Learn
Let AI agents analyze large data files without downloading them

To let AI agents analyze large data files without downloading them, keep the files in a shared drive that agents can search on the server and read in parts.
The agent runs grep or a full-text search across the files to find what's relevant, reads the header and the rows or lines it needs instead of the whole file, and writes its findings back to the drive as a file the team can check. In Fabric, that works for CSV exports, XLSX spreadsheets and plain text files such as logs, whether the agent connects through MCP, the CLI or the mounted desktop drive.
How people do this today
Many data questions start with an export: a year of orders from the shop, a CRM dump, a week of server logs. The common ways of putting an AI agent on that file each run into the same problem, which is that the file is large and the agent sits somewhere else.
The first option is uploading it to a chat. That's fine for small files, but chat tools set limits. OpenAI's file uploads FAQ currently puts the cap for CSV files and spreadsheets in ChatGPT at roughly 50MB, depending on the size of each row, and a busy year of transaction data passes that quickly. The upload also creates a copy that drifts from the original the moment someone re-exports.
The second is to download the file to a laptop and point a coding agent at it. That works for one person on one machine. When a colleague's agent needs the same file, it gets its own copy, and the two analyses start from different versions.
The third is to paste a sample into the conversation and hope it's representative. Sometimes it is. Often the interesting rows, the refunds or the one error that happened at 3 a.m., aren't in the first two hundred lines.
None of these is wrong for a small, one-off question. They get awkward when the files are big, when several people or agents work on them, or when the question comes back every month.
Which files this works for
Fabric stores and reads several kinds of data file. The table below covers what an agent can do with each, based on what Fabric supports today.
File type | What agents can do in Fabric |
|---|---|
CSV | Search across files on the server; read part of a large file, such as the header and a range of rows; read the full contents through MCP |
XLSX | Fabric's AI and connected agents can read spreadsheet contents |
Text and log files | Search with grep, regular expressions or full-text search on the server; read a section, such as the last part of a log |
JSON | Stored as a document and searchable as text |
Google Sheets | Available through a Google Sheets connection |
For a closer look at spreadsheets in particular, see spreadsheets in Fabric and how to ask AI questions about a spreadsheet. Fabric's own assistant can also answer questions about a folder of files; this page is about agents you run yourself, such as Claude Code, Codex or ChatGPT.
Set up the folder
Start with one folder per dataset. Put the raw exports in a raw subfolder and leave them as they came out of the source system. Agents will be told to read from there and never write to it.
Next to the exports, add a short notes file. It saves every agent from guessing what the columns mean, which is where most analysis mistakes start. Something like this is enough:
Add a results folder for anything the agents produce. Each result gets its own Markdown file with the question, the method and the answer, so a person reading it later can tell how a number was reached. Ask your files, don't build a dashboard makes the case for this habit of keeping raw exports and a notes file side by side.
Connect the agent
There are four routes in, and the choice depends on where the agent runs. Agents in Claude, ChatGPT, Cursor, VS Code or Claude Code can use Fabric's MCP server, which signs in with OAuth and can search the drive and read file contents. Scripts and coding agents can also use the CLI or the API. How AI agents connect to a shared drive has setup details for each, and giving ChatGPT or Claude access to your spreadsheets covers the spreadsheet case.
If a coding agent runs on a computer with the Fabric desktop app, it can also use the mounted drive. Fabric appears as a drive on Mac, Windows and Linux, and files take no disk space and stream when they're opened. The agent reads and writes there as it would in any folder, so it can run its own Python script over a CSV in the shared drive and save the output to results.
Work through the file in parts
A good analysis on a big file usually happens in three moves, and they map onto what Fabric lets agents do.
Look before computing. Ask the agent to read the header and a few hundred rows from different parts of the file. Agents can read part of a large file without fetching the whole thing, so this is quick even on an export that runs to millions of rows. It's enough to check that the columns match the notes file, spot a date format that changed halfway through the year, or find the duplicate rows the notes warned about.
