in your communities
Microsoft AI for Business Leaders

Embrace Responsible AI
Principles and Practices
© Copyright Microsoft Corporation. All rights reserved.

May 2024
Agenda

Prepare for the implications of responsible AI Principles of responsible AI

Establish a system for AI governance Take actions for AI governance

AI governance engagement Responsible AI at Microsoft

© Copyright Microsoft Corporation. All rights reserved.

Prepare for the Implications of Responsible AI

Societal Implications of AI
The use of AI technology has broad impacts on society, raising complex and
challenging questions about the future we want to see

The Importance of a Responsible Approach to AI

With benefits come unintended and unforeseen consequences:

• Threats: Novel threats require a constant evolution in our approach to

responsible AI

• Biased outcomes: AI may reinforce societal or other biases without

deliberate planning and design

• Sensitive use cases: Another illustration of our responsibility to mitigate

unintended consequences is with sensitive technologies like facial
recognition

© Copyright Microsoft Corporation. All rights reserved.

Principles of Responsible AI

Challenge or Risk Description Example

Fairness Bias can affect results A loan-approval model discriminates by gender due to bias in the
data with which it was trained

Reliability & safety AI may cause inadvertent harm A generative AI model may produce output that includes
inaccurate or dangerous advice

Privacy & security Private data could be exposed A medical diagnostic bot is trained using sensitive patient data,
which is stored insecurely or revealed in AI output

Inclusiveness Solutions may not work for everyone A generative language model may not reflect perspectives from
some segments of society

Transparency Users must trust a complex system An AI-based financial tool makes investment recommendations –
what are they based on?

Accountability Who's liable for AI-driven decisions? An innocent person is incorrectly convicted of a crime based on
evidence from facial recognition – who's responsible?

© Copyright Microsoft Corporation. All rights reserved.

Reflect and Discuss

What other examples

can you think of for
each of the
responsible AI
principles?
© Copyright Microsoft Corporation. All rights reserved.
Establish a System for AI Governance

Chief Ethics Officer Ethics Office

Centralized decision-making Dedicated ethics team from
enables organizations to quickly different levels of the
develop policies around ethics organization that is solely
while ensuring there's focused on ensuring the ethical
accountability for each decision principles are being followed by
all employees

Ethics Committee
Provides perspectives from people with a wide range of diverse
backgrounds and expertise, unbiased opinions from external
members, and buy-in from senior leaders across the company

© Copyright Microsoft Corporation. All rights reserved.

Take Actions for AI Governance
No matter which AI governance approach you take…

Make Resources Available

A handbook, a manual, or a training session can enable employees to understand and follow
responsible AI principles

Create a Centralized AI Inventory

Document all AI models and systems to prioritize efforts and support compliance audits

Develop Tools
Build tools to automate system monitoring and reporting for compliance issues

© Copyright Microsoft Corporation. All rights reserved.

AI Governance Engagement

Engagement with AI Systems Developers Engagement with External Stakeholders

First-party AI systems Collaborate with other organizations to ensure best practices while maximizing the

Your ethical governance system should: potential of AI to deliver broad benefits:

• Review and provide advice on AI system development • External enterprises

• Encourage employees to surface ethical concerns • Public organizations

• Provide risk mitigation guidance • Governments

• Create processes to monitor the AI systems • Nonprofits

Your developers should: Consider participating in industry partnerships and initiatives:

• Be given detailed and thorough AI development guidance • Microsoft is a member of the Partnership on AI (PAI), a group of researchers,
nonprofits, nongovernmental organizations (NGOs), and companies dedicated to
• Use guidelines and checklists for specific AI technologies, such as face
ensuring that AI is developed and utilized in a responsible manner
recognition or generative AI

Third-party AI systems

Your ethical governance system should:

• Learn about the third party's commitment to responsible AI design to ensure it

aligns with your own principles

© Copyright Microsoft Corporation. All rights reserved.

Responsible AI at Microsoft

Centralized Governance Decentralized Governance

Senior Leadership Team Responsible AI Champs

Sets the company's AI principles, values, and human rights commitments Employees nominated by their leadership teams from within key engineering
and field teams to advise on responsible AI

Office of Responsible AI
Every Employee
Implements and maintains our commitment to responsible AI governance by
working with stakeholders across the company We expect every Microsoft employee to:

• Develop a general understanding of our AI principles

Aether Committee • Report and escalate sensitive uses

• Contact their Responsible AI Champ when they need guidance on

(AI, Ethics, and Effects in Engineering and Research)
responsible AI

Advises the senior leadership, the Office of Responsible AI, and other teams
across the company

© Copyright Microsoft Corporation. All rights reserved.

