I Built an AI-Powered Excel Analyst Because Spreadsheets Still Take Too Much Manual Work

An Excel file can contain thousands of rows of useful information.

But getting from that file to a meaningful answer can still take hours.

You open the spreadsheet, inspect the columns, look for missing values, clean inconsistent data, calculate KPIs, create charts, investigate trends, and finally turn everything into something another person can actually understand.

The analysis itself is often not the hardest part.

The repetitive preparation around it is.

That was the problem I wanted to explore when I built Excel Auto-Analyst, a Streamlit-based application for analyzing Excel and CSV datasets.

Live app: https://excel-auto-analyst-ne9ocshgvqtvqtitbapbjs.streamlit.app/

Source code: https://github.com/Shweta-Mishra-ai/excel-auto-analyst

The problem

Data analysis workflows often start with something deceptively simple:

«“Here is the Excel file. Find out what is happening.”»

In practice, that request expands quickly.

Before answering a single business question, an analyst may need to:

• Understand the dataset structure
• Identify missing values
• Detect problematic columns
• Check data types
• Clean the dataset
• Calculate descriptive statistics
• Identify useful KPIs
• Explore relationships between variables
• Build visualizations
• Interpret trends
• Prepare a report

None of these tasks are individually complicated.

The problem is repetition.

When the same workflow happens across dozens of spreadsheets, a significant amount of analyst time can disappear into tasks that are necessary but not particularly insightful.

I wanted to build something that could automate more of that workflow while keeping the analyst in control.

What Excel Auto-Analyst does

The application starts with a simple input:

Upload an Excel or CSV file.

From there, the workflow moves through several stages.

1. Data profiling

The first step is understanding what was uploaded.

Instead of immediately generating charts, the application first looks at the structure of the dataset.

This includes information such as:

• Number of rows
• Number of columns
• Data types
• Missing values
• Duplicate records
• Basic dataset characteristics

This matters because analysis built on a misunderstood dataset can produce misleading conclusions.

Profiling should happen before interpretation.

1. Data cleaning

Real-world spreadsheets are rarely perfect.

Columns may contain missing values, inconsistent formats, duplicated records, or values that require preprocessing before analysis.

Excel Auto-Analyst includes a cleaning stage so the user can prepare the dataset before moving into deeper analysis.

An important part of this workflow is maintaining visibility into what happens to the data rather than treating cleaning as an invisible operation.

That distinction matters.

Automation should not mean:

«“Trust the black box.”»

It should mean:

«“Here is what happened to your data, and here is what you can inspect.”»

1. KPI and statistical analysis

Once the dataset is prepared, the next question becomes:

What does the data actually tell us?

The application moves beyond simply displaying rows and columns and provides analytical outputs such as KPIs, statistics, and patterns that can help users understand their dataset.

This is where spreadsheet analysis starts becoming more useful.

Instead of manually writing the same calculations every time, the application can provide a starting point for exploration.

The goal isn't to replace statistical reasoning.

The goal is to reduce the repetitive work required to reach the point where reasoning becomes useful.

1. Visual exploration

Numbers are useful.

But patterns often become much easier to see when they are visualized.

Excel Auto-Analyst generates interactive visualizations to help users explore the dataset from different perspectives.

Charts can make things such as:

• Category differences
• Trends
• Distributions
• Relationships
• Outliers

much easier to inspect.

This is particularly useful during the exploratory phase of analysis, where the objective isn't necessarily to prove a hypothesis yet.

Sometimes you first need to discover what questions the dataset is capable of answering.

1. AI-assisted insights

This is the part where the application moves from traditional analytics toward an AI-assisted workflow.

Instead of requiring the user to manually interpret every visualization, the application can generate AI-assisted observations from the analysis.

The important distinction is that AI should sit after the data and analytical processing, not blindly replace them.

A useful architecture is:

Data → Profiling → Cleaning → Analysis → Visualization → AI interpretation

rather than:

Data → LLM → Answer

That difference is important.

If an LLM receives raw spreadsheet data and is simply asked to “analyze it,” the system has much less control over how calculations are performed and how conclusions are produced.

A structured analytical pipeline provides more opportunities for validation and inspection.

The architecture

The project is built around a relatively straightforward pipeline:

Excel / CSV
↓
File Upload
↓
Data Profiling
↓
Data Cleaning
↓
Statistical / KPI Analysis
↓
Visualization
↓
AI-Assisted Insights
↓
Report Export

The interface is implemented with Streamlit, which makes it possible to expose the analytical workflow through a browser without requiring users to build a local dashboard themselves.

Python handles the underlying data-processing and analytical workflow.

The architecture is intentionally focused on the workflow rather than creating another generic chatbot around a spreadsheet.

Why I didn't want another “chat with your CSV” app

There are already many applications where you can upload a dataset and ask an LLM questions.

That approach is useful, but I wanted to explore something slightly different.

A data analyst doesn't only ask questions.

They also need to know:

• What is in the dataset?
• Is the data clean?
• Are there missing values?
• Which variables matter?
• What metrics are meaningful?
• Which patterns are actually present?
• Can the result be visualized?
• Can the analysis be reproduced or communicated?

That is why I approached Excel Auto-Analyst as an analysis workflow, rather than simply a conversational interface.

The AI layer is one component.

It isn't the entire system.

What I learned building it

One of the biggest lessons was that automating analysis is harder than generating charts.

A chart can be produced quickly.

A useful analytical workflow requires much more consideration.

You need to think about data quality, assumptions, edge cases, user experience, and how the output will be interpreted.

There is also a fundamental difference between:

“The system generated an answer.”

and

“The system generated an answer that an analyst can inspect and reason about.”

For AI-assisted data analysis, I think the second goal is much more important.

Where this could go next

There are several directions I want to explore further.

For example:

• More advanced anomaly detection
• Better natural-language data exploration
• Stronger validation of AI-generated insights
• More analytical templates
• Larger dataset support
• Reproducible analysis pipelines
• More export formats
• Better evaluation of generated insights

The interesting problem isn't simply making AI produce more analysis.

It's making automated analysis more reliable, inspectable, and useful.

Try it

If you work with Excel, Python, analytics, or data science, you can try the application here:

Live app:
https://excel-auto-analyst-ne9ocshgvqtvqtitbapbjs.streamlit.app/

GitHub:
https://github.com/Shweta-Mishra-ai/excel-auto-analyst

The project is open source, and I'm interested in feedback from people who work with real datasets rather than only toy examples.

The question I'm exploring is simple:

How much of the repetitive spreadsheet analysis workflow can we automate without taking the analyst out of the loop?

What part of your own data-analysis workflow would you automate first?