Can I Build an AI Sales Bot with Dataverse and Copilot Studio?

(1) Share

Report

Posted on by CU03121435-0

Hi Copilot Studio forum,

I’m working on a project to streamline sales processes for my company, and I’d like your input on whether Dataverse and Copilot Studio are the right tools for the job.

Project Overview

I’ve been tasked with developing an AI sales bot to assist our sales team. The goal is to help salespeople quickly navigate complex product configurations during customer calls or email interactions.

The bot should be able to:

Guide the user through product configurations in real-time.
Ask clarifying questions when inputs are incomplete.
Provide accurate configuration options along with exceptions or special notes.

Here’s a sample interaction I envision:

Salesperson: "I have a customer interested in a Citroen without a lift."
Bot: "Which body type are they interested in? The options are: 'open platform' or 'tilt platform without sides.'"
Salesperson: "They want a 'tilt platform without sides.'"
Bot: "Great! The length will be 4000mm, and the width is 1900mm. Note: A special tow hitch is required."

My Setup and Resources

Experience:
- Azure Portal services like AI Search and Logic Apps.
- Powershell
- Excel and Power Query
- No experience with SQL databases or complex coding (I rely on tools like ChatGPT for support).
Resources:
- Working solo without IT team support.
Data:
- Product data is currently in 15 Excel tables (~300 rows and 6 columns each).
- Tables contain configurations, dimensions, and comments.
Timeline:
- Deliver an MVP within 8 weeks, with a full system ready in 10-20 weeks.

My Proposed Solution

Data Storage:
- Normalize Excel files using Power Query and upload them to Dataverse.
- Use Dataverse as the central repository for all product configurations, exceptions, and related data.
Bot Interface:
- Build the bot using Copilot Studio for a conversational interface.
- Connect Copilot Studio to Dataverse for querying product configurations.
Logic and Workflows:
- Use Power Automate to handle data processing, clarifications, and fetching relevant comments from Dataverse.
Exceptions and Comments:
- Store product-specific notes (e.g., "Special tow hitch required") in Dataverse and surface them dynamically in bot responses.

Key Questions

Dataverse Feasibility:
- Can Dataverse handle dynamic queries like "Show all configurations for Citroen without a lift" and return options for further filtering (e.g., body type)?
- Is Dataverse practical for storing and querying this kind of relational data efficiently?
Copilot Studio:
- Can Copilot Studio handle dynamic clarifications, like asking for missing inputs ("Which body type?")?
- How seamlessly does Copilot Studio integrate with Dataverse for querying structured data?
Scalability:
- As more configurations and exceptions are added over time, will this setup scale effectively?
Alternative Approaches:
- Should I consider a hybrid solution (e.g., combine Power Platform with Azure services like AI Search), or stick solely to Power Platform/Copilot Studio?

What I’ve Considered So Far

Azure-Based Solution: Normalize tables with Power Query, store them in Azure Blob Storage, and use Azure Cognitive Search for querying data and using Azure Open AI for NLP chat.
Power Platform Solution: Use Power Query for normalization, Dataverse for storage, and Copilot Studio/Power Automate for logic and interactions.

Given my constraints (limited coding, no SQL, tight timeline), which approach would you recommend? Has anyone here implemented a similar bot using Dataverse and Power Platform? Or would it be cheaper to hire external experts.

Any advice or guidance would be greatly appreciated!

Thanks in advance!

Categories:

General topics

I have the same question (0)

All responses (2)

Answers (0)

Suggested answer

Artur Stepniak 1,539 Super User 2025 Season 2 on at

Like (1)

Report

Hello,

really interesting question, thanks for that! Overall, the analysis that you've performed is very good in my opinion, congratulations. :-) Answering your questions:

1. I think yes, it should work without any special configuration, but that would depend on the prompts that you're going to input - how complex are they etc. Basically when you define Dataverse as a datasource for the bot, it probably uses ODATA query to grab the data and then generate a response based on the data retrieved. You should keep that in mind when prompting - the prompt should be constructed so that the model could understand to which columns it is referencing. You'd probably also need to manipulate the system prompt to achieve desired results.

2. You'd need to test as it mainly depends on the datasource that you're connected to. The best way to check is to create a sample table and see how it performs. It's a new technology, so everyone are testing - it's not possible to give unambiguous answer for now. :-)

3. It should, keep in mind though that a lot is changing on the platform and giving you 100% certainty is impossible.

4. I'd go with Power Platform/Copilot Studio environment as the task is possible within existing low-code environment. I'd avoid Azure as you'd need to perform a more complex setup to prepare the solution.

Overall, the deadlines are too tight given the fact that the products are new and a lot is changing day-by-day. The expectations from the business to prepare a fully functional product, based on not fully stable environment, is not a good approach. I understand though that the pressure is big now.

The choice that you'll make also depends on the data quality. If it's poor, then you'd need to work with the business to prepare a more reliable dataset.

To sum it up, I'd try to prepare MVP by myself and see how it performs. Your lack of coding experience is not a problem as you're going to use low-code solutions. Lack of knowledge about SQL could be a problem in database design, but what helps you is the low-code approach in Dataverse. Then, after you prepare an MVP, you could decide if you'll manage to prepare a fully functional product or if it's better to transfer it to the external party. That's how I'd see it.

In case of any other questions, let me know. If the answer helped you, mark it, so that others can benefit from it.

Best regards,

Artur Stepniak

Was this reply helpful? Yes No
CU03121435-0 4 on at

Like (0)

Report

@Artur Stepniak

Hello again,
Thank you for your detailed response.

I have been experimenting with Copilot Studio, inputting data correctly into Dataverse and configuring instructions to tailor the bot’s behavior. While some queries yield accurate responses which I am very happy about, I frequently encounter situations where the bot does not deliver the expected results despite multiple adjustments. Furthermore, resolving one type of issue often creates unintended problems with other types of prompts.

The nature of my interaction with the bot is primarily to ask questions about possible configurations of products, such as vehicle builds. Ideally, the bot should respond by either asking follow-up questions to narrow down the options or directly providing specific information, such as dimensions or specifications.

Currently, I rely on Dataverse as the knowledge source, which the bot queries to extract relevant information, paired with automatic "conversational boosting" to enhance its generative responses. My primary focus is on configuring the "Overview" and "Instructions" sections to refine the bot’s behavior and responses.

One key issue I’m currently facing is the bot’s regularly mention of irrelevant metadata despite explicit instructions to avoid such behavior. For example, I explicitly instruct: "Never mention creation dates, record counts, or other irrelevant metadata."

Yet, the bot sometimes produces responses like:

"For this product configuration, there are two records for the model [Product Identifier]."

It is vital for my setup that this won't happen. Otherwise I will be unable to use Copilot for my chatbot as this undermines the expected functionality and leads to user frustration. My goal is to ensure that the bot provides concise, relevant answers without introducing unnecessary or incorrect details. If you have any suggestions for addressing these persistent challenges while maintaining functionality for diverse prompts, I would appreciate your guidance. Also the 2000 characters limit is very strict.

Was this reply helpful? Yes No