Reflect and Discuss

What challenges can you identify for

establishing a system of responsible AI
governance in your organization?
© Copyright Microsoft Corporation. All rights reserved.
Knowledge Check

1 2 3

Which of the following How should you establish a Which action should your
capabilities of Azure Machine responsible AI governance responsible AI governance
Learning can help mitigate bias system? leadership take?
issues related to the principle of
• Appoint a chief ethics officer with • Create and share assets to help
fairness?
sole responsibility for devising and employees understand and adopt
• Models can be trained using data that enforcing responsible AI principles responsible AI principles
is stored securely in cloud data • Allow employees to follow their own • Start by removing all access to IT
stores instincts and trust them to make systems and re-allow access only to
• Models can be analyzed to quantify ethical decisions employees as they are individually
the impact of individual data features • Take an approach that combines approved by the chief ethics officer
on predictions centralized and decentralized • Replace all employee smartphones
• Models can be deployed as governance and works best for your with secure laptop devices that are
containerized services in Microsoft organization capable of running the latest AI
Azure models
Summary

Establishing a system for responsible AI governance is an essential element of successfully adopting AI in an organization. Industry
standards, government regulations, and an organization's own ethical principles all help determine an appropriate responsible AI approach.

Now that you have completed this module, you should be able to:

• Prepare for the implications of responsible AI

• Describe principles of responsible AI
• Establish a system for AI governance
• Take actions for AI governance
• Engage across teams and organizations to implement responsible AI principles
• Take inspiration from how responsible AI is approached at Microsoft

To learn more, see [Link]

© Copyright Microsoft Corporation. All rights reserved.

Copilot in Project Operations

Boost
Productivity with
AI-Powered
Assistance
Task plans • Risk assessments • Status reports — all powered by
intelligent automation
What is Copilot for Project Operations?

AI-powered assistant integrated directly into Dynamics 365 Project

Operations

Simplifies planning, accelerates risk identification, and streamlines

reporting through conversational AI

• Natural language task generation

• Intelligent risk assessment
• Automated status reporting
• Enhanced decision-making support
Enabling Copilot
Prerequisites & Setup Requirements

Before You Begin

• Working knowledge of Project Operations
• Valid Copilot licence assignment
• Administrator access rights
• Data region consent (if required)
Task Plans & Risk Assessments
Accelerate Project Initiation

Task Plan Generation Risk Assessment

AI analyses project name and Identifies potential risks across

description to suggest scope, schedule, and budget
comprehensive work breakdown dimensions
structure
Provides mitigation strategies
Fully editable and customisable to based on project metadata and
match your methodology historical patterns

Time Savings: Generate initial drafts in minutes rather than

hours, allowing project managers to focus on strategic oversight
and refinement
Status Reports & Chat Experience

Automated Status Reporting

Copilot aggregates project data — work progress, financials, tasks, effort tracking — into
comprehensive reports

• Internal and external report formats

• Charts and graphs automatically generated
• Editable summaries for context refinement

"Generate a task plan for website redesign "What are the top risks for Project Alpha?" "Create a status report for the executive
project" team"
Important Limitations
What Copilot Doesn't Do Automatically

Manual Refinement Required Context Verification Essential Compliance & Governance

AI-generated tasks lack automatic resource Risk assessments and status reports may miss Ensure proper data governance policies are in
assignments, dependencies, and detailed nuanced project context place
checklists
Validate AI output before sharing with Verify tenant compliance with data-region
Always review and adjust to match your stakeholders or clients settings and organisational requirements
project's specific needs

Treat Copilot as an intelligent assistant, not a replacement for project management expertise and judgement
When to Use Copilot
Optimal Scenarios & Implementation Guidance

Project Initiation Periodic Risk Reviews

Generate quick work breakdown structures for new projects Regular risk identification and mitigation planning

Frequent Reporting Fast Project Updates

Weekly or monthly status updates for stakeholders Quick adjustments to plans and rapid communication

