An example of how I use AI

From Income Statements
to Sankey Diagrams

Jaeyoon Yu

August 27, 2026

How people use AI today

Dot chart showing the estimated level of AI use across the global population

Each dot = 3 million people.
8.3 billion people in total.

Never used generative AI — 71% Uses free AI chatbots — 28% Uses AI coding agents — 0.36%

Income Statement

Nebius Group N.V.Q2 2026

Unaudited condensed consolidated statement of operations · USD millions

Three months ended June 3020252026
Revenues105.1582.3
Operating costs and expenses
Cost of revenues(30.1)(133.6)
Product development(42.8)(191.0)
Sales, general and administrative(68.2)(173.9)
Depreciation and amortization(75.2)(259.7)
Total operating costs and expenses(216.3)(758.2)
Loss from operations(111.2)(175.9)
Interest income3.624.1
Interest expense(4.8)(119.1)
Gain from revaluation of investments597.4—
Income / (loss) from equity method investments(6.3)12.7
Other income, net24.681.5
Net income / (loss) before income taxes503.3(176.7)
Income tax expense(0.8)(13.7)
Net income / (loss) from continuing operations502.5(190.4)

Source: Nebius Q2 2026 earnings release.

Sankey Diagram for an Income Statement

App Economy Insights income statement Sankey diagram

This is what I kept seeing: an income statement, drawn as a flow.

A Sankey diagram shows where the money comes from, where it goes, and what remains.

Full access costs about $200 per year — and I still cannot choose the companies or figures I want.

Can I generate this for any company, any year, and any course?

Claude Code and Codex work directly with the project files.

CClaude Code
$ claude
> Read Direction.md and inspect this folder.

✓ Read 8 examples in Input/
✓ Built the data pipeline
✓ Wrote diagrams to Output/

A coding agent can see the instructions, examples, scripts, and results together.

</>Codex
$ codex
> Build the Sankey explorer locally.

✓ Pulled SEC XBRL data
✓ Validated the flows
✓ Created the HTML explorer

The work stays in a local folder that I can inspect, revise, and rerun.

Knowledge of Python and HTML would be helpful.
It is not required.

Claude Code and Codex vs. web-based chats

Web-based chat

Files→Web chat→Copy /
download
  • Inputs are attached or pasted into a conversation.
  • Outputs live in the response until I copy or download them.
  • My control over the working files is limited.

Claude Code or Codex

Input/
folder→Claude Code
or Codex→Output/
folder
  • They work with the local files I choose as inputs.
  • Scripts, data, figures, and pages stay as inspectable outputs.
  • I can revise the process and rerun it.

The benefit: much more control over the inputs and outputs.

I started with only two things

9-IS-to-Sankey/
├── Input/ — example Sankey images
└── Direction.md — what I wanted built

Markdown (.md) is a simple plain-text file for writing instructions, using familiar headings and lists.

Then the project grew into this

9-IS-to-Sankey/
├── Input/ — example Sankey images
├── Output/ — diagrams and interactive explorer
├── Direction.md — data and diagram instructions
└── Slide-generation.md — instructions for this deck

Step 1 — Collect examples

I saved eight Sankey examples in \Input, because showing a format is faster than describing it.

Inside the Input folder — descriptive filenames

Input/ — 8 files
├── NBIS-26Q2.png
├── CBRS-26Q2.png
├── DIS-26Q3.png
├── GOOGL-26Q1.png
├── PLTR-26Q2.png
├── UBER-26Q2.png
├── LLY-26Q2.png
└── AMZN-24Q3.png

Step 3 — Direction.md contains this instruction

Consider the following companies: COSTCO, WALMART, and TARGET.
Do this for 2018–2025.
I would like to generate Sankey diagrams like the examples in the \Input folder.
Create one diagram for each company and save it in the \Output folder.
Then create one HTML file in the root folder that lets me navigate the diagrams.
The HTML file should include an animation showing how the Sankey diagram has changed
  for the selected ticker.
Include three separate line charts for total revenue, operating income, and net income.
Include three separate line charts for gross, operating, and net profit margins.
Include tables of the collected data for me to review.
Include links to the inline-XBRL 10-K filings.

Behind the scenes, the agent retrieves XBRL data for COSTCO, WALMART, and TARGET from SEC EDGAR, then turns it into Sankey diagrams.

SEC EDGAR10-K filings </> XBRL data3 companies · 2018–2025 Sankey diagramsone per company-year
Generated Costco FY2025 Sankey diagram

Out came 24 diagrams — 3 companies × 8 years — every one built from the 10-K XBRL.

Costco’s Sankey changes with the business.

Animated sequence of Costco income statement Sankey diagrams from fiscal 2018 through 2025 FY2018
FY2018

Use the slider to inspect a year, or press Play to animate the full series.

The same Costco data also produces additional figures.

Total revenue, FY2018–FY2025

20182025$275B$0B $142B$275B

FY2025 at a glance

$275.2BTotal revenue
$10.4BOperating income
$8.1BNet income
12.8%Gross margin
3.8%Operating margin
2.9%Net margin

These are the figures the explorer generates from the same SEC XBRL data as the Sankey diagram.

Screenshot of the generated HTML explorer

And one page to navigate them: pick a ticker, drag the year, press play.

The explorer keeps every diagram connected to its source data, XBRL tag, trend, and inline 10-K.

Statement data view Trend charts view

Takeaway

  1. Claude Code and Codex are especially strong for automation.
  2. You have much more control over the inputs, outputs, and revisions.
  3. The work is managed locally in a folder you can inspect and rerun.

Action plan if you’re interested

  1. Install Claude Code.
  2. Make a Markdown (.md) file that describes what you would like to do.
  3. If you repeat something often, consider constructing Claude Skills.

Thank you

→ / ← or click to move · F for full screen