Four ways to use an AI on a data file, and where the file goes in each
There are four common ways to use an AI on a spreadsheet or a CSV file. They differ in where the model runs, where the file is read, what you need before you start, and what each one returns. This page describes each, so that you can match the tool to the file in front of you.
1. An online AI assistant
You open a chat service in your browser, attach the file, and ask. The model runs on the provider’s servers, and the file is uploaded to them. Many of these assistants write and run a short program on the uploaded copy to compute their answer; others read the file as text.
You need an account, and often a paid plan for larger files. In return, you get large models, fast answers and a written reply, sometimes with the code that was run.
The file goes to the provider. What happens to it there is set by the service’s terms and by your plan, and those differ from one service to the next.
2. A model runner on your own computer
You install an app that downloads open models and runs them on your machine. The model runs locally, and it reads what you give it. To ask questions about a table, you either paste part of it into the chat or connect the model to your file yourself, usually with a short script in Python or SQL.
You need a computer with enough GPU memory, or unified memory on a Mac, for the model. To ask about a whole table, you also need to write or adapt code. You control the model and the code, and you can change the setup as you need. Whether the answer is right depends on the model and on the code you write around it.
The file is read on your computer, as long as the model and your script run there.
3. A tool in the browser
Some web pages load a database engine into the page itself and read your file inside the browser tab. Where the model runs depends on the tool: a few run a small model in the page, and others use a model on a server.
You need a recent browser and, for the tools with a model in the page, a computer strong enough to run it. The tool needs no installation, so you can start at once.
The file is read inside your browser tab. A tool that uses a model on a server also sends your question there, sometimes with a sample of the data, and the tool’s own description says what it sends.
4. A desktop app with the model inside
You install an app that includes its model and its database engine. You open the file in the app and ask in plain language; the model writes a query, the engine runs it on your file, and you get the result with the query that produced it. Totsum Data Explorer works this way: the model runs through llama.cpp and the database engine is DuckDB, both inside the app.
You need the hardware for the model and, when sales open, a license that you pay for one time. The model is about 7 GB, and the installer downloads it one time. Totsum requires a GPU with roughly 12 GB of VRAM, or a Mac with an M4 Pro chip or better (an M3 Max or M3 Ultra also qualifies) and at least 18 GB of unified memory.
You get the rows behind each figure, the query, and a report for the person who checks your work. When Totsum can check its retelling, the query is also retold step by step in plain language. Answers take longer than with an online assistant, often minutes rather than seconds, because the model runs on your own machine.
The file is read on your computer. Totsum can work on a computer with no network connection.
The four side by side
| Kind of tool | Where the model runs | Where the file is read | What you need | What it returns |
|---|---|---|---|---|
| Online assistant | The provider’s servers | The provider’s servers | An account | A written answer, sometimes the code |
| Model runner | Your computer | Your computer, through your own code | Hardware, and code | What your code returns |
| Browser tool | The page or a server, depending on the tool | Your browser | A recent browser | A table or a chart, depending on the tool |
| Desktop app with the model inside | Your computer | Your computer | Hardware | A table or a chart, the rows behind it, the query |
How to choose
If the file is public, an online assistant gives you a larger model and a faster answer.
If you want to start at once without installing anything, a browser tool reads the file inside your browser tab, and its description says where its model runs.
If you want to keep the file on your computer and you write code, a model runner gives you the most control.
If you want to keep the file on your computer and would rather not write code, a desktop app with the model inside lets you ask in plain language and still read the query it ran. It needs the hardware, and each answer takes longer.
Whichever way you choose, an answer can be wrong. A tool that shows the rows and the query behind an answer makes a mistake easier to find. How to ask a spreadsheet questions in plain language describes how to check an answer.
When the file belongs to a client
Whether a client’s file may go to an online service depends on your contract with the client. A contract that forbids uploading the file to a cloud AI may still allow an AI that runs on your own machine.
Totsum Data Explorer is in a private beta, free in its first phase. You can apply for the beta; the free trial opens after it.