M365 Copilot Upskilling Curriculum | Module 8
Module 8: Building Agents in
Microsoft 365 Copilot
As organizations mature in their artificial intelligence journey, they transition from utilizing broad,
general-purpose AI tools to deploying highly specialized agents. While the standard Copilot for Microsoft
365 environment excels at general productivity, corporate tasks often require localized knowledge,
specific behavioral logic, and structural guardrails. This module introduces the concepts and workflows
necessary to design, build, test, and publish specialized AI agents within the Microsoft 365 ecosystem,
enabling professionals to streamline high-frequency, complex departmental processes.
Agent landscape: Agent Builder (no-code, natural language),
SharePoint agents (grounded on a site/library), Copilot Studio
(connectors, actions, multi-step logic), Power Automate (deterministic
automation); choosing between them
To successfully architect specialized solutions, professionals must understand the Microsoft 365 agent
landscape. An agent is essentially a dedicated, customized instance of Copilot designed to execute a
narrow, predefined scope of work. Rather than forcing a user to manually paste instructions and locate
reference materials in every chat session, an agent pre-packages these parameters into a reusable,
always-available interface. Microsoft provides a spectrum of tools to construct these agents, ranging
from simple, no-code solutions built through conversational dialogue to sophisticated, pro-developer
workspaces that connect to external enterprise databases.
The entry point for specialized design is Agent Builder, a conversational interface built directly into
Microsoft 365 Copilot for Business and Enterprise users. Formerly accessible as Copilot GPT Builder in
consumer environments, Microsoft retired the consumer tool to focus entirely on commercial and
corporate environments where data privacy is guaranteed under Microsoft Entra ID authentication.
Agent Builder operates on a natural language paradigm: instead of writing code, the builder describes
the desired agent's behavior to an AI-powered setup wizard. For example, an HR professional can tell
the builder, "Create an onboarding assistant that explains medical benefits based on our policy PDFs,
utilizing a supportive and polite tone," and the system will automatically draft the technical instruction
block, profile details, and configuration settings behind the scenes.
For teams focused strictly on document repositories, SharePoint agents offer a highly localized, zero-
configuration solution. These agents are built directly on top of specific SharePoint sites, folders, or
document libraries. A SharePoint agent is automatically grounded to the file systems where it resides,
Page 1
M365 Copilot Upskilling Curriculum | Module 8
meaning it inherits all local metadata, folder hierarchies, and administrative restrictions. If an operations
manager stands up a SharePoint agent on a library containing standard operating procedures, any team
member with access to that library can instantly query the agent to summarize engineering
specifications, explain safety procedures, or retrieve historical incident narratives without any manual
prompt engineering.
When an agent must bridge multiple software systems or execute complex conditional paths, Copilot
Studio serves as the primary professional development platform. It provides a visual, graphical
development environment containing advanced capabilities such as custom API connectors, multi-step
logic nodes, custom variables, and integrated telemetry analytics. In Copilot Studio, builders do not rely
solely on probabilistic AI; instead, they can define strict conversation flows, mandate that certain
questions trigger specific database lookups (such as pulling an active customer record from Salesforce),
and enforce precise transactional outcomes.
Finally, true rules-based, transactional processes require Power Automate. While Large Language
Models operate on probabilistic pattern matching—meaning they predict the next logical token but can
exhibit variance or "probabilistic drift"—Power Automate is entirely deterministic. It runs strictly defined
sequential actions with zero guesswork. An effective enterprise design often pairs a conversational
Copilot agent with a deterministic Power Automate flow. The conversational agent gathers customer
requests, synthesizes the requirements, and then hands the structured parameters over to a Power
Automate flow to execute a secure approval, update an ERP database, or distribute a formal report.
Page 2
M365 Copilot Upskilling Curriculum | Module 8
Figure 8.1: Selecting the Right Tool for the M365 Agent Workflow
To guide the selection process, builders must analyze the characteristics of their workflow across
complexity, integration needs, and the required certainty of the output. The table below outlines how to
map typical corporate use cases to the appropriate Microsoft 365 tool:
Tool Category Ideal Business Primary Paradigm Real-World Grounded
Scenario Example
SharePoint Agents Departmental No-code, static Facility Expansion FAQ
document search, document grounding Agent grounded on a
policy lookups, construction
localized FAQs construction folder
Agent Builder Role-specific writing No-code, natural RFP Response Agent
assistants, customized language prompts built in Copilot Chat
campaign ideators for sales proposal
teams
Copilot Studio Database lookups, Low-code visual Northwind Business
multi-step logic gates, canvas + action APIs Insights Agent pulling
CRM writebacks real-time warehouse
stock data
Power Automate Database writing, Deterministic logic, Standardized audit
secure sign-offs, rules-based triggers notifications triggered
transaction records on negative quality
disposition
Anatomy of an agent: name/description, instructions, knowledge
sources, starter prompts, actions, guardrails
Constructing an effective AI companion requires a deep understanding of the anatomy of an agent.
Regardless of the tool selected for building, every custom agent is composed of six foundational blocks.
Designing each element with precision directly determines the accuracy, safety, and adoption rate of the
solution within a corporate department.
1. Name and Description: The first structural block establishes the identity of the agent. The name
must be unique, intuitive, and easy to remember because users invoke agents using the "@" symbol
Page 3
M365 Copilot Upskilling Curriculum | Module 8
in chat. The description is equally critical: it operates as a meta-routing signal for Copilot. When a
user asks a general query, Copilot scans the descriptions of all active agents in the organization's
catalog to determine if the query should be routed to a specialized assistant, such as the Regulatory
Inquiry Assistant Agent.
