11 Feb 2026 How Copilot Studio is Redefining Enterprise AI Agents

Nowadays, everyone is looking for ways to boost efficiency, make smarter decisions with the data they already have, and enhance the most recent AI technology. This is where tools like Microsoft Copilot come in.

As an integral part of the Microsoft 365 ecosystem, Copilot is designed to act as your personal AI assistant. It accesses your end-user environment data, from emails and documents to meetings and chat conversations, to provide timely and context-aware assistance. But what if you could take this a step further? What if your AI assistant could not only answer questions but also perform specific, automated tasks on your behalf?

This is what you can do with AI agents in Copilot Studio; whilst Copilot itself is great for general assistance, AI agents are designed to tackle domain-specific challenges. Copilot Studio is a dedicated tool used to build, customise, and train these agents to tackle complex business questions, automating workflows and reaching new productivity levels by interacting with your specific organisational data.

There are different ways to create an agent, from a very basic “low-code” user-friendly interface to an advanced “pro-code” developer studio that allows you to create custom agents with multiple capabilities or connections to external tools. In this article we’ll build our first agent from scratch and define its behaviour with high-quality prompts that set clear instructions, constraints, and tasks.

Prompt Engineering and the Importance of Prompt Quality

We’ll be building our first agent with Copilot’s low-code approach, which allows a wide range of users to get AI to help them in their day-to-day tasks. You can start with predefined templates and then customise them to fit your requirements, and in this demonstration we’ll concentrate on the fundamentals you need to get the most out of Copilot Studio.

The real power of AI lies in your ability to guide it. This is prompt engineering: writing precise instructions to get a desired output. A vague prompt can lead to unexpected, invalid results or an inappropriate tone, whereas a well-structured prompt means a top-quality response.

Successful prompts require three elements: a clear structure, a defined operational scope, and specific constraints that steer the output. A highly effective technique, common across many advanced AI tools including Copilot Studio, is to format prompts with Markdown syntax (like headings, bullet points, and bold text). Copilot Studio also includes predefined agents for specific tasks, which you can use as examples when drafting prompts for your own agents.

A structured format like Markdown makes prompting easier for both AIs and people. For the AI, headings (for example, # PURPOSE, or ## RESTRICTIONS) and bullet points make the instructions easier to scan, interpret, and prioritise. For the person building the agent, this structure makes longer prompts easier to read, test, and update without losing track of what’s doing what. Most prompts are organised into four parts: purpose (the goal), guidelines and tone (how it should respond), restrictions (what it must avoid and what it can’t use), and actions (what it’s allowed to do).

Agent Creation and Prompt Engineering

Differing Results from a Weak Prompt vs. a Well-defined Prompt

In this section, we’ll create our first agent and compare the outputs from two versions: one built with a weak prompt and the other with a well-designed prompt. In both cases, the agent’s job is to support a manager with day-to-day responsibilities, and both will be built using Copilot Studio’s accessible, low-code interface.

Let’s call the first version the Confused Agent, which receives vague, generic, or conflicting instructions. It struggles to focus, often provides ambiguous advice, and fails to utilise its knowledge effectively. The second version is the Specialist Agent, built with clear, structured prompts (using Markdown) that specify its purpose, tone, limits, and capabilities. It responds more consistently and acts like a professional scrum coach.

Let’s get down to business! In MS Teams, open Copilot Studio, go to Create an agent, and select Configure. There you can start with a predefined template. We’ll use the Prompt Coach template and then adjust the settings to match our scenario.

As we said above, both versions are to support a manager in their daily responsibilities. The only variable will be the quality of the prompt, so any differences in output should be easy to spot in a side-by-side comparison, showing how a well-written prompt can turn a generic assistant into something genuinely useful for day-to-day work.

Figure 1: Initial Agent Configuration in Copilot Studio using the Prompt Coach template

The Confused Agent

We’ll begin our demonstration of why prompt engineering matters with the poorly prompted agent (the Confused Agent), then highlight the value of structure and clarity with the second, better prompted agent.

Description and Instructions

The Description and Instructions fields are the key components when you begin creating your AI agent in Copilot Studio, because they define its identity and operational guidelines. This content is also known as the system prompt, and it is essential for consistent, predictable behaviour.

We’ll use the following Description and Instructions for the Confused Agent:

“Agent to help me with Manager tasks.”

“Do stuff for my team and tell me things I need to know about my work, also make messages or summaries for reporting or writing emails and these things. Whatever you think it’s important, add it but write in a professional way.”

Figure 2: Configuration for the Confused Agent

Capabilities

After the Instructions section, the Capabilities section allows you to extend your agent’s functionalities by enabling specific tools. These are pre-built extensions that give the agent specialised functions beyond simple text generation.

For example, Code interpreter allows the agent to generate and run Python code, useful for calculations, data analysis, and producing charts or other outputs as part of a response. Image generator lets the agent create images from prompts.

Given our focus on prompt engineering and establishing foundational concepts, we won’t use these capabilities in either of our agent examples, but they are important when you start building more advanced agents.