Best Practice Approach

• Use AI outputs as starting drafts, not final deliverables

• Involve team members in reviewing generated plans and risks
• Combine AI efficiency with human expertise and contextual knowledge
• View Copilot as a productivity amplifier, enhancing — not replacing — PM judgement
Key Takeaways & Next Steps
Copilot accelerates project setup and reporting
Task plans, risk assessments, and status reports generated in minutes

Human oversight remains essential

Always review, refine, and validate AI-generated outputs

Natural language makes Project Operations more accessible

Chat-style interface reduces learning curve for new users

Recommended Next Steps

1. Enable Copilot feature in your environment

2. Train team members on responsible AI usage
3. Pilot on a small, low-risk project
4. Establish review processes for AI-generated content
5. Review AI outputs carefully and document adjustments needed
6. Scale adoption based on pilot results
Copilot in Project Operations

Boost
Productivity with
AI-Powered
Assistance
Task plans • Risk assessments • Status reports — all powered by
intelligent automation
What is Copilot for Project Operations?

AI-powered assistant integrated directly into Dynamics 365 Project

Operations

Simplifies planning, accelerates risk identification, and streamlines

reporting through conversational AI

• Natural language task generation

• Intelligent risk assessment
• Automated status reporting
• Enhanced decision-making support
Enabling Copilot
Prerequisites & Setup Requirements

Before You Begin

• Working knowledge of Project Operations
• Valid Copilot licence assignment
• Administrator access rights
• Data region consent (if required)
Task Plans & Risk Assessments
Accelerate Project Initiation

Task Plan Generation Risk Assessment

AI analyses project name and Identifies potential risks across

description to suggest scope, schedule, and budget
comprehensive work breakdown dimensions
structure
Provides mitigation strategies
Fully editable and customisable to based on project metadata and
match your methodology historical patterns

Time Savings: Generate initial drafts in minutes rather than

hours, allowing project managers to focus on strategic oversight
and refinement
Status Reports & Chat Experience

Automated Status Reporting

Copilot aggregates project data — work progress, financials, tasks, effort tracking — into
comprehensive reports

• Internal and external report formats

• Charts and graphs automatically generated
• Editable summaries for context refinement

"Generate a task plan for website redesign "What are the top risks for Project Alpha?" "Create a status report for the executive
project" team"
Important Limitations
What Copilot Doesn't Do Automatically

Manual Refinement Required Context Verification Essential

AI-generated tasks lack automatic Risk assessments and status reports may
resource assignments, dependencies, and miss nuanced project context
detailed checklists
Validate AI output before sharing with
Always review and adjust to match your stakeholders or clients
project's specific needs

Compliance & Governance

Ensure proper data governance policies are
in place

Verify tenant compliance with data-region

settings and organisational requirements

Treat Copilot as an intelligent assistant, not a replacement for project management expertise and
judgement
When to Use Copilot
Optimal Scenarios & Implementation Guidance

Project Initiation
Generate quick work breakdown structures for new projects

Periodic Risk Reviews

Regular risk identification and mitigation planning

Frequent Reporting
Weekly or monthly status updates for stakeholders

Fast Project Updates

Quick adjustments to plans and rapid communication

Best Practice Approach

• Use AI outputs as starting drafts, not final deliverables

• Involve team members in reviewing generated plans and risks
• Combine AI efficiency with human expertise and contextual knowledge
• View Copilot as a productivity amplifier, enhancing — not replacing — PM judgement
Key Takeaways & Next Steps
Copilot accelerates project setup and reporting
Task plans, risk assessments, and status reports generated in minutes

Human oversight remains essential

Always review, refine, and validate AI-generated outputs

Natural language makes Project Operations more accessible

Chat-style interface reduces learning curve for new users

Recommended Next Steps

1. Enable Copilot feature in your environment

2. Train team members on responsible AI usage
3. Pilot on a small, low-risk project
4. Establish review processes for AI-generated content
5. Review AI outputs carefully and document adjustments needed
6. Scale adoption based on pilot results
Deploy & Configure Copilot
for Sales

Architecture,
Setup & Best
Practices
For IT Administrators and CRM Consultants
What is Copilot for Sales?
An AI-powered add-on that transforms sales workflows by
integrating intelligent assistance directly into Outlook,
Teams, and Dynamics 365.