2. Custom Instructions: Operating as the core programmatic blueprint, custom instructions define
the agent's behavior, tone, style, and logical boundaries. Written in plain English, these instructions
should employ a modular framework, establishing the agent's specific professional role, the
structured tasks it must execute, the context of the department, and strict constraints. Instructions
are parsed by the Large Language Model at the beginning of every interaction, acting as a permanent
system prompt that enforces conversational consistency.
3. Knowledge Sources: To ensure responses are grounded in verified corporate data, builders define
knowledge boundaries. Knowledge sources can include specific OneDrive folders, SharePoint sites, or
external web-indexed URLs. By restricting the grounding boundary, builders prevent the agent from
retrieving outdated public web records. For example, a procurement agent grounded to a specific
SharePoint library containing active vendor pricing sheets will strictly reference those sheets to
calculate expenses, bypassing any extraneous pricing data on the open web.
4. Starter Prompts: Often referred to as conversation starters, these appear as clickable suggestion
tiles when a user opens the agent's chat window. They serve an educational purpose: rather than
staring at a blank prompt box, non-technical users can immediately click pre-formatted options like
"Summarize the latest quality alert audit" or "Generate a standard RFP response template based on
our referenced guidelines."
5. Actions (Connectors and Plugins): While standard agents are passive readers of information,
actions enable them to interact with the physical and digital software environment. Utilizing API
plugins and Power Automate flows, an agent can perform transactional operations, such as writing a
new customer ticket to the IT help desk, executing database lookups, or checking real-time
warehouse inventory levels across a global supply chain.
6. Guardrails: To ensure compliance with corporate policy, builders hardcode behavioral boundaries.
Guardrails specify what the agent must not do, defining topics that are out of scope, preventing the
use of confidential pricing parameters, and mandating human verification gates. A secure quality
agent, for instance, might possess guardrails stating: "Under no circumstances will you make final
quality dispositions or replace official systems of record; you must always direct users to human
engineers for safety-critical approvals."
Page 4
M365 Copilot Upskilling Curriculum | Module 8
Figure 8.2: Architectural Anatomy of a Microsoft 365 Agent
Building, testing, publishing: iterative building, testing against known-
answer questions, recognising poor grounding, permissions,
publishing/sharing/ownership, lifecycle management
The creation of an enterprise-grade agent follows a disciplined, step-by-step methodology. In an
industrial environment where data integrity and compliance are paramount, agents cannot be launched
without rigorous quality control. The pipeline spans four distinct phases: iterative design, validation
testing, security routing, and lifecycle maintenance.
The development phase relies on iterative building. Builders utilize the live preview panels in Agent
Builder or Copilot Studio to chat with the agent side-by-side as they modify its instructions and
knowledge sources. A minor change in system prompt wording—such as instructing the agent to always
lead with structured tables or restricting its vocabulary to plain language—takes effect immediately,
letting the builder observe how phrasing adjustments influence the generated response.
Once a draft configuration is established, builders must conduct validation testing against known-
answer questions. Rather than asking general questions, the testing team must prepare a matrix of 10
to 15 highly specific queries where the exact, correct answers are already known and verified. For
instance, in testing a Regulatory Inquiry Assistant Agent, a builder should ask: "What was the compliance
submission deadline listed for the 2024 environmental audit in Section 4.2?" If the agent provides the
Page 5
M365 Copilot Upskilling Curriculum | Module 8
exact date with a correct citation to the corresponding document, the configuration is stable. If the
agent provides a plausible-sounding but incorrect date, the builder has exposed a critical hallucination.
A critical skill for any AI coordinator is recognizing poor grounding. Weak grounding occurs when an
agent fails to anchor its claims to the provided document sources and begins relying on general training
data. Grounding failure is easy to spot in conversation: the agent will omit clickable citation markers, use
overly general language, write statements that begin with "Typically, organizations..." instead of
"According to the handbook...", or fail to display the specific PDF files used to compile its answer. To
resolve poor grounding, builders must refine instructions, eliminate nested folders, or clean up tabular
structures in the referenced source files.
Before publishing an agent, builders must verify that existing security permissions are fully respected. A
common point of anxiety among non-technical professionals is the fear that an agent might accidently
expose confidential corporate records. Microsoft 365 agents strictly operate within the context of the
user. They utilize standard Role-Based Access Control (RBAC) protocols and respect SharePoint
administrative folder permissions. This means if a standard operations specialist queries a Facility
Expansion FAQ Agent, the agent will refuse to display confidential executive land-acquisition budgets if
the specialist does not have direct read permissions to the source spreadsheet. The agent can only see
and synthesize what the specific querying user is legally authorized to see.
Once approved, the agent enters the publishing, sharing, and ownership phase. Builders can choose to
publish their agent to their personal workspace, share it with a select Microsoft Teams collaboration
group, link it to a department-wide SharePoint site, or submit it to the global enterprise tenant catalog
for administrative approval. Ownership must be clearly defined within the department: who holds
accountability for the accuracy of the agent's instructions, and who manages its file configurations as
business processes change?
This structural governance is handled via lifecycle management. An agent is not a static project; it
requires continuous monitoring. As manufacturing guidelines, safety protocols, and procurement vendor
contract prices change, the underlying document knowledge sources must be actively updated. The AI
coordinator should set up a recurring monthly check to review user feedback, audit the relevancy of
grounded folders, remove duplicate files, and ultimately retire agents that have reached the end of their
operational lifecycle.
Page 6
M365 Copilot Upskilling Curriculum | Module 8
Figure 8.3: Sequential Lifecycle of Agent Design, Testing, and Governance
By mastering these three dimensions—architecting within the proper tool landscape, modeling all
structural anatomical components, and enforcing a disciplined testing-to-lifecycle pipeline—teams can
build highly reliable, secured agents that securely automate repetitive daily tasks and enhance
operational decision-making across Microsoft 365.
Page 7