Use copilot to analyze desktop flow activity (preview)

Article
11/14/2023

[This article is prerelease documentation and is subject to change.]

Understanding automation performance is key to achieving operational excellence and reliability goals, regardless of the size of the automation estate, team or role within the organization. To reach those goals requires advanced and dynamic monitoring capabilities that provide you with valuable insights that highlight areas of success and identify potential bottlenecks, trends and areas for improvement. Having more detailed insights allows you to make informed decisions that optimize your automation processes, leading to increased efficiency and effectiveness.

Important

This is a preview feature.
Preview features aren’t meant for production use and may have restricted functionality. These features are available before an official release so that customers can get early access and provide feedback.

The latest advancements in AI provide us with unprecedented opportunities to explore new automation health monitoring use-cases that could include anything from simple data exploration to anomaly detection, smart recommendations, and even self-healing bots.

With copilot now able to analyze desktop flow activity we're taking the first step in a new direction, allowing you to democratize access to insights by asking copilot desktop flow activity-specific questions using natural language.

Important

This capability is powered by Azure OpenAI Service.
Copilot is a new technology that is still being developed. It is optimized for use with English language and has limited support with other languages. As such, parts of it might appear in English rather than your preferred language.
Read the responsible AI FAQs for copilot in desktop flow activity (preview) to learn more about this new copilot experience.
More FAQs: Responsible AI FAQs for Power Automate,FAQ for copilot data security and privacy in Microsoft Power Platform

Prerequisites

A work or school account with access to a Power Automate environment that's based in the United States.
During initial preview, you must have an environment in the United States region to use this feature. If you don’t have access to an environment that's based in the United States, you can ask your administrator to create a new environment in Power Platform admin center and select United States as its region.
Check known limitations for more information.

How does it work?

This copilot experience is powered by the Azure Open AI service and is capable of translating user prompts into valid Dataverse FetchXML queries. Initially, these queries are focused on and optimized for desktop flow activity, such as runs, flows, errors, and machines.

High-level process

Once the user inputs a valid prompt, copilot generates a valid FetchXML query based on the input.
If the generated FetchXML is valid, the query is then executed against the Dataverse backend under the current user's security context to retrieve matching data. This ensures that users only see data that they're already authorized to access.
Copilot then determines the most suitable output visualization, such as a table, pie chart, bar chart, or line chart, to effectively present the insights and data to the user.

What are FetchXML queries?

Microsoft Dataverse FetchXML is a language used for retrieving data from a Dataverse database. It's designed to be easy to create, use and understand. For example, you might want to ask Dataverse to give you a list of all flow runs for a specific flow. The FetchXML query is the way you phrase that question so the database understands it and can give you the right results.

Prompting best-practices

Be specific: The more specific you are with your prompt, the better the AI will understand and respond. If the AI isn't producing the desired output, don't worry, try again by adjusting your prompt.
Experiment with prompts: If you're not getting the results you were expecting, try rephrasing your prompt or provide more context.
Provide feedback: If the AI produced great or unsatisfactory responses, let us know by selecting the thumbs up or down with an option to provide more feedback via the Tell Microsoft what you liked about this feature link that appears underneath.

Prompt examples

Examples of prompts that can be used as starter prompt for your own use-cases are explained in this section. Some of these prompts might not be applicable or return incorrect results, since the accuracy might be influenced by model understanding or the actual prompt and the data available to you based on your permissions. We recommend that you review and validate the returned results and FetchXML query. More information: Validate FetchXML query results generated by copilot.

Runs

Which flows ran the most last week?
What were yesterday’s top five flows by number of completed runs?
What was the average run duration of the '[insert your flow name here]' flow during last semester?

Errors

Show me the most frequent run errors during last month.
Show me a distribution of successful versus failed flows during last quarter.
What were the number of failed runs during the week prior to the last one?

Machines

Which bots had the most run failures today?
Which machines are in maintenance mode?
What are the machines with the most run failures?

Makers

Show me the top flows by number of runs together with their owner info.
Who were the top 10 users running flows during last month?
When and by whom were desktop flows modified last week?

Multi-turn prompts

In the context of AI, multi-turn prompts allow you to have an ongoing conversation with the copilot, where it remembers the context of the previous messages in the conversation. It's not just answering one-off questions; it's engaging in a dialogue with you, where each response is based on what's been said before.

Note

When engaging in a multi-turn conversations, note that copilot keeps track of the five most recent questions only. This means that copilot starts clearing the prompts that were entered first and only keeps the latest five. To improve response quality, we suggest limiting your follow-up questions to four and then restart the chat. More information: Clearing previous prompt context to start over.

Example

Turn	Prompt and reply
	User: show me a distribution of successful vs failed flows during last quarter
	Copilot: Here's the distribution of successful vs failed flows during the last quarter.
	User: what was the top error of those that failed?
	Copilot: Here's the top error of those that failed.
	User: on which machine names did they fail the most?
	Copilot: Here are the machine names where the most failures occurred.
	User: of those that succeeded what were their average run duration?
	Copilot: Here's the average run duration of the flows that succeeded.

