部落格

Ask your files, don't build a dashboard



For most one-off questions about your business, a folder of raw exports and a short notes file will get you a better answer than a new dashboard. Here's how to set that folder up, and where dashboards still earn their place.



When you have a question about your numbers, the fastest route to an answer is often to put the raw exports in one folder, write a short notes file explaining them, and ask an AI agent. The agent reads the notes, searches and reads the exports, and answers the question you have today. A dashboard answers the questions someone predicted you'd have when it was built, and those are frequently not the same.

This is an opinion, and it has limits, which are covered further down. For a small team, a founder or an operations lead without a data team, it's the better default.


The dashboard that answered last month's question

Picture the operations lead at a small online shop. In October, refunds go up. She asks for a dashboard so she can see why. Someone connects the store's data to a reporting tool, builds charts for refunds by week, by product and by country, and shares it two weeks later.

By then the question has moved on. Refunds by product show one item is the main source, but what she wants to know now is whether those orders came from the new ad campaign, whether they shipped from the second warehouse, and whether the customers mentioned sizing in their refund reasons. None of that is on the dashboard. Each answer means another round of requests, and the dashboard joins the list of tabs nobody opens.

The data needed to answer all three questions was in the exports she already had: orders, refunds with free-text reasons, and a spreadsheet of ad spend by campaign.


What to keep instead

The setup is a folder. Put the raw exports in it as they come out of each system, CSV or XLSX, with the date in the file name. Don't clean them first, and don't merge them into one master sheet. Then add one Markdown file that explains them.

# notes.md

orders-2026-10.csv: one row per order. order_total is in cents.
refunds-2026-10.csv: refund_date is when the refund was issued,
not when the order was placed. Join to orders on order_id.
reason is free text typed by customers.
ad-spend-q4.xlsx: one tab per campaign. "Autumn" launched 6 Oct.
Warehouse B started shipping on 1 Oct. Orders from it have
fulfillment_location = "B".
The 3 Oct orders export was missing EU orders; re-exported as
orders-2026-10-eu-fix.csv.
# notes.md

orders-2026-10.csv: one row per order. order_total is in cents.
refunds-2026-10.csv: refund_date is when the refund was issued,
not when the order was placed. Join to orders on order_id.
reason is free text typed by customers.
ad-spend-q4.xlsx: one tab per campaign. "Autumn" launched 6 Oct.
Warehouse B started shipping on 1 Oct. Orders from it have
fulfillment_location = "B".
The 3 Oct orders export was missing EU orders; re-exported as
orders-2026-10-eu-fix.csv.
# notes.md

orders-2026-10.csv: one row per order. order_total is in cents.
refunds-2026-10.csv: refund_date is when the refund was issued,
not when the order was placed. Join to orders on order_id.
reason is free text typed by customers.
ad-spend-q4.xlsx: one tab per campaign. "Autumn" launched 6 Oct.
Warehouse B started shipping on 1 Oct. Orders from it have
fulfillment_location = "B".
The 3 Oct orders export was missing EU orders; re-exported as
orders-2026-10-eu-fix.csv.

That file is the important part. Raw columns are ambiguous. An agent seeing order_total of 4999 can't know whether that's $4,999 or $49.99, and it has no way to learn that one export was incomplete. Ten lines of notes remove the guesses an agent would otherwise make silently. They also help the next person who opens the folder, which is a reason to write them even without agents.

Over time, add a line whenever you learn something about the data. When a question turns up a quirk ("refunds before 2 Oct use the old reason codes"), put it in the notes, so the next answer starts from it.


How an agent answers

In Fabric, that folder is available to whichever agent you prefer. Claude, ChatGPT, Cursor and coding agents such as Claude Code and Codex connect through Fabric's MCP server, and an agent on your computer can open the folder on the mounted desktop drive like any other. How AI agents connect to a shared drive covers the options.

Ask the refund question and a good agent works in a recognizable order. It reads notes.md. It searches the refunds file for "size" and "fit", which Fabric runs on its side across the files, so the agent doesn't have to download them to search. It reads the rows that match, joins them to orders on order_id, checks which ones shipped from warehouse B, and compares the order dates to the campaign start. Large exports can be read in parts, so an agent that needs only part of a big export doesn't fetch the whole file. Why AI agents give up searching your files explains why that matters for speed, and let AI agents analyze large data files without downloading them goes further into big files.

You can also skip the external agent. Open the folder in Fabric and open the AI assistant; it's limited to that folder automatically and works from what's in it, including CSV and XLSX files. Its answers cite their sources, and you can save an answer as a note in the same folder. How to ask AI questions about a spreadsheet walks through it.