Key Capabilities
• Contextual email and meeting summaries
• Automated CRM data enrichment
• Real-time insights during customer interactions
• Productivity acceleration across sales processes
Architecture Overview

Data Sources
Dynamics 365, Outlook, Microsoft 365, CRM entities

AI Processing
Azure OpenAI endpoints, secure data processing

Security Layer
Permissions, consent flows, tenant controls

Client Integration
Outlook add-in, Teams app, Sales Hub
Deployment Targets
Copilot for Sales integrates seamlessly across multiple touchpoints in the sales workflow.

Outlook Integration Teams Meetings Sales Hub

AI-powered email summaries, Meeting insights, action items, Enhanced CRM with contextual AI
suggested responses, CRM updates conversation summaries assistance, data enrichment
Prerequisites & Initial Setup
01

Verify Regional Availability

Confirm Azure OpenAI endpoint availability or consent for cross-region data movement

02

Licence Verification
Ensure appropriate licences assigned to users

03

Admin Consent
Configure tenant-level settings in Power Platform Admin Centre

04

Enable in Environment
Activate Copilot in Sales Hub and Dynamics 365 environment
Configuration & Customisation

Core Configuration Advanced Customisation

Enable AI Features
01

Activate AI in Sales Hub

Navigate to settings and enable Copilot features

02

Enable Auditing
Turn on auditing for leads, opportunities, accounts tables

03

Configure Access Controls

Set permissions at tenant, Entra group, or environment level

04

Verify Deployment
Test with pilot group, validate AI context features
Best Practices & Governance

Always Review AI Outputs

Treat summaries and suggestions as assistance requiring human validation before use

Ensure Data Privacy Compliance

Align deployment with organisational policies and regulatory requirements

Controlled Rollout Strategy

Start with small user group, monitor behaviour, audit changes, verify data flows

Manage Consent Settings

Regularly review and document consent flows and cross-region data movement approvals
Limitations to Consider
Regional Restrictions Data Residency
Azure OpenAI availability May require cross-border data
varies by geography processing consent

Feature Dependencies
Requires specific licences and environment configurations

Important: AI-generated content should always be reviewed for

accuracy and appropriateness before sharing with customers or
stakeholders.
Your Action Plan

Verify Prerequisites Enable & Configure

Check regional availability, licences, admin access Activate in tenant, configure forms, set permissions

Pilot & Refine Scale Deployment

Test with small team, gather feedback, adjust settings Roll out to broader organisation with governance in place

Next Steps: Review Microsoft Learn documentation, schedule pilot programme, establish feedback loop with users
Adopt FinOps on Azure

Cloud Cost
Management
& Optimisation
What is FinOps on Azure?
FinOps Definition Why It Matters
Cultural practice bringing finance, technology and business • Cloud costs can escalate rapidly
teams together to manage cloud spend efficiently • Unpredictable spending patterns

Azure-specific approach to maximise cloud value whilst • Need for visibility and accountability
maintaining cost control • Optimisation opportunities often missed
Core FinOps Capabilities
Cost Usage Tracking
Monitor and analyse Azure resource consumption patterns

Cost Allocation
Chargeback and showback to teams and business units

Optimisation
Rightsizing, savings plans, and efficiency improvements

Governance
Policies, controls, and compliance frameworks

Critical enabler: Resource tagging and organisational structure for

transparency and accountability
Getting Started: First Steps
01

Audit Current State

Identify cloud usage and establish cost baseline across all Azure resources

02

Define Structure
Set up subscriptions, management groups, tagging policies, and ownership models

03

Enable Azure Tools

Activate Cost Management, Azure Policy, monitoring solutions, and reporting
dashboards

04

Establish Governance
Create standards, approval workflows, and cost allocation frameworks
Practical Optimisation Strategies
Quick Wins
• Rightsizing: Match VM types to actual workload needs
• Eliminate waste: Shut down idle or unnecessary resources
• Leverage commitments: Use reservations and savings plans
strategically
• Optimise storage: Select appropriate tiers and lifecycle policies

Key principle: Balance cost reduction with business value delivery —

cheaper isn't always better
Continuous Review is Essential
FinOps isn't a one-time project

Evaluate Usage Optimise Resources

Review consumption patterns and Implement efficiency improvements
spending trends and cost reductions

Iterate & Improve Allocate Costs

Refine processes and governance Assign spending to teams and projects
continuously accurately
Building FinOps Culture

Cross-Team Shared Accountability

Collaboration
Create cost ownership at every
Unite finance, operations, and level of the organisation
development teams around
shared cost goals

