Complex Model (Plus Deck) in 60 Mins with Claude

There’re 3,141 finance pros in my community, and most of them start with no idea how to use Claude effectively in Excel and PowerPoint.

But lots of them now build complex models that used to take 2-3 (plus the deck) in minutes.

To teach you to do the same, I'm hosting this free 60-min masterclass with Ramp. You'll see:

• How to build a financial model from scratch in Excel using Claude, (with graph updates as assumptions change).
• The audit-tab method to catch errors (well before anyone questions your numbers).
• How to turn it all into a branded PowerPoint deck in 5-10 minutes (without touching a slide)

You'll leave with the exact workflow my community is using to offload 80% of their time to AI, so they can focus on the work that gets them recognized.

​Save your free spot here before another ‘quick scenario’ lands in your inbox ;)

Uber spent its whole 2026 AI budget by April

So for the last 12 months, the top question I got in masterclasses was "Which AI tool is best?”.

But, in the last few months it's been changing. Now more and more finance leaders are asking me, "How do we make our usage last and keep the cost down?"

Imagine you set a budget for the year, and by April it is gone. That is what happened at Uber. [1]​

Heavy users were each using $500 to $2,000 a month. [2]​

Plus, Microsoft (yes, Microsoft) just took its own engineers off Claude Code to reduce usage. [3]​

They are not the only ones. Meta built a leaderboard that ranked 85,000+ staff by how much AI they used. The top person used 281 billion tokens in a month. They called it "Token Legend." Then they shut it down. [4]​

So the most advanced AI companies on earth spent a year teaching people to use more.

Now they are teaching them to use the right amount.

Today, I will help you avoid the expensive lesson they all made.

“I’ve reached my usage limit already”

If you’ve not seen the 'I've run out of AI' messages from your team yet you will.

And as finance, this is not just about your own team’s usage. You will be the one explaining the total AI bill to the board (especially if you have developers!).

AI usage is becoming more and more of a concern.

Sam Altman even predicts this. He said:

Plus, Goldman Sachs is already predicting global AI consumption will grow 24× by 2030, as more companies hand work to agents. [5]​

So, you may not be using your budget in 4-months. But, you need to have a smart way of managing your usage now, before it causes you big problems in the future.

I am a finance guy first. So when I see Meta rank 85,000 people by how much AI they use, to me this says that they are not using AI efficiently.

For example - This study of 22,000 developers said AI got 34% more tasks done, but bugs went up 54% and reviews took 5× longer. [6]​

What’s super important to understand here is. More usage does not mean more value.

If you reward "more," you just end with "more."

What you really want is the most efficient use of AI to get the job done.

Tim Ferriss would call this the ‘Minimum Effective Dose’ (MED) - only this time we’re not talking about one workout for 15-mins twice per week to get jacked.

We are talking about smarter use of AI whilst maintaining quality and accuracy.

Your Usage Footprint

Now, to be clear: not every team is facing usage limits right now.

Some AI licenses have no limit. But Claude has usage limits built in, and ChatGPT is adding them. This will become more of a trend.

If you are not being stopped by limits yet, my rule of thumb is simple: always use the best model. There is no point limiting capability if there is no restraint.

But if your limits start showing up (or you want to be ahead of the curve). Here is the method.

I call it the Usage Footprint. Before you run a task, you ask one simple question: what is the smallest footprint that still ensures quality and accuracy?

Your usage is not one number. It’s the model you pick, whether thinking is on, whether you use an agent, and whether a connector gets used.

Turn each one up and the footprint grows. In one study, running a task as a full agent used up to 1000× more tokens than handling it as a simple chat (without much change in the quality of output). [7]​

Now, this is not about using less AI. I use a thinking model for about 95% of my own work, because most of my work is real analysis.

The point is to match the tool to the job.

You would never put your best, most expensive analyst on data entry. AI is the same.

How to Optimize Your AI Footprint with 4 Dials

By the end of this, your team will use far less for the same work. Here are the dials I would use, in order.

Dial 1: Stop sending your emails to the most powerful model.

Match the model to the job. Use a fast, low consumption model for emails, notes and tidying up text. Save the top model for real analysis and anything you would put your name on. Big models cost more per use and write more, so the wrong default costs you twice (More on the current models further down).

• Tidying up a payment-chasing email or a note to the board → fast, cheap model.
• Turning your finished numbers into plain words for the pack → fast, cheap model.
• Building a 5-year model or testing "what if" scenarios → top model.

Dial 2: Turn off "thinking" for anything a junior could do in one go.

Thinking mode makes the AI reason before it answers, which consumes more. Great for planning a hard model. A waste for "tidy up this table." Leave it on for the hard 20%, off for the easy 80%.

• Fixing the column headers on a trial balance export → thinking off.
• Pulling the key dates and amounts out of one vendor contract → thinking off.
• Working out a tricky deferred-revenue or consolidation question → thinking on.

Dial 3: Only hand a task to an agent when it truly needs many steps.

Agents like Claude Cowork or ChatGPT's Agents do not give one answer. They run lots of steps on their own. Perfect for a big job, like a full reconciliation. Overkill for a quick question. So save agents for the big stuff, and check their work like you would a junior's.

Examples:

• "What is our cash position today?" from one file → just chat, no agent.
• A one-line variance comment on this month → just chat, no agent.
• Reconciling a whole folder of bank statements against invoices → agent.

Dial 4: Watch when connectors get called, not just which ones you switch on.

Connectors like Gmail or Google Drive are not the problem. The problem is they get used more than you think, and each time they are ‘called’ (not just used) it adds to your usage.

Two fixes:

1. First, paste this line into your team's project or system prompt: "Only use a connector when the task really needs it."
2. Second, stop pasting Google Drive links for simple / short docs. When you paste a link, the AI has to call the connector to open the file. Sometimes it’s easy (and just as fast to copy/paste)

Examples:

• Sharing a small budget file with the AI → upload it, do not paste the Drive link.
• You only need one email → paste that email in, do not let it search the whole inbox.
• A project prompt that says "only pull live ERP data when I ask" → stops it calling every time.

Pro tip - Send the above to your team, or use AI to convert it into an SOP, it’s up to you.

The thing that is important is that your governance around usage is clear, so that you don’t end up with 24-1000x the consumption in the future (without knowing how you got there).

Current Model Table

The One Thing To Remember

AI usage is more than a percentage bar. It has four dials: model, thinking, agents, and connectors.

Turn all four up and one task can consume 1000× that chatting alone. Right-size them and the same task costs 1×.

Uber learned this four months into the year, the expensive way. You learnt it this afternoon, for free.

So before your next task, ask one question: what is the smallest footprint that still gets me high-quality accurate outputs?

Get that one question into your team's habits, and you’ll be the ones still using AI to get ahead when other companies are shutting it down ;)

Best,

Your AI Finance Expert,

- Nicolas

P.S - Is anyone on your team saying they run out of AI too fast? Hit reply and tell me (I read all replies.)

P.P.S - Using Claude efficiently is the difference between wasting budget and multiplying it. Here's how a pro does it → How to Use Claude in Excel Like a Pro (2026)​

​