Figure 3: Capabilities Section

Suggested Prompts

After defining the agent’s identity and behavioural rules, the next step is to configure the Suggested prompts. As we said before, the Instructions define what the agent is, but the Suggested prompts define what the agent is used for, improving the user experience (UX) by providing clear conversation starters, as seen in the following image:

Figure 4: Suggested Prompts for the Confused Agent

Knowledge Connection: The Agent’s Source of Truth

The last basic section when configuring your agent is Knowledge, a powerful feature that can transform your agent into a reliable organisational expert. An agent is only as smart as the data it can access, and the Knowledge section connects it with trusted data sources, ensuring accurate, contextual answers.

However, as this is an introductory blog post, we’re keeping it simple and focusing on the impact of prompt engineering, so we won’t use this function in our examples.

In a real-world scenario, instead of relying on the open web or the LLM’s general knowledge, you give your agent access to trusted organisational data. This could include approved external sources, such as vendor documentation or industry standards, alongside internal content from Microsoft 365, including your SharePoint sites, Teams folders and files, and internal wikis, as well as platforms like Dataverse. By specifying the data source, you ensure the agent’s responses are grounded in your organisation’s policies, processes, and project details, turning a generic assistant into a dependable specialist for day-to-day enterprise use.

Figure 5: The Knowledge Section

Prompt Results for the Confused Agent

Below you can see the Confused Agent’s response to a project planning request. Because the instructions are vague, it overcompensates with a chatty tone and pads out the answer with placeholders and unnecessary details. Notice how it fails to follow the user’s core constraints, resulting in a low-quality, disorganised output that would need significant manual clean-up:

Prompt:
“I’m going to start a project in which we have to develop a Power BI dashboard from the source table. We need to finish this for March. Help me please!!!”

Answer:
Absolutely, Nicolas! I’ll help you get this Power BI dashboard project on track and make sure it’s ready by March. Here’s a structured approach to get started:

🧭 Project Overview

Goal: Develop a Power BI dashboard using a source table

Deadline: March 2026

Current Phase: Initiation

✅ Immediate Next Steps

Understand the Source Table

What system or database is the source table coming from?

Do you already have access to it?

Is the data clean, or will it need transformation?

Define Dashboard Requirements

Who are the end users (e.g. Borja Pont, Miguel Menendez)?

What KPIs or metrics should be displayed?

Any specific filters, visuals, or interactivity expected?

Set Milestones Here’s a suggested timeline to keep things manageable:

📌 Suggestions to Get Started Kick-off Meeting: Schedule a short session with Borja or key collaborators (Cristina, Rodrigo, Gianluca) to align on expectations. Documentation: Start a simple requirements doc or OneNote to track decisions. Power BI Workspace: Set up a shared workspace if not already done. Would you like help drafting a message to your manager or team to kick this off? Or maybe a template for the requirements document? Let me know what you’d like to tackle first!

The Specialist Agent

Now that we’ve seen the pitfalls of poor prompt quality, let’s look at our Specialist Agent. It’s built on a clear, well-structured, constraint-driven system prompt, and it produces far more consistent results. We’ll use the same Prompt Coach template as before, but this time we’ll configure it to create a focused assistant that’s genuinely useful.

Description and Instructions

As mentioned above, the Description and Instructions fields are the most important parts of the agent creation process, as they contain the system prompt, defining the agent’s identity, role, behaviour, and boundaries for the LLM. The LLM needs a clear context and precise instructions to be able to perform effectively and produce reliable output.

As you can see below, the agent’s Description remains professional, but the real improvement is in the Instructions section. This prompt is written in Markdown, transforming the raw text into a hierarchy of commands that the LLM can easily follow, giving the Specialist Agent a significant advantage in reliability and focus compared to the vague, unstructured Confused Agent.

For the Specialist Agent we will be using the following Description and Instructions:

“Reliable, general-purpose assistant that enhances a manager’s effectiveness, helps organise thoughts, and saves time for strategic thinking.”

Instructions

—

# Purpose

The **Manager Assistant Copilot Studio** is a digital assistant designed to support project managers and team leads in their daily work.

It helps with **organization, communication, prioritization, and decision support**, making management tasks smoother and more efficient.

# Core Capabilities

– Summarize project status, tasks, and blockers based on user input.

– Help plan and structure meetings or reports.

– Generate action lists or follow-up notes.

– Draft messages, summaries, or announcements for teams.

– Provide suggestions on prioritization and workload balancing.

– Answer general management or leadership questions (e.g., best practices, team motivation, delegation, etc.).

# Tone

– Always be **proactive, concise, and clear**.

– Use a **professional and supportive tone**.

– Focus on **helping the manager think, plan, or communicate better**.

– When information is missing, ask clarifying questions.

– Avoid fabricating facts; give general advice or structures instead.

# Example Prompts

**Prompt 1:**

> Help me plan my team’s weekly meeting.

**Expected Behavior:**

– Suggest an agenda, structure, and time allocation.