Narrow with search. Before running anything heavy, have the agent search for the rows that matter. Fabric runs grep, regular expressions and full-text search on the server across the whole drive, so an agent can find every line mentioning chargeback across twelve monthly files, or every ERROR with a particular request ID across a week of logs, without downloading them first. Why AI agents give up searching your files explains why server-side search makes such a difference to how agents behave. For logs specifically, see error logs in Fabric.
Compute where the numbers can be checked. For totals, averages and anything else that needs every row, a coding agent should run a script over the file instead of reading rows and adding them up in its head. Agents are good at writing that script and less reliable at doing long arithmetic in text. The script and its output go into results beside the answer, so anyone can rerun it.
Fabric provides the storage, the search and the partial reads. The analysis itself, meaning which columns to join, how to treat refunds and whether a spike is real, comes from the agent and the person directing it. Keep that split in mind when you read the results.
Several agents on one dataset
Larger questions split well across agents. One agent per month of orders, each writing a short summary to results/by-month/, and a final agent that reads those summaries and writes the yearly view. Because the files sit in one shared drive, the final agent sees each monthly summary as soon as it's saved. Run a research project with a hundred agents uses the same pattern on interviews and papers, and AI agents for video production applies it to footage.
If two agents append to the same summary at the same moment, Fabric refuses a save that's based on an older version of the file. The agent re-reads it and reapplies its change, so neither set of findings is lost.
Give the agents only what they need. With Agent Vault, each agent gets read access to raw, write access to results, and no permission to delete. Access can last for a single task and then ends automatically, and every read and write is logged. Agent Vault is on the Pro and Team plans. Even with careful permissions, mistakes happen; deleted files go to a trash you can restore from, and earlier versions of each file are kept.
Questions that come back every month
Some questions repeat: last month's refund rate, the top error types this week, which customers went quiet. For those, Fabric's background agents run on a schedule you set, with access limited to the folders you choose, and write their results as documents in the workspace. Recurring data analysis is one of the uses they're designed for. Background agents are on the Pro and Team plans.
The wider idea behind all of this is in your AI agents need a shared drive, not a group chat: when the data, the notes and the results sit in one folder that every agent can reach, nobody has to pass files between them. If you're starting from scratch, what is a shared drive for AI agents? covers the basics. Fabric prices per person, so agents don't need seats; see the team plans.
Frequently asked questions
Can AI agents analyze a large CSV file without downloading it?
Yes. In Fabric, an agent can search CSV files on the server and read part of a large file, such as the header and a range of rows, without fetching the whole thing.
What file types can agents analyze in Fabric?
CSV and XLSX spreadsheets, plain text files such as logs, and JSON. Google Sheets come in through a connection.
Can ChatGPT or Claude read my spreadsheets in Fabric?
Yes. Once connected through Fabric's MCP server, both can search your Fabric drive and read file contents, including CSV and XLSX files.
How do agents search inside large log files?
Fabric runs grep, regular expressions and full-text search on the server across the whole drive. An agent can find matching lines in a week of logs and then read just the section around them.
Is it accurate to let an AI agent do the math on a big dataset?
It's safer to have the agent write and run a script for totals and averages, and save the script with the result. That way the numbers can be checked and rerun.
Can several agents work on the same dataset at once?
Yes. Each can work on its own part, such as one month of data, and write results to the shared drive. If two save to the same file at once, Fabric refuses the save based on the older version and the agent reapplies its change.
How do I stop an agent from changing the raw data?
With Agent Vault, give the agent read access to the raw data folder and write access only to a results folder. Leave out the permission to delete.
Can I schedule a recurring analysis?
Yes. Background agents in Fabric run on a schedule you choose, with access to the folders you pick, and save their results as documents. They're on the Pro and Team plans.
Related pages
You might be interested in:

How to hand off work between AI agents

Share skills with your whole team through a shared drive

How AI agents connect to a shared drive: MCP, CLI, API or a mounted drive

Let AI agents analyze large data files without downloading them

Build a CRM your agents can use (it's just a folder)

What Is a Shared Drive for AI Agents?

What Is a Virtual Data Room?

How to let clients proof and choose photos online