Claude Code for CPAs: A First Day With Anthropic's Coding Agent

Most coverage of Claude Code is written by software engineers, for software engineers. This is a first-day account written for the audience that's actually being underserved by it: accountants who don't code but spend their week wishing some piece of repetitive work would do itself.

What I built in the first 90 minutes

A small agent that takes a folder of vendor invoice PDFs, extracts vendor name, invoice number, date, and total, and writes a clean CSV. Did this work before? Sort of — manual data entry into Excel, ~15 minutes per stack of 10 invoices. The agent does it in under a minute and leaves a flagging column for invoices it couldn't parse confidently.

I wrote zero lines of code. I described what I wanted, watched Claude Code write a Python script using pdfplumber, run it, fail on a few invoices, ask me to look at the failure cases, adjust, and produce a working version.

The mental shift

Traditional accountant's relationship with code: "this is the IT department's job; I describe what I want and they build it." Time-to-result: weeks.

Accountant with Claude Code: "I describe what I want; Claude writes and runs the code; I verify." Time-to-result: an afternoon.

The IT-department dependency doesn't go away for production-grade work — anything touching client data or shared systems still needs proper review. But for the daily small tasks that used to be unsolvable because they weren't worth a ticket, the agent unlocks a huge productivity surface.

Setup, in plain language

1. Install Node.js (a JavaScript runtime). Free download.
2. Open a terminal (Terminal on Mac, PowerShell on Windows). Type one command — npm install -g @anthropic-ai/claude-code — wait a couple minutes.
3. Get an API key from console.anthropic.com. Anthropic gives you free starter credit; topping up is pennies for casual use.
4. Type claude in any folder. You're in.

Total setup: about 20 minutes for a non-technical user.

The five real CPA use cases I tested

1. CSV cleanup — vendor master, customer master, employee roster. Trivial.
2. Bank reconciliation prep — match a bank export against a GL export, produce an exception list. Worked first try on simple cases.
3. Tax document organization — rename and sort 200 tax documents by client name and document type. Beautiful.
4. Variance commentary first drafts — produce 3-sentence variance explanations from a numbers file. Drafts needed editing but saved time.
5. Audit sample selection — randomly select 25 transactions from a 10,000-row file with stratification by amount. Worked, with documentation of the methodology.

What surprised me

• It's legitimately good at Excel-like work in Python. If you can describe what you'd do in Excel ("group by vendor, sum amounts, filter to top 20"), Claude Code can write the Python equivalent and run it on far larger datasets than Excel handles cleanly.
• It explains what it's doing. Every script comes with comments. You can ask "why did you do it this way?" and get a real answer.
• It catches its own errors. When a script fails, it reads the error message and tries something different rather than giving up.

What still needs care

• Don't feed it client data without firm approval. Same rule as any external AI service.
• Verify the output. The agent will produce confident wrong answers. Spot-check before relying on output.
• Don't let it write to live systems without explicit human approval at every step. Read-and-prep is safe; write-back is risky.
• Watch your API costs. Long agent sessions on big files can rack up tokens. Set a monthly budget cap in the Anthropic console.

The career framing

Three years from now, an accountant who's comfortable driving Claude Code (or whatever the dominant agent is by then) will outproduce a peer who isn't — by a margin too large for managers to ignore. The barrier to picking it up is one afternoon. The compounding return is a decade of more productive work.

Try it on a personal task this weekend. The setup cost is small; the unlock is real.