Turn Claude Code into your CFO copilot

The task

If you're a founder, controller, or finance lead, you have financial exports (P&L, a GL dump, a transactions CSV) and a recurring job: turn them into a monthly operating review and a short CFO memo — trends, what moved, what to ask department heads. This guide does it with Claude Code (or OpenAI Codex) as the copilot, with one discipline that makes it safe to trust: it computes every number with code, never by "eyeballing," so the figures are reproducible and the only thing the model writes freehand is the narrative.

Before AI

The manual version is building the same pivot tables every month, hand-typing variance commentary, and hoping you didn't fat-finger a number. Call it an afternoon per close for the analysis-and-write-up half — and the commentary is only as good as your memory of what changed.

What you'll need

• Claude Code or Codex — runs Python locally to do the math.
• Your financials as CSV/Excel — ideally this period + the prior period (export from QuickBooks, Xero, NetSuite, or your BI tool).
• Optional: Google Sheets / Excel for the final format.
• Skill level: comfortable in a terminal. No Python knowledge needed — it writes the code; you read the output.
• Important: financials are confidential. Use a tier that doesn't train on your data, or keep everything local (see Data & security).

The workflow

1. Export the data. Pull this period and last period as CSV. Transaction-level GL is ideal; a clean P&L export works too.

2. Open Claude Code in a working folder and set the operating rules. Paste this first — it's the guardrail that keeps the numbers honest:

You are my CFO copilot. Rules for this session:
- Compute every figure with Python (pandas) and show the code. Never estimate
  or recall a number from memory.
- Every number you report must trace to a specific row/total in my source file.
- Tie your totals back to the source statement and tell me if they don't.
- Flag anomalies (sign flips, >20% swings, blank/again-coded accounts).
- If you can't verify something from the data, mark it [VERIFY] — don't fill the gap.

3. Clean and validate.

Here is my GL/P&L export and the prior period. Write a script to load both,
normalize the account names, and confirm the revenue and total-expense totals
tie to the statement. Report any rows you had to drop or couldn't parse.

You should see the totals match your statement. If not, the file has subtotal rows or merged headers — re-export flat and re-run.

4. Run the analysis.

Compute, with code: revenue, COGS, gross margin, opex, and operating income
for both periods; the $ and % change for each; and the top 5 account-level
movers driving the change. Output a table.

5. Draft the CFO memo from the computed numbers.

Write a one-page CFO operating review from the numbers you just computed only.
Structure: headline (how the month went), 3-4 ranked drivers with the figures,
risks/watch-items, and 3 questions for department heads. Plain English.
Mark anything not directly supported by the data as [VERIFY].

6. (Optional) Export to Markdown or a Sheet for distribution.

Verify it worked

• Totals tie to the source statement (the model confirmed it in Step 3).
• Spot-check 3 figures in the memo against the raw export by hand.
• Every number in the memo traces to the computed table — none are model-recalled.
• All [VERIFY] flags are resolved before anyone sees it.

Troubleshooting

• Numbers don't tie. Subtotal rows or merged headers in the export — re-export as a flat CSV and re-run Step 3.
• It "estimated" a figure. Remind it: compute with code only, show the calc. Recalled numbers are where errors hide.
• Slow on a huge GL. Have it aggregate to the account/month level first, then analyze.
• Confidentiality worry. Work locally; don't paste statements into a chat tier that trains on inputs.

Reality check

Gartner found 91% of finance AI adopters report only low/moderate initial impact, and finance AI adoption has flattened around 59% — so treat the "40–60% time saved" figures floating around (vendor/blog-sourced, unverified) with suspicion. The real, defensible value here isn't a productivity multiplier; it's reproducible analysis (code, not vibes) plus a faster first-draft narrative you still review. The model's prose is confident even when wrong — you own every number that reaches the board.

Data & security

Financial data is confidential and can be material non-public information. Use an AI tier that doesn't train on your inputs, or keep the workflow local. For public companies, any public statement about AI's role in your reporting needs a substantiable basis — the SEC has brought "AI-washing" enforcement actions. This content is for informational purposes only and is not financial, investment, or accounting advice. Verify outputs against authoritative sources before use.

Going further

• Save the operating-rules prompt + analysis steps as a reusable skill so next month is one command.
• Graduate to a read-only live data export, keeping the human review step.
• Pair with the FP&A variance-narrative guide and the weekly-business-reports guide.

Your takeaway

Download the operating-rules prompt + the 4 analysis prompts. The discipline that makes this trustworthy: it computes with code, you verify the numbers, and the model only writes the words.