Analyze data · Recorded September 28, 2026

A small dataset.
A checkable answer.

Pi Agent analyzed 12 synthetic orders, separated refunded rows and produced a report with reproducible calculations.

Actual run · Synthetic input
This tests one local CLI task. Pi Agent needs installation, a configured model provider and permission to run Python. It is not a test of a browser-based spreadsheet product or a general data-analysis benchmark.

The task

Calculate paid revenue, refunded order value, paid-order average and revenue by channel. Exclude refunded orders from paid revenue. Use code for the arithmetic and avoid unsupported business conclusions.

The input is synthetic and contains no personal or customer data. Currency units are unspecified.

order_id,channel,amount,status
A001,Organic,120,paid
A002,Email,80,paid
A003,Social,65,paid
A004,Organic,150,paid
A005,Email,40,refunded
A006,Social,90,paid
A007,Organic,110,paid
A008,Email,100,paid
A009,Social,75,refunded
A010,Organic,130,paid
A011,Email,60,paid
A012,Social,50,paid
Download the original CSV

What the agent delivered

MetricResult
Paid orders10
Paid revenue955
Refunded order value115
Paid-order average95.5
Organic / Email / Social510 / 240 / 205

The output contains a Markdown report and a Python script using only the standard library. It gives exact formulas and a reconciliation: 955 paid + 115 refunded = 1,070 total row value.

Download the agent’s original report · Download the calculation script

How we checked it

We ran the generated script and separately recomputed totals from the original CSV with decimal arithmetic. We checked both refunded order IDs, all three channel totals and the paid-order count. The calculations matched.

Run configuration: Pi Agent with the configured deepseek/deepseek-v4.1-flash model through its provider adapter; local read, shell and file-write tools enabled.

Human work and limitations

We supplied a deliberately small, clean dataset and an explicit definition of paid revenue, then reviewed the generated files. No output edits were needed for the arithmetic. This run does not test messy spreadsheets, ambiguous column meanings, private data handling or large files. The execution was supervised by the site-building assistant, not a separate human research panel.

Cost: the configured model-provider account was used; a reliable per-run billed amount was not available. No cost estimate is presented as a measurement.

Try the workflow

  1. Prepare a small CSV and define exactly which rows should count.
  2. Ask the agent to produce a report and the calculation code.
  3. Run the code in a safe local environment.
  4. Independently check a few totals before acting on the report.

Pi Agent official website ↗ · Compare spreadsheet analysis products →