Transparency & Reporting

Enforce tagging standards, cost allocation, and regular reporting
cadence

"FinOps demands collaboration across teams — not just DevOps or IT,

but finance and business stakeholders too"
The FinOps Iteration Workflow

Baseline Costs
Establish current spending and usage patterns

Plan Actions
Define optimisation and governance initiatives

Implement Changes
Execute planned improvements and policies

Monitor & Assess

Track results and measure impact

Report & Iterate

Share outcomes and plan next cycle

Set measurable goals for each iteration — track progress against defined metrics
Implementation Checklist

Foundation Ongoing Practice

01 01

Audit Azure Spend Schedule Regular Audits

Analyse current costs and resource usage Establish cost review and optimisation cadence

02 02

Define Resources & Tags Assign Responsibility

Establish organisational structure and ownership Define roles, accountability, and reporting workflows

03 03

Enable Monitoring Tools Foster FinOps Culture

Activate Cost Management and governance solutions Build cross-functional collaboration and awareness
Key Takeaways
Understand Organise Apply Tools
Gain visibility into cloud cost and Structure resources, define roles, Leverage Azure Cost Management
usage patterns implement tagging and governance solutions

Iterate Govern
Review and optimise regularly through FinOps cycles Enforce policies, accountability, and transparency

Start Your FinOps Journey Today

Maximise cloud value whilst maintaining cost control on Azure
Finance Agents (Copilot for
Finance)
Automate Reconciliation, Email Triage, and Customer
Workflows
What Are Finance Agents?

AI-Powered Assistants Email Intelligence Excel Reconciliation

Embedded directly in Outlook and Summarize threads, draft responses, Match transactions, identify
Excel to automate routine finance surface customer context and invoice discrepancies, suggest adjustments
tasks data with AI assistance

Business Outcomes: Reduce manual data entry, accelerate month-end close, improve customer response times
Prerequisites & Setup Requirements
Licensing Admin Roles

• Microsoft 365 E3/E5 or Business Premium • Global Administrator or Dynamics 365 Administrator
• Copilot for Microsoft 365 license • Power Platform Administrator for environment setup
• Dynamics 365 Finance or compatible ERP connector • Security permissions for data source connections

Important: Users must consent to data access for ERP and CRM systems before agents can surface financial records
Key Capabilities
1 2 3

Summarize and Triage Emails Customer Details Enrichment Outstanding Invoices Insights
View aging summaries, payment
Automatically detect finance-related Pull credit limits, payment history, terms, recent transactions without
messages, extract key details, open disputes from ERP into Outlook leaving inbox
prioritize by urgency sidebar

4 5

Compose and Reply with Context Excel Reconciliation Accelerators

Draft professional responses referencing invoice numbers, Match bank statements to ledger entries, identify
amounts, due dates variances, propose corrective entries
Workflows in Outlook
01

Summarize Threads
Generate concise summary of multi-party email chains involving invoices or payment disputes

02

Draft Replies with Referenced Data

Agent suggests response text including invoice totals, due dates, payment instructions

03

Insert Customer Details

Surface account balance, credit status, primary contacts in sidebar pane

04

View Outstanding Invoices

List all open invoices for sender's organization, sortable by age or amount
Workflows in Excel
Reconciliation Experience

• Partial Matching: Identify transactions that match on date and

reference but differ in amount
• Amount Tolerances: Configure acceptable variance thresholds for auto-
match suggestions
• Matching Without Monetary Key: Use description, vendor name, or
custom fields when amounts unavailable

Agent highlights unmatched rows, proposes candidates, logs decisions for

audit trail
Deploy Finance Agents

Enable Features Select Pilot Group

Turn on Copilot for Finance in Microsoft 365 Admin Center Start with 10-20 accounts payable or receivable users

Collect Feedback Scale Rollout

Weekly surveys on accuracy, time savings, pain points Expand to broader finance team after validation

Rollout Guidance: Plan 4-6 week pilot, document use cases, refine permissions before general availability
Architecture Overview
Data Sources
Microsoft 365 mailboxes, Excel files, Dynamics 365 Finance, third-party ERP via
connectors

AI Orchestration
Azure OpenAI Service processes prompts, applies grounding data, returns contextualized
responses

Policy Controls
Data Loss Prevention, Conditional Access, Information Protection labels enforce
governance

