Copilot Studio: Generate HTML Battle Cards Without Consuming Messages

Preparing a client call on a technical subject in less than ten minutes, without sacrificing documentation rigor: that's exactly what the combination of Copilot Studio, the Researcher agent from Microsoft 365 Copilot, and Power Automate enables. This article is aimed at M365 administrators and solution architects looking to optimize message consumption while automating the production of structured content.

Why HTML is the ideal output format for an AI agent

Large language models (LLMs) have been trained massively on web content, which is structured in HTML. In practice, this means these models generate HTML with significantly higher reliability than other output formats. A response in plain Markdown can pose interpretation issues depending on the client; an HTML file is immediately readable in any browser, with no dependency on a third-party rendering engine.

The other concrete advantage: HTML is saved as a single file, can be converted to PDF via Power Automate (action OneDrive for Business – Convert file), and integrates seamlessly into a SharePoint or Teams workflow.

Solution architecture in three blocks

The architecture relies on three distinct components, each with a precise role:

1. Researcher — the search agent included in Microsoft 365 Copilot, which queries the web and official documentation (including Microsoft Learn) without consuming Copilot Studio messages.
2. A custom Copilot Studio agent — responsible only for HTML formatting based on search content provided as input.
3. A Power Automate flow — orchestrates the two, passes data from one block to another, and saves the produced HTML file to OneDrive.

The separation between the search phase and the HTML generation phase is the key point of the architecture: it avoids calling browsing capabilities from the Copilot Studio agent, which would consume messages.

Creating the HTML generation agent in Copilot Studio

The agent has a unique and well-defined role: receive a block of search text, structure it, and return pure HTML.

1Based on the research provided, create a one-page HTML battle card style document.
2The document must contain:
3- What it is (definition)
4- An analogy to explain the concept
5- Core concepts
6- Common misconceptions and how to correct them
7- When to use it / when not to use it
8- A quick example
9- Relevant documentation links (Microsoft Learn preferred)
10 
11Respond only with pure HTML — no markdown, no explanation, no code block fence.
12The HTML must be self-contained (inline CSS, dark mode preferred).

Building the Power Automate flow

This is where the message savings lie. The Copilot node in Power Automate can call Researcher without consuming Copilot Studio messages. The same node can then call your custom agent.

1Only write two to three pages on the topic: [research_topic]
2Do not ask extra questions. Research broadly on the topic.
3Focus mainly on Microsoft knowledge but explore further if relevant.
4Provide links from Microsoft Learn when available.

Observed results and performance

On a test covering the topic Custom Connectors Power Platform, the complete sequence proceeded as follows:

The HTML file produced contained: concept definition, analogy (the universal travel adapter), five core concepts, authentication options, common design errors, platform limitations, and a list of direct links to Microsoft documentation (overview, creation, DLP, certification).

Implementation

The following script illustrates how to call your Copilot Studio agent directly via Microsoft Graph for testing outside Power Automate flows. It requires the Microsoft.Graph module and access with the minimum Copilot Studio User role.

1# Required module: Microsoft.Graph (>= 2.x)
2# Install-Module Microsoft.Graph -Scope CurrentUser
3# Minimum permission: CopilotStudio.Conversations.ReadWrite (delegated)
4 
5# Parameters to adapt
6$TenantId     = "<your-tenant-id>"
7$ClientId     = "<your-app-registration-client-id>"
8$AgentId      = "<id-of-your-copilot-studio-agent>"
9$ResearchText = "Paste here the content returned by Researcher"
10 
11# Interactive connection (delegated flow)
12Connect-MgGraph -TenantId $TenantId -ClientId $ClientId `
13    -Scopes "CopilotStudio.Conversations.ReadWrite"
14 
15# Creating a new conversation with the agent
16$ConversationBody = @{
17    agentId = $AgentId
18} | ConvertTo-Json
19 
20$Conversation = Invoke-MgGraphRequest -Method POST `
21    -Uri "https://api.copilotstudio.microsoft.com/environments/{environmentId}/bots/$AgentId/conversations" `
22    -Body $ConversationBody `
23    -ContentType "application/json"
24 
25$ConversationId = $Conversation.id
26Write-Host "Conversation created: $ConversationId"
27 
28# Sending search content to the agent
29$MessageBody = @{
30    content = $ResearchText
31} | ConvertTo-Json
32 
33$Response = Invoke-MgGraphRequest -Method POST `
34    -Uri "https://api.copilotstudio.microsoft.com/environments/{environmentId}/bots/$AgentId/conversations/$ConversationId/activities" `
35    -Body $MessageBody `
36    -ContentType "application/json"
37 
38# Retrieving the HTML response
39$HtmlContent = $Response.value | Where-Object { $_.type -eq "message" } | Select-Object -Last 1 -ExpandProperty text
40 
41# Saving the HTML file locally
42$OutputPath = ".\battle-card-output.html"
43$HtmlContent | Out-File -FilePath $OutputPath -Encoding UTF8
44Write-Host "HTML file saved: $OutputPath"
45 
46# Verification: open in default browser
47Start-Process $OutputPath

Troubleshooting common errors

The OneDrive node displays "Need setup" Double-click the node to force open the configuration panel. The Power Automate editor doesn't always open it automatically on first insertion.

The agent returns JSON instead of raw HTML Verify that the Output type parameter of the Copilot node is set to Text response and not the default value. Also add in the agent's instructions: Respond only with pure HTML — no markdown fence, no explanation.

Researcher doesn't bring back Microsoft Learn links Explicitly add provide links from Microsoft Learn when available to the Researcher node's prompt. Without this instruction, the agent sometimes prioritizes generic sources.

Execution time exceeding 10 minutes Power Automate imposes a 30-minute limit on manually triggered flows, but Researcher can be slow on broad topics. Reduce the request to "one to two pages" to keep search time under 5 minutes.

The produced HTML file is empty Verify that the expression referencing the agent output is correct. In the Power Automate expression editor, use outputs('Agent_step_name')?['text'] to extract the raw text field.

What you can do right now

This architecture is reproducible in less than an hour on any M365 tenant with Copilot Studio and M365 Copilot licenses. Here are the concrete actions to chain together:

• Create the agent Create Battle Card in Copilot Studio with the HTML formatting instructions above.
• Publish it and note its identifier to reference it in Power Automate.
• Build the flow with the two Copilot nodes (Researcher → your agent) and the OneDrive action.
• Test on an internal topic — for example, a component of your technical stack that you often need to explain to stakeholders.

If your organization manages a tight Copilot Studio message budget, the point of this architecture goes beyond simple sheet generation: it demonstrates that it is possible to delegate the costly part (web search) to a component included in the M365 license, and reserve Copilot Studio messages for high-value transformations. This logic applies to any workflow where an enrichment step precedes a generation step.