Six tenets, applied across the lifecycle
Before an agent is built, its behavior must be bounded by Microsoft's Responsible AI principles. They are not a single checkbox but an integrated discipline spanning design, development, deployment, and operations.
| Tenet | What it requires |
|---|---|
| Fairness | Treat all people equitably; avoid harmful bias and discrimination in decisions. |
| Reliability & safety | Perform predictably in expected and unexpected conditions; monitor and validate continuously. |
| Privacy & security | Protect personal data across the lifecycle; secure models from misuse; meet compliance. |
| Inclusiveness | Empower everyone; build accessibility and inclusive design in; avoid excluding any group. |
| Transparency | Let users understand AI logic, data use, decisions, and limitations. |
| Accountability | People — not machines — own outcomes; maintain governance and human oversight. |
Key concept · RAI is a lifecycle, not a step
Apply the tenets at every phase: design (define risks, users, safety scenarios), development (fairness testing, security reviews, transparency docs), deployment (human oversight, pre-production validation, monitoring), and operations (continuous improvement, retraining, policy adaptation).
Task, autonomous, and prompt-driven
Every Copilot Studio agent shares a modular anatomy — goals, skills, actions, knowledge, context, and safety rules — but the three archetypes differ in what starts them and how much they decide on their own.
Task agents
Complete a specific, structured action on request — create a record, summarize an inquiry, trigger a workflow. Designed around one outcome-based goal, with skills (language understanding, data interpretation, planning, execution), actions (connectors/APIs/Dataverse), governed knowledge, and explicit rules.
Autonomous agents
Run independently, started by triggers (user input, system changes, scheduled events) rather than waiting for a prompt. They reason over instructions and grounded knowledge, execute actions, and publish into Teams or enterprise workflows.
Prompt & response agents
Conversational agents combining NLU, topic logic, and event-driven behavior. Generative answers (NLU Boost) act as a fallback when topic matching fails; system topics handle greeting/escalation/errors; condition nodes branch; triggers fire on events.
Key concept · Six-part agent anatomy
- Goals — the mission (actionable, observable, testable, tied to a business outcome).
- Skills — capabilities: language understanding, data interpretation, planning, execution.
- Actions — executable operations via connectors, custom APIs, Dataverse, or cloud flows (each with inputs, output schema, auth, error handling).
- Knowledge — governed grounding: SharePoint, SOPs, product manuals, Dataverse, knowledge articles.
- Context — enterprise data and signals that shape decisions.
- Safety & rules — guardrails that constrain what the agent may and may not do.
Choosing the mechanism inside a prompt-driven agent is a design decision in its own right:
| Mechanism | Use when |
|---|---|
| Generative answers (NLU Boost) | Question isn't covered by a topic; answer must come from knowledge sources. |
| System topics | Greeting, fallback, escalation, errors, disambiguation. |
| Condition nodes | Branch the dialog on variables or values (VIP vs standard, form input, context). |
| Event triggers | Agent must react to an event (file uploaded, row added, schedule) rather than a query. |
Automating and shaping agent behavior
Agent flows run natively inside Copilot Studio to automate multi-step work — every flow is a trigger (manual, scheduled, event, or another agent) plus actions (send email, update data, retrieve info). They can be authored in natural language or refined on a visual canvas, and — unlike Power Automate cloud flows — consume Copilot Studio message capacity rather than a separate license.
| Feature | Agent flow | Cloud flow |
|---|---|---|
| Licensing | Copilot Studio message capacity | Power Automate license |
| Best for | AI-driven conversational automation | Enterprise integration automation |
| Builder | Natural language + visual canvas | Connector-driven workflow design |
Key concept · Prompt actions as reusable instruction blocks
A prompt action is a reusable block that makes an agent respond, reason, transform, summarize, extract, generate, or classify consistently. Author them with the Prompt Coach template — Goal → Context → Instructions/Rules → Examples (optional) → Output Format — and embed them in topics, agent flows, fallback logic, or business processes (e.g. "summarize case details before escalating"). Keep them modular; one task per prompt.
Copilot Studio's four authoring surfaces — Describe, Configure, Test, Publish — carry an agent from intent to deployment, with grounding (SharePoint, M365, websites) and capabilities like Code Interpreter and image generation added along the way.