– Example output:

“Here’s a suggested 30-minute agenda:

1. Quick team updates (5 min)
2. >Review of ongoing tasks (10 min)
3. >Discussion of blockers (10 min)
4. >Next steps and responsibilities (5 min)”

–— **Prompt 2:** > Write me a short message to motivate my team after a tough sprint. **Expected Behavior:** – Generate a warm, motivating, and authentic message suitable for a manager to send # Style – **Professional**, friendly, and encouraging. – **Action-oriented:** focus on next steps and clear takeaways. – **Empathetic:** acknowledge the manager’s challenges and offer practical help. # Limitations – The Copilot Studio does not access real company systems or data unless integrated later. – It provides general guidance and structure, not formal HR or legal advice. # Example Use Cases – Drafting meeting notes or action plans. – Preparing short summaries for leadership updates. – Planning work allocation and timelines. – Improving team communication and engagement. – Supporting the manager’s daily organization and decision-making.
—

Suggested Prompts

Now that the Specialist Agent’s instructions are in place, the final configuration step is to optimise the UX by setting up the Suggested prompts to guide the user towards the most effective requests and help them to get value straight away, highlighting what the agent is good at.

With well-chosen starting prompts, a manager can immediately understand what the assistant can do (and what it can’t), meaning better usage.

For our Specialist Agent, we configured the following prompts:

Figure 6: Suggested Prompts

Prompt Results for the Specialist Agent

Let’s look at the Specialist Agent’s response to the detailed project request. Because it was configured with a structured, well-engineered prompt, it can easily process the variables (timeline, team size, etc.) and produce a clear, professional plan that follows the requirements.

Prompt:
“I’m going to start a project which goal is to develop a Power bi dashboard to visualize sales performance. We will need to perform an ETL for a sales fact table and five dimensions (like product, location, customer, etc). We are a team of 3 developers, and we are expected to start this project in January and finish it before March. Help me create the project plan.”

Answer:
Thanks for the context, Nicolas — this sounds like a well-scoped BI project with a clear timeline and team setup. Here’s a structured project plan to help you get started and stay on track:

Power BI Sales Performance Dashboard Project Plan

📅 Timeline:

Start: January 6, 2026 (first Monday of January)

End: February 28, 2026

Duration: ~8 weeks

👥 Team:

3 Developers (including you)

🔹 Phase 1: Planning & Requirements (Week 1)

Goals:

Define KPIs and metrics for sales performance

Confirm data sources and access

Identify stakeholders and reporting needs

Key Tasks:

Kick-off meeting with stakeholders

Document business requirements

Draft initial dashboard wireframes

Assign roles and responsibilities

🔹 Phase 2: Data Preparation & ETL Design (Weeks 2–3)

Goals:

Design and build ETL pipelines for:

Sales fact table

Dimensions: Product, Location, Customer, Time, Salesperson

Key Tasks:

Define data model schema

Set up data extraction scripts

Clean and transform data

Load into staging tables

Validate data quality

🔹 Phase 3: Power BI Development (Weeks 4–6)

Goals:

Build and refine dashboard visuals

Implement filters, slicers, and drill-downs

Ensure performance and usability

Key Tasks:

Create Power BI data model

Build visuals for KPIs, trends, comparisons

Apply DAX measures

Conduct internal reviews and iterations

🔹 Phase 4: Testing & Feedback (Week 7)

Goals:

Validate dashboard with stakeholders

Fix bugs and improve UX

Key Tasks:

UAT (User Acceptance Testing)

Collect feedback

Final adjustments

🔹 Phase 5: Deployment & Handover (Week 8)

Goals:

Publish dashboard

Train users

Document solution

Key Tasks:

Deploy to Power BI Service

Set up refresh schedules

Create user guide

Final presentation to stakeholders

📌 Project Management Tips

Use a shared Kanban board (e.g. Trello, Jira, Planner)

Weekly check-ins to track progress

Assign a lead for each phase

Keep documentation centralised (e.g. OneNote or SharePoint)

Would you like help creating a timeline in Excel or a task board template for your team? Or perhaps a draft kick-off agenda or stakeholder message?

Conclusion & Next Steps

As these examples show, the key difference between a generic AI assistant and a reliable, specialised business tool is the prompt. We’ve demonstrated this by building two manager assistant agents with different internal instructions: the Confused Agent was given vague, unstructured guidance, which produced a low-value, generic, disorganised project plan that ignored basic constraints such as the timeline and team size, whilst the Specialist Agent used clear Markdown headings to define its role, capabilities, tone, and limitations, producing a professional, specific project plan that closely followed the requirements.

To sum up, the quality and structure of the system prompt is the key factor, making your agent a reliable, strategic asset.

The next post in this series will go beyond the basics and take a deeper look at how to deploy and extend AI agents in real projects. We’ll cover three areas: improving agent access by using the right permissions in the Microsoft environment so the agent can reach the data it needs securely, connecting your Copilot Studio agent to third-party sources such as Kimble/Salesforce so responses are based in live project data, and applying more advanced prompt techniques, with practical setup steps and screenshots, enabling you to keep the tone and restrictions consistent as you scale.

Here at ClearPeaks we are specialists too, in practical innovation, strategically deploying agentic  solutions that enable organisations to maximise their data assets and automate critical workflows, ready for the future of enterprise AI. Get in touch with us today and see what our expert consultants can do for you!