Influencing the output format

You can influence copilot's output format by asking for explicit output types like "show me failed vs. succeeded flow run distribution as a bar chart." This likely produces the following outcome:

Clearing previous prompt context to start over

If you wish to reset the conversation with copilot you can select the three dots ... next to the copilot name, and then select New chat.

Validate FetchXML query results generated by copilot

The following steps guide you through the process to validate (and potentially reuse) FetchXML queries in Power Automate cloud flows.

Step 1: Make a copy of the FetchXML query

After submitting your query to the copilot, you get a reply that includes a link labeled Show code. Select this link and then select the copy icon located in the upper right corner of the FetchXML box to copy the code.

Step 2: Create cloud flow and test FetchXML query

Navigate to the Power Automate portal and select My flows from the left-navigation menu.
Continue by selecting + New flow on the command bar, and then select Instant cloud flow from the dropdown menu.
Enter a flow name, select Manually trigger a flow, and then select Create.
The cloud flow designer appears. Find and then select the + New Step button.
On the search bar that appears, enter Dataverse, and then select the Dataverse connector from the results.
Various actions are displayed. Scroll through until you find and select the List rows action.
Within the List rows action, select the Show advanced options link.
A FetchXML query field appears. This is where you input the copied FetchXML query that copilot previously generated.
After pasting in your FetchXML, select Save.
Test your flow by selecting Test.
Follow the prompts on your screen to start your flow manually to review its results.

Step 3: Understanding the results

Let's assume you asked the copilot 'how many failed vs succeeded flows did we have last month?' This produces a FetchXML query similar to the following:

<fetch version="1.0" mapping="logical" aggregate="true" count="3" page="1">
    <entity name="flowsession">
        <attribute name="flowsessionid" alias="flowsession_count" aggregate="count" />
        <attribute name="statuscode" alias="flowsession_statuscode" groupby="true" />
        <filter type="and">
            <condition attribute="completedon" operator="last-x-months" value="1" />
        </filter>
    </entity>
</fetch>

If data matches the given FetchXML query, the List rows Dataverse action configured in step 2 returns data in a format called JSON (JavaScript Object Notation), which is essentially a method used to present data in a well-organized manner, making it easy to read and write digitally.

For distribution-based questions like previously mentioned, data is grouped by one or more fields (statuscode), together with an aggregation (count) that returns the number for each group (that is, failed, succeeded, and so on).

Each of the returned records contains fields such as:

flowsession_count: The number of times the workflow ran.
flowsession_regardingobjectid: The unique identifier for the flow run.
flowsession_statuscode: The status of the flow run (for example, Failed).
workflow_name: The name of the flow.

If you want to know how many times a specific flow ran, look at the flowsession_count column of the record where workflow_name is your flow name.

Understanding copilot replies on problematic prompts

This table shows default responses that are returned when the copilot is unable to understand your question, intent, or generate a valid answer.

Copilot reply	Details
Sorry, something went wrong. Please try again.	Indicates that an unexpected error occurred. Rephrase your question and try again.
Sorry, I couldn’t understand your question. Please rephrase it and try again. I’m able to answer questions that are about the data on this page. For more examples of prompts that you can ask the copilot, you can visit the prompt example section on our documentation page.	Indicates that your question couldn't be translated into a valid FetchXML query. Rephrase your question and try again.
Sorry, Copilot is at capacity and temporarily unavailable — please try again in a little while.	Indicates there are resource constraints on the backend. Retry your question after a short time.
Sorry, your message contains potentially harmful content. Please ensure your input is appropriate and try again.	Indicates that your question might include potentially harmful content and has been blocked by the backend service. Remove any potentially harmful content from your question and try again.
Sorry, I was not able to generate a valid answer based on your question. Please rephrase it and try again. I’m able to answer questions that are about the data on this page. For more examples of prompts that you can ask the copilot, you can visit the prompt example section on our documentation page.	Indicates that the generated FetchXML is invalid or that the query failed when copilot tried to execute it. Rephrase your question and try again.
Sorry, your search includes too many results. Please refine your query and try again. For examples on how to limit search results returned by the copilot, visit our documentation page.	Indicates that the filter(s) applied to your query exceed current aggregation limits in FetchXML. Add more appropriate filters such as asking for yesterday's or last month's data to your query to ensure that it returns data within those limits.

Known issues and limitations

The following list contains known limitations of the copilot in desktop flow activity.

Copilot is a new technology that is still being developed. It's optimized for use with English language and has limited support with other languages. As such, parts of it might appear in English rather than your preferred language.
Copilot is currently only available in Dataverse environments based in the United States.
Copilot might return wrong or incomplete data and FetchXML queries.
Copilot is initially only capable to answer questions about desktop flow activity such as errors, machines, and past and current runs.
In multi-turn conversations, copilot keeps context of the last five question only. If you encounter wrong or incomplete results, consider resetting the conversation. More information: Clearing previous prompt context to start over.
For queries that return large result-sets, copilot might not be able return or render these.

Share via