Tenant Boundaries
All processing within your Microsoft 365 tenant; no data shared across organizations
Important Limitations

Human Review Required Data Access Scope

Always validate agent-generated email drafts and Agent can only surface records user already has
reconciliation suggestions before finalizing permission to view in source systems

Domain Nuance Gaps Consent Requirements

May not understand company-specific terminology, Users must explicitly grant access to ERP data; revocation
invoice formats, or reconciliation rules without tuning disables agent features
Best Practices & Governance
Responsible Use Guidelines Technical Controls

• Treat agent output as draft; verify figures and context • Permission Scoping: Limit ERP connector access to
• Do not share sensitive customer data via agent-generated necessary tables
emails without approval • DLP Policies: Block sharing of credit card, SSN, or
• Escalate anomalies to finance manager regulated identifiers
• Auditing: Enable Microsoft 365 audit logs for agent usage
and data access

Change Management: Conduct hands-on training sessions, provide job aids, establish support channel for questions
Key Takeaways & Next Steps
Enable Pilot
Configure Copilot for Finance in Admin Center, connect ERP Deploy to small group, measure time savings and accuracy
data sources improvements

Collect Feedback Scale

Weekly check-ins, document edge cases, refine prompts Roll out to full finance organization, integrate into standard
and permissions operating procedures

Improvement Cadence: Monthly review of agent usage analytics, quarterly update of grounding data and prompt templates
Build Effective Agents with
Microsoft Copilot Studio
Plan, Design, and Validate Generative Agents
Use Cases
Customer Support
Answer policy questions, process returns, escalate complex issues to human
agents

HR Self-Service
Onboarding checklists, benefits enrollment guidance, PTO balance lookups

Operations Automation
Facility requests, IT ticketing, procurement workflows triggered by conversation

Finance Inquiries
Expense report status, vendor payment timelines, budget approvals

Why Agents vs. Traditional Apps: Conversational interface reduces training time, natural
language reduces form complexity, AI handles ambiguity
Plan for Agents
01 02 03

Problem Framing KPIs Guardrails

Define specific pain point: high ticket Measure resolution rate, time to answer, Identify topics agent should not handle:
volume, repetitive inquiries, manual user satisfaction, deflection from live financial advice, legal interpretation,
lookup tasks agents medical guidance

04 05

Data Sources Knowledge Scope

List knowledge bases: SharePoint sites, Dataverse tables, public Determine domain boundaries: internal policies only, or include
websites, file libraries product documentation and FAQs
Topics & Conversation Design
Trigger Phrases Small Talk & Fallbacks

• User utterances that start a specific topic flow • Graceful handling of greetings, thank-yous, off-topic questions
• Examples: "reset password", "track my order", "request time off" • Fallback topic when confidence is low: "I didn't understand, can you rephrase?"

Intents & Entities

• Intent: User goal extracted from message

• Entity: Key data points like date, amount, product name
Design Your Agent

Persona & Tone Escalation Rules

Professional vs. friendly, formal vs. casual language, brand When to transfer to human: unresolved after 3 turns,
voice alignment explicit user request, sensitive data involved

Safety Filters Content Moderation

Block profanity, personal attacks, requests for regulated Azure Content Safety checks for harmful content, hate
information speech, violence before responding

Inclusive Design: Test with diverse user personas, support multiple languages, accommodate accessibility tools
Build Solutions
Connecting Data Sources

• Dataverse: Native integration with tables, views, and business rules

• SharePoint: Index document libraries, surface knowledge articles
• Websites: Crawl public URLs, embed FAQ content

Actions

• Power Automate Flows: Trigger workflows for approvals, notifications, data updates
• Connectors: Integrate with ServiceNow, Salesforce, SAP, custom APIs

Prompt Templates & Grounding Data

Provide agent with system message, example Q&A pairs, authoritative documents to reduce hallucinations
Build a Conversational Agent
Multi-Turn Flows Disambiguation
Agent asks clarifying questions, remembers context across Present options when multiple interpretations possible:
messages "Did you mean Product A or Product B?"

1 2 3

Confirmations
Verify user intent before executing action: "Should I submit
this expense report?"

