ai-ledger

Stop Using AI Like It Is Excel

How accountants should actually use Claude, ChatGPT, and Copilot together

By Svetlana Toohey Published March 2026

Your AI Team

A Quick Story

The first time I used AI for a real accounting workflow, I did what most of us do.

Opened a chat box. Pasted data. Asked it to “analyze.”

Then I waited for magic.

What I got instead:

That is when it clicked. I was not using AI wrong. I was thinking about it wrong.

The Wrong Question

Most accountants are asking:

“Should I use ChatGPT or Claude?”

That is like asking: “Should I use Excel or email?”

They serve different purposes. And the right answer is almost always: use both, for different things.

If you have already explored how Claude’s different interfaces work, Ways to Use Claude covers that in detail. This article goes further: how do you assign the right AI to the right job?

The Right Mental Model

AI is not a tool. It is a team.

And just like your accounting team, not everyone does the same job – and not everyone should.

When you assign a staff accountant to a task that requires a controller’s judgment, things go wrong. The same principle applies to AI models. Each one has a role.

Meet Your AI Team

Claude Models – Your Core Accounting Team

Model Accounting Analogy Best For
Opus Technical accounting expert Complex logic, edge cases, multi-step reasoning, planning
Sonnet Senior accountant Building workflows, writing scripts, implementation
Haiku Staff accountant / automation Fast execution, repeating established tasks, simple analysis

Opus is the model you bring in when the problem requires deep thinking. A complicated revenue recognition question, a multi-entity consolidation with intercompany eliminations, or designing the architecture of a new workflow. It is slower and more expensive, but it gets the hard problems right.

Sonnet is your day-to-day workhorse. It writes Python scripts, builds reconciliation logic, structures data pipelines, and implements the plans that Opus designs. Most of the articles in this series were built with Sonnet.

Haiku is fast and lightweight. Once you have a script that works, Haiku can run it across files, summarize data, or handle routine analysis that does not require deep reasoning. Think of it as the model you use when the thinking is already done and you need execution.

For a deeper look at how Claude’s interfaces work (web, IDE, and API), see Ways to Use Claude.

ChatGPT – Your Whiteboard Partner

ChatGPT is best used for:

Think of ChatGPT as the colleague you walk to the whiteboard with. You sketch ideas, talk through problems, and refine your thinking. But you would not hand it a live workbook and ask it to run your close.

ChatGPT is a starting point, not a production system.

GitHub Copilot – The One That Works Inside Your Files

GitHub Copilot is where things change.

Copilot is not just another chat tool. It lives inside Visual Studio Code. It reads your files. It understands your structure. It works inside your workflow.

If you have already set up VS Code using the guidance in Getting the Right Tools Installed, adding Copilot is the next step.

How Copilot Actually Thinks

Most people treat Copilot like ChatGPT. That is a mistake.

Copilot prioritizes:

This means the way you organize your project directly affects how well Copilot can help you. A messy folder with random file names gives Copilot nothing to work with. A structured project gives it everything.

Your folder structure is your prompt.

/project
  /data
    /raw
    /processed
  /src
  /outputs
  plan.md
  status_update.md

This is not just organization. This is how you teach AI how to work with you.

If you have read Why Claude “Forgets”, you already know why plan.md and status_update.md matter. Copilot uses these files the same way – they become the context that makes AI useful across sessions.

The GitHub Superpower Most People Miss

Here is where it gets powerful.

Instead of explaining your workflow every time, you can import it.

Inside VS Code, you can ask Copilot:

“Pull the folder structure and workflow patterns from the PythonMuse Workflow Kit and set up my workspace.”

And within seconds:

The PythonMuse Workflow Kit is a ready-to-use project template on GitHub. It includes the folder structure, CLAUDE.md instructions, plan.md and status_update.md templates, and starter skills for bank reconciliation and margin analysis. Click “Use this template” on GitHub to create your own copy, or ask Copilot to import it.

The PythonMuse AI Ledger repository and the AI Accounting Framework are also public on GitHub. Copilot can find them, read them, and apply their patterns to your project.

You are not prompting anymore. You are deploying a system.

Skills: The Next Layer

A “Skill” is simply a reusable workflow with instructions that AI can follow consistently.

Instead of writing this every time:

“Compare bank to GL, find differences, save results…”

You create a Skill file:

/skills/reconciliation/SKILL.md

The Skill file tells AI exactly what to do, what data to expect, and where to save results. The Skills, Agents, and Models module in the AI Accounting Framework explains how to structure these files.

Two ready-to-use examples are available now:

Copy either into your project, point AI at your data, and you have a repeatable workflow.

How you use it today:

“Load the reconciliation skill from /skills/reconciliation and apply it to files in /data/raw. Follow plan.md and update status_update.md with results.”

Where this is heading:

“Use PythonMuse reconciliation skill.”

And Copilot will find it, load it, and execute it.

This is the shift from one-time analysis to repeatable workflows that we covered earlier – but now with a distribution layer that makes your workflows portable.

Model Orchestration: Assigning the Right AI to the Right Job

This is where everything clicks.

Model Orchestration Workflow

Figure: A four-step accounting workflow with the right model assigned to each stage.

Example: Bank Reconciliation Workflow

Step Model Task
1. Plan Opus Design a reconciliation workflow comparing bank statement and GL. Identify key steps and edge cases.
2. Build Sonnet Write Python script to perform reconciliation using files in /data/raw. Save results to /outputs.
3. Execute Haiku Run reconciliation script across all files and summarize differences.
4. Explain ChatGPT Summarize reconciliation results in plain English for CFO review.

You do not need a better AI. You need to assign the right AI to the right job.

This is the same principle behind the three-tier classification we covered in Article 10 – exploratory, repeatable, and audit-ready. The model you choose depends on which tier the task falls into.

Bringing It All Together

This is exactly why PythonMuse focuses on:

The shift is not from one AI tool to another.

The shift is from prompts to systems.

Try This: Start Here

  1. Open Visual Studio Code
  2. Install GitHub Copilot
  3. Clone the PythonMuse Workflow Kit (or click “Use this template” on GitHub)
  4. Ask Copilot to explain the structure
  5. Edit plan.md for one workflow you do repeatedly
  6. Run it

If you need help getting started with VS Code and Python, Getting the Right Tools Installed walks through the full setup.

Final Thought

You do not need to learn everything at once.

Start with one workflow you hate.

And ask: “What would this look like if it was repeatable?”

That is your first system.

*Related: The Power of Skills and Agents Ways to Use Claude Getting the Right Tools Installed From One-Time Analysis to Repeatable Workflows AI Accounting Framework*
This site is open source. Improve this page.