Power Platform Community Forum Thread Details

Hi everyone,

I’m facing an issue in Copilot Studio where the agent is summarizing input JSON instead of directly calling a tool, even though the trigger comes from Power Automate.

🔹 My End-to-End Flow

SharePoint → File created trigger
Power Automate → Get file content (base64)
Power Automate → Trigger Copilot agent
Copilot should call tool → "Run workflow from JSON"

🔹 What I Send from Power Automate to Copilot

{
  "instruction": "Run workflow from JSON",
  "fileContentBase64": "<base64 content>",
  "fileName": "workflow_schema_2_step.json"
}

🔹 Copilot Instruction

If instruction = "Run workflow from JSON":

- Call the tool "Run workflow from JSON"
- Pass fileContentBase64 and fileName
- Do not summarize
- Do not generate any response

🔹 Expected Behavior

Copilot should:
➡️ Immediately call the tool
➡️ Pass inputs exactly as received
➡️ NOT generate any text response

🔹 Actual Behavior ❌

Instead of calling the tool, Copilot:

Decodes/interprets the base64 JSON
Parses the workflow
Generates a detailed summary

🔹 What I Tried

Making instructions stricter
Explicit “do not summarize / do not respond”
Using clear condition (instruction = ...)
Verifying tool schema and parameters

Still, the agent prefers generating a response instead of calling the tool.

❓ Questions

How can I force Copilot to directly call a tool when triggered from Power Automate?
Is there a way to make Copilot behave like a pure execution agent (no reasoning/output)?
Is this limitation related to:
- Copilot topic design?
- Tool configuration?
- Or how Power Automate triggers the agent?

🔹 Alternative Approach (Considering)

{
  "command": "EXECUTE_TOOL",
  "tool": "Run workflow from JSON",
  "fileContentBase64": "...",
  "fileName": "workflow_schema_2_step.json"
}

Any suggestions or best practices would be really helpful 🙏

Thanks!

Categories:

Autonomous agents

Calling actions from Copilot Studio

Hi @CU15041523-0,

Copilot Studio summarizes JSON instead of executing the tool
If you are triggering a Copilot Studio agent from Power Automate and seeing that it summarizes or explains your JSON instead of calling the tool, you are not misconfigured. This is a platform design limitation, not a mistake in your setup.

What is actually happening
Copilot Studio is a conversational AI, not a deterministic orchestration engine.

Reasoning-first execution: Every request sent to Copilot (including from Power Automate) goes through an LLM reasoning step. The model decides whether to respond conversationally or call a tool.

No silent / force-execution mode: There is currently no supported way to force Copilot to immediately invoke a tool or suppress text generation.

High-signal JSON encourages summarization: Structured data like JSON strongly nudges the model to explain or summarize it, even when explicitly instructed not to.

“Do not summarize” is advisory, not enforceable: Instruction text influences behavior but does not override the LLM’s decision-making.

This behavior is by design and consistent across Copilot Studio.

What does work (recommended architectures)
Option 1: Reference Key strategy (most reliable)
Do not send the full JSON payload to Copilot.

Store the JSON in Dataverse, SharePoint, or another store.
Send only a reference ID to Copilot.
Inside the tool or flow, retrieve the JSON using that ID.

Why this works:
Copilot sees low-context input and has nothing useful to summarize, so it proceeds to call the tool.

Option 2: Intent-only Copilot
Use Copilot only to decide what should happen, not to execute it.

Ask Copilot whether a workflow should run.
Return a flag (Yes/No or workflow name).
Let Power Automate perform all execution and data handling.

Result: You get AI decision-making without risking JSON interpretation or hallucination.

Option 3: Skip Copilot for execution
If your scenario requires:

Guaranteed execution
Zero conversational output
Deterministic behavior

Then Copilot should not be part of the execution path.
Use Power Automate, child flows, Logic Apps, or Azure Functions directly.

Why command-style payloads do not work
Sending payloads like:
{

"command": "EXECUTE_TOOL",

"tool": "Run workflow from JSON"

}
does not force execution. Copilot treats this as conversational input, not a system directive.

Key takeaway
Copilot Studio should decide what to do, not reliably do it.
Use Copilot as the decision layer and Power Automate as the execution engine.

Treat Copilot as the concierge and Power Automate as the engine.
Don’t ask the concierge to be the engine.

✅ If this answer helped resolve your issue, please mark it as Accepted so it can help others with the same problem.
👍 Feel free to Like the post if you found it useful.