Inclusive Design Considerations: Keyboard navigation, screen reader support, plain language summaries for complex responses
Validate & Iterate
Test Plans

• Unit tests for each topic branch

• End-to-end scenarios with sample user inputs

• Load testing with concurrent users

Analytics & Transcripts

• Conversation completion rate

• Escalation frequency

• Sentiment analysis on user feedback

Important Limitations

Over-Reliance on Generative Output

Users may trust incorrect answers; implement confidence thresholds and
disclaimers

Hallucinations Without Proper Grounding

Agent invents plausible-sounding but false information when knowledge base is
incomplete

Throttling & Quotas

Azure OpenAI requests per minute limits, connector call throttling during peak
usage

Connector Constraints
Some third-party systems have rate limits, authentication complexities, or
incomplete APIs
Best Practices & Governance
1 2 3

Data Minimization Retention Policies RBAC

Only surface data necessary for Define conversation log retention Role-Based Access Control for who
conversation; avoid exposing entire period aligned with compliance can edit topics, publish agents, view
databases requirements analytics

4 5

ALM with Solutions Environments Strategy

Application Lifecycle Management: package agents in Separate Dev, Test, Production environments; promote
solutions, track versions, migrate across environments changes via managed pipelines
Key Takeaways & Next Steps
1 2

Start Small Validate Value

Single use case, narrow scope, measurable KPI Pilot with 20-50 users, collect satisfaction scores,
document time savings

3 4

Scale Patterns Establish Metrics

Replicate successful topics across departments, Resolution rate, escalation rate, user satisfaction as
standardize prompt templates success gates

Quality Gates: Minimum 70% resolution rate, maximum 10% escalation rate before production rollout
Manage Agents in Microsoft
Copilot Studio
Environments, Permissions, Monitoring, and Administration
Environments in Copilot Studio
Dev/Test/Prod Strategy Region & Data Residency

• Development: Sandbox for building and unit testing topics • Choose environment region matching compliance requirements

• Test: Integration testing with production-like data • Data at rest stored in selected geography

• Production: Live environment for end users • Cross-region replication for disaster recovery

Promote changes via solution export/import or ALM pipelines

Generative AI Settings

Enable/Disable Generative Safety Settings Content Filtering

Features
Configure Azure Content Safety Block specific keywords, patterns, or
Per-environment toggle for AI- thresholds: low, medium, high phrases from appearing in agent
powered topic suggestions, dynamic sensitivity for hate, violence, self- responses
responses, content generation harm

Important: Disabling generative features reverts agent to scripted topic flows only
Agent Topic Permissions
Who Can View, Edit, Publish Topics

• System Administrator: Full control over all topics and settings

• System Customizer: Create and edit topics, cannot publish without
approval
• Basic User: Read-only access to published topics, view analytics

Least-Privilege Patterns: Grant edit permissions only to content owners,

require peer review before publish, separate dev and prod access
Entities & Flow Permissions
Managing Entities Connecting & Securing Flows

• Define custom entities: product names, locations, department codes • Power Automate flows run under service account or user context

• Control who can create, edit, delete entity lists • Use least-privilege connections: read-only where possible

• Share entity definitions across topics • Audit flow execution logs for suspicious activity

Variables & Actions

• Scope variables to topics or make global

• Permissions for invoking Power Automate flows

Authentication
AAD/Entra ID OAuth Connections
Single sign-on with Microsoft Entra ID, inherits user's Third-party service authentication via OAuth 2.0, store
permissions for data access tokens securely in environment

Token Scopes Consent Flows

Request minimum scopes necessary: read vs. write, Admin consent for tenant-wide access, user consent for
specific resource access individual permissions
Monitor & Diagnose
Conversation Transcripts CSAT

• Review individual sessions for • Customer Satisfaction score via

quality assurance post-conversation survey
• Identify where users dropped • Track trends over time,
off or escalated correlate with topic changes
• Flag transcripts for training
Error Handling
data
• Fallback routes when topic fails
Analytics

• Daily active users, total • Escalation to human agent

conversations • Logging and alerting on system
• Topic engagement rates errors
• Average session duration
Administer & Manage
01 02 03

Export/Import Solution Packaging ALM

Package agent as solution, export from Include dependencies: entities, flows, Application Lifecycle Management:
source environment, import to target connections, environment variables version control, change tracking, rollback
capability

04 05

Versioning Rollback
Semantic versioning for major, minor, patch changes; maintain Restore previous solution version if issues detected post-
release notes deployment

Best Practice: Test in non-production environment, obtain stakeholder approval, schedule deployment during low-usage window