Back office ops · Production
Building a ChatGPT-powered business analytics assistant using multi-agent ReAct workflows
The problem
LLMs including ChatGPT are unreliable for quantitative reasoning on structured data, yet most critical business information resides in SQL databases, making analytics automation challenging.
Workflow diagram · grounded in source
1
User poses analytical question
trigger
“The objective is to allow users to ask complex analytical questions of business data”
2
Data Scientist agent plans approach
ai_action
“the agent plans how to solve an input question. For non-trivial problems, agents might require multiple intermediate analysis steps leading to unanticipated yet advanced outcomes. Observations gained during these intermediate steps may c…”
3
Data Engineer retrieves SQL data
integration
“This agent is responsible for performing data acquisition from the source system (in this case a SQL database). The data engineer agent receives instructions from the data scientist agent.”
4
Data Scientist performs analytics
ai_action
“This stage encompasses the performance of everything from simple computation such as aggregations to statistical analysis and Machine Learning”
5
Results presented with visualizations
output
“ChatGPT to deliver the answers in the best possible format, complete with rich visualizations that make it easier for users to comprehend the results”
Reported outcome
The article presents a methodology and reference implementation that turns ChatGPT into a business analytics assistant, enabling users to query structured data and receive visualized answers without advanced technical skills.
Reported stack
ChatGPTStreamlitPlotlyPython ConsoleSQLiteSQL Server
Source
https://medium.com/data-science-at-microsoft/automating-data-analytics-with-chatgpt-827a51eaa2c
Read source ↗Frequently asked questions
What did this team achieve with this AI workflow?
The article presents a methodology and reference implementation that turns ChatGPT into a business analytics assistant, enabling users to query structured data and receive visualized answers without advanced technical…
What tools did this team use?
ChatGPT, Streamlit, Plotly, Python Console, SQLite, SQL Server.
How is this back office ops AI workflow structured?
User poses analytical question → Data Scientist agent plans approach → Data Engineer retrieves SQL data → Data Scientist performs analytics → Results presented with visualizations.