Power Platform Community Forum Thread Details

Hi All,

I've been tasked with doing some preliminary R&D on how to best utilise Copilot with our in-house CRM data (which is stored in an AWS MySql database). Any guidance in input would be greatly appreciated.

The general objective would be to have our CRM data (customers, sales history, orders, quotes, project management, notes) available for Copilot to analyse and provide contextually accurate answers/outputs. Some examples of the types of usecases we'd be hoping to achieve:

Provide a list of customers in New York, with between 20-100 employees, where we haven't contacted them in the last 2 years.
[Copilot provides excel/csv/table]
Find the Installation Date for customer "Bobs Tires" for their pending AWS Server Project.
[Copilot returns the corresponding installation date attribute from our Project Management module for the given customer and project]
Provide a list of our top 10 customers by Gross Profit for the 2023 Calendar year.
[Copilot provides excel/csv/table]
Which Sales Rep has sold the most of Product X.
Which sales order is associated with Reference ''14VSC2S"
etc...

I good start point would be.... is this even possible in the way I've described? What barriers are we likely to face if trying to develop a bot/GPT that is able to understand and interrogate our data structure in such a way?

Some more specific questions:

We already use the PowerBI gateway service to get data into Fabric, which is then used to power multiple dashboards and reports. My presumption is that there will be some type of data pipeline we can setup to get our raw data into the Azure stack for Copilot to index. What is the best way to maintain this data? Is it a periodic/scheduled refresh? How does this data get indexed? I guess my question is relating to "how do we keep the data up-to-date", and is this even practical?
I have reasonable experience in ETL and structuring import queries. How does one go about maintaining/training the Copilot models semantic interpretation? Is this best achieved through logical renaming of database fields?

I'm sure more things will come to mind, but thought this was a reasonable starting point.
As I said earlier, any feedback would be greatly appreciated. 😀

Categories:

General topics

Hey there @Samwell89 ,

The RAG patterns implemented currently in Copilot Studio work over unstructured data sources (e.g. documents stored in SharePoint/the copilot itself), so your use case will require a different approach.

This is only a high-level description of the design, but you would have to: 1. convert a natural language query to a query syntax recognized by your CRM (e.g. a SQL Statement). Leverage an LLM, such as Azure OpenAI to do that, 2. Run the query 3. Convert the results back to natural language (again, using an LLM)

All of that can be done during runtime (e.g. with the Power Automate SQL connector and Azure OpenAI) - no need to set up data pipelines.

I think there are some tentative plans to support Gen Answers over structured data sources, but I don't think anything has been confirmed. In the meantime, you could try the custom approach above.