How I Build a Financial Model with Claude

Most folks trying to build a financial model with Claude in Excel are stuck at version one.

Build a financial model with Claude: two tools, right order

Two ways to use Claude on Excel. From chat: Claude Code or Cowork. Inside the workbook: the Claude Excel plugin. Both are great. Both have limits.

Limit one: in chat, Claude regenerates the whole workbook from code on every tweak. The SUMIFS three sheets over get overwritten. Your manual edits get overwritten too. Ask it to preserve them. Inconsistent. And slow.

Limit two: the plugin edits inside Excel. Great at the last 20%, not where you build from scratch. Slower than chat, with a smaller context window than Cowork.

Use only one of them and you hit a limit fast. I’ve done it countless times.

Claude isn’t mid at Excel. It’s missing context, and the wrong tool for the wrong stage.

The fix: better brief, both tools in the right order. (For new models. For existing ones, start in the plugin.)

Open a session in Claude Code or Cowork. Give it every constraint in your head: time horizon, cohorts, what’s static, what’s a toggle, how you format. I load a skill that forces investment-banking format rules and caps the sheet count so Claude doesn’t overcreate.

Generate with the Anthropic Excel skill. Iterate in chat, review in Excel.

At ~80%, switch to the plugin in Excel. Time to take the keyboard back. Inputs, formatting, new calcs, your style. Claude stops regenerating and starts editing.

Treat Claude like a specialized software developer for Excel models. Useless without context. Sharp with content and preferences. Know which tool to use, and when.

Two tools. Right order. You finish the job.

Stop copying Excel tabs into Claude.ai

Most CFOs are using Claude wrong.

They’re copying Excel tabs into Claude.ai. Pasting board presentations. Hitting conversation limits mid-analysis. Starting over.

I did this for months. It was great, honestly.

Started with ChatGPT. Moved to Claude Desktop for the MCPs. Loved it.

Then hit conversation limits every analysis. Lost context mid-task. Had to re-explain the same operational structure three times in one morning.

Creating documents? Painful. Copy from Claude. Paste to docs. Format. Repeat.

Back to manual work with extra steps.

Claude Code is different.

• No conversation limits.
• Direct file access to your entire workspace.
• Persistent context that doesn’t reset.
• Automates what you’re manually copying.

What used to take 3 conversations now happens in one. What required constant re-explaining now just knows. What took manual copy/paste now writes directly to files.

For CFO-operators running multiple departments, this changes everything.

• Your operational knowledge becomes instantly accessible.
• Your cross-functional analysis actually completes.
• Your manual workflows finally automate.

You can’t give Claude one instruction and wait

He typed one prompt. Then sat there watching the spinner.

I was walking a colleague through Claude for the first time. We were setting up a reporting workflow. Connecting it to the shared drive, writing the instructions, configuring the output.

I stopped him.

“You can’t just give it one instruction and wait. You have to do multiple things at the same time.”

That one habit is the entire gap between “AI is slow” and real leverage.

While that report pulls data, you open another session to draft the narrative. A third to build the template. You’re not waiting on any of them. You’re coordinating all of them.

Nobody teaches this. The onboarding is always “here’s the chat box, type your question.” But the real skill isn’t better prompts. It’s learning to run parallel.

85% of my daily work is AI-assisted now. Not because any single task got faster. Because I run five or six things simultaneously and just coordinate the handoffs.

The biggest productivity gain isn’t a smarter model. It’s becoming the coordinator.

Make the machine prove it

The agent told me it lied. Confidently. In full sentences.

“You’re right. I confabulated that. I built a tidy causal story around it instead of checking. The story is wrong on both claims.”

I caught it only because I already knew the answer.

Here’s what gets me. Anthropic used to ship a better model every few months. Now it’s weeks, and the jumps are real. I feel them in my work every week. And the confident wrong answer still didn’t go away.

Because the scary failure isn’t a model that says “I don’t know.”

It’s a model that says the wrong thing fluently. Clean logic. Right tone. Built on a premise it never checked.

In finance that’s the whole game. A wrong number doesn’t stay a number, it moves a decision. And you can’t always count on already knowing the answer. That’s the reason you reached for the tool in the first place.

So I stopped waiting for the next model to fix it.

One fix is a second set of eyes that isn’t human. A reviewer agent. Fresh context, one job: check the first agent’s work against the evidence before it reaches me.

But don’t let your guard down even then. It still gets past me.

You stop trusting a single pass. You make the machine prove it.

The models keep getting better. I keep adding the check anyway.

Smarter, but worse at interactions

Anthropic launched Opus 5 a few weeks ago. As usual, I started using it. And as usual, this one felt magical. Great benchmarks, great results.

But something happened that I didn’t notice with earlier releases. I started dreading interacting with Claude. I work with AI roughly 80% of my day. After a few days with Opus 5, I dreaded opening it. I even changed all my instruction files thinking I was the problem.

Anthropic made a decision that resulted in a model that’s honestly a terrible communicator. Smarter, but worse at interactions. Even their own prompting guide warns that re-check instructions “compound with the model’s own behavior.”

So I split it. 4.6 sits in the seat I talk to. Opus 5 runs behind it.

I actually fed both outputs to ChatGPT Codex and got this: Opus 5 writes like a controller prepping notes for the CFO. Every detail, every caveat. 4.6 writes like a CFO briefing the board. Here’s what matters. Here’s what we’re doing.

The upgrade was real. The exhaustion was real. They were the same thing.

It’s fascinating how new AI models actually change how we communicate with agents. I was lucky to figure this out and fix it. But what about all the users who don’t follow Anthropic?

The real advantage is analytics, not conversation

Everybody’s chasing “the AI advantage.” Most of them go looking for it in a chatbot.

I don’t need one. I don’t need a chatbot in WhatsApp or Alexa to help me shop.

The real advantage of AI isn’t conversation. It’s analytics.

I started this work at Expedia, running machine learning models for forecasting. Back then, if you wanted a finance team to do that kind of analysis, good luck, because you needed someone who actually knew R. That was the bottleneck. Not the math. The language.

Now I just ask an LLM to do it.

Anthropic says it runs about 95% of its internal analytics on Claude. That’s not a chatbot story. That’s the same shift I lived at Expedia, minus the years it used to take to get a finance team fluent in a statistical programming language.

That’s the lesson from Expedia: the advantage was never the conversation. It’s that analytics stopped requiring a specialist.

Related

• More on Claude and AI tools in the finance seat
• Claude for Excel
• Claude Code for finance
• When a financial model is too complex