Measuring value and driving uptake
An agent that no one uses delivers no value. The module closes by separating two interdependent ideas: success criteria (what outcomes the solution should deliver) and adoption goals (how the organization embraces and scales them). Success criteria must be business-aligned, measurable, outcome-driven, time-bounded, and feasible with available data.
| Category | Example indicator |
|---|---|
| Business value | Reduce processing time by 40%. |
| Operational efficiency | 30% fewer manual tasks in a workflow. |
| User experience | 20% faster task completion. |
| Quality & accuracy | AI recommendation accuracy ≥ 85%. |
| Risk & compliance | 100% of outputs audited for governance. |
| Scalability | Handle 10k requests/day. |
Key concept · CAF AI adoption — four pillars
- Establish business outcomes — northstar objectives tied to strategy (satisfaction, cost, revenue, decision speed).
- Identify measurable AI scenarios — high feasibility × high impact.
- Assess AI feasibility — data quality, security, skills, model appropriateness.
- Establish program governance — responsible AI standards, lifecycle management, metrics, data protection.
Adoption itself spans four readiness areas — organizational (leadership alignment, AI literacy, change enablement), data (quality, availability, governance), technical (infrastructure, integration, lifecycle), and user (value demonstration, training, feedback loops). People and process must evolve alongside technology.
Module 5 quiz
Fifteen questions on responsible AI, the three agent archetypes and their anatomy, agent flows and prompt actions, and success/adoption planning. Answer first, then expand Show answer.
Part A · Multiple choice
Which is NOT one of Microsoft's six Responsible AI tenets?
- A Fairness
- B Profitability
- C Transparency
- D Accountability
Show answer
B — Profitability. The six are fairness, reliability & safety, privacy & security, inclusiveness, transparency, and accountability.
What primarily distinguishes an autonomous agent from a task agent?
- A It has no goals
- B It is started by triggers and reasons/acts independently rather than waiting for a prompt
- C It cannot access knowledge sources
- D It runs only in Excel
Show answer
B. Autonomous agents fire on triggers (user input, system change, schedule) and execute independently; task agents complete a specific requested action.
In a prompt-driven agent, which mechanism answers a question that no topic matches by drawing on knowledge sources?
- A Condition node
- B System topic
- C Generative answers (NLU Boost)
- D Event trigger
Show answer
C — Generative answers (NLU Boost). It's the fallback when topic matching fails, grounded in websites, SharePoint, Dataverse, or connectors.
Which is NOT part of the six-part agent anatomy?
- A Goals
- B Skills
- C Billing tier
- D Safety & rules
Show answer
C — Billing tier. The anatomy is goals, skills, actions, knowledge, context, and safety & rules.
An agent flow always consists of which two building blocks?
- A A model and a dataset
- B A trigger and one or more actions
- C A prompt and a persona
- D A license and a connector
Show answer
B. Every agent flow has a trigger (how it starts) and actions (what it does).
The Prompt Coach template structures a prompt action into which sections?
- A Intro, Body, Conclusion
- B Goal, Context, Instructions/Rules, Examples, Output Format
- C Trigger, Action, Condition
- D Fairness, Safety, Privacy
Show answer
B. Goal → Context → Instructions/Rules → Examples (optional) → Output Format, for predictable, aligned output.
Which are the four authoring surfaces in Copilot Studio?
- A Draft, Review, Approve, Ship
- B Describe, Configure, Test, Publish
- C Plan, Build, Deploy, Retire
- D Ingest, Train, Evaluate, Serve
Show answer
B — Describe, Configure, Test, Publish.
Which pillar is NOT one of the CAF AI adoption pillars?
- A Establish business outcomes
- B Identify measurable AI scenarios
- C Maximize model size
- D Establish program governance
Show answer
C — Maximize model size. The four pillars are business outcomes, measurable scenarios, feasibility assessment, and program governance.
Part B · True or false
Responsible AI is a single implementation step completed at design time.
Show answer
False. It's an integrated discipline across the whole lifecycle — design, development, deployment, and operations.
Grounding an agent in approved organizational knowledge improves accuracy and reduces incorrect responses.
Show answer
True. Agents should reference governed knowledge sources rather than generating without context — a core reliability practice.
Agent flows require a separate Power Automate license to run.
Show answer
False. Agent flows run natively in Copilot Studio and use its message capacity — no separate Power Automate license.
Success criteria and adoption goals are interdependent and should reinforce each other.
Show answer
True. Success criteria define the outcomes; adoption goals ensure the organization realizes them at scale.
Part C · Short answer
Name the three agent archetypes in this module and the one thing that most distinguishes each.
Show answer
Task (completes a specific requested action), autonomous (started by triggers, acts independently), and prompt & response (conversational, combining NLU, topics, conditions, and event triggers).
List the four adoption-readiness areas an organization must prepare.
Show answer
Organizational readiness, data readiness, technical readiness, and user adoption.
Give three of the five characteristics that well-defined success criteria must have.
Show answer
Any three of: business-aligned, measurable, outcome-driven, time-bounded, and feasible with available data and systems.