If you want a chart, ask an agent that can run code, and it can usually draw one in its own environment from the same files. Save the image and a two-line explanation back into the folder. The chart then sits next to the data and the notes it came from, so anyone can see how it was made.


Keep the answers, too

The folder improves as you use it. Save each useful answer as a short note: the question, the answer, the files and filters it used. A month later, "why did refunds go up in October?" has a written answer that anyone, or any agent, can find with a search.

When the same question comes up every week, that's a sign it deserves something more regular. Fabric's background agents, on the Pro and Team plans, run on a schedule you set, with access to the folders you choose, and write their results as documents in your workspace. A Monday agent that reads the new exports and writes a one-page summary of refunds, with the notes file as its guide, covers what many small teams build a dashboard for.


Where dashboards still earn their place

None of this argues for deleting the dashboards you already rely on. Some jobs suit them better.

Situation

Better fit

A question you'll ask once or twice

Ask the files

A question that changes as you learn

Ask the files

A number many people check every day

Dashboard

Operational monitoring with alerts

Dashboard

One agreed definition of a metric for the whole company

Dashboard or a governed report

Data too large or too live for exports

Your warehouse and its BI tools

A weekly summary for a small team

Scheduled agent writing to a folder

A dashboard is good at showing the same number the same way to many people, every day, without anyone asking. That consistency matters for things like revenue reported to a board, where two people getting slightly different answers from two different prompts would be a problem.

Alerts on live operational data, such as an error rate or stock running out, need a system watching continuously, which a folder of exports doesn't provide. And when data lives in a warehouse at a scale where exports stop being practical, the warehouse and its reporting tools are the right home for the numbers. Your data warehouse handles the numbers makes the case for where structured data belongs, and do you need a database or a library? draws a similar line.

Agents also make mistakes with numbers. They can misread a column, drop rows in a join, or answer confidently from a sample. Ask for the working: which files, which filters, how many rows. Spot-check one figure against the raw export. The notes file reduces these errors, and checking the working catches most of the rest, but an unchecked number from an agent deserves the same caution as an unchecked number from anyone.


Setting it up in Fabric

Create one folder per area you ask questions about, such as sales, refunds or marketing spend. Drop exports in as you pull them, or add them from your phone with the share sheet. Write notes.md. If several agents will read the folder, give them read-only access to it with Agent Vault on the Pro and Team plans, and a separate folder they can write answers into. If two agents update notes.md at once, Fabric refuses the edit based on an outdated version and the agent reapplies it, so no one's addition disappears. A wrong delete can be undone from the trash, and earlier versions are kept.

The same pattern works for more than numbers. Build a CRM your agents can use applies it to customer records, and your agents, across all your devices shows how the folder you set up on a laptop is the one Claude reads on your phone.

For the broader idea of a team and its agents working from one set of files, read your AI agents need a shared drive, not a group chat and what is a shared drive for AI agents?


Frequently asked questions

Can AI analyze CSV files without a dashboard?

Yes. An AI agent can read CSV and XLSX exports directly, search them, and answer questions about them. A short notes file explaining the columns makes the answers more reliable.

What should go in a data notes file?

What each file contains, what the ambiguous columns mean (units, date definitions, join keys), known gaps or bad exports, and dates of changes such as a new campaign or warehouse.

Should I clean my data before giving it to an AI agent?

Keep the raw exports as they are and describe their quirks in the notes file. Cleaned copies are fine as extra files, but the originals are what you check answers against.

Can ChatGPT or Claude read spreadsheets stored in Fabric?

Yes. Connected through Fabric's MCP server, they can search Fabric and read file contents, including spreadsheets. Fabric's own AI assistant can also read CSV and XLSX files in a folder.

When is a dashboard better than asking an AI?

When many people need the same number every day, when you need alerts on live data, or when the whole company needs one agreed definition of a metric.

Can an AI agent read a very large CSV?

In Fabric, agents can read part of a large file, such as part of a large CSV, without downloading the whole thing, and search across files on the server.

How do I check that an AI's numbers are right?

Ask it to show which files, filters and row counts it used, then compare one figure with the raw export yourself.

Can I get a regular report without building a dashboard?

Yes. Fabric's background agents, on the Pro and Team plans, can run on a schedule, read the latest exports in a folder and write a summary document.


Related pages


The workspace that thinks with you.

Ready when you are.

The workspace that thinks with you.

Ready when you are.

The workspace that thinks with you.

Ready when you are.