Better finance results, with 90% less AI usage (your 5 smart switches)

This is the second part in my series on AI ROI. Because the best way to provide ROI is to achieve the same results, with less AI usage.

“Why is Claude getting more and more stupid?”

Someone asked me this recently.
​
I told them that the chat got really heavy. When Claude's memory of a conversation gets too full, it starts forgetting your instructions, and your answers get worse.

This is also the reason why you keep hitting your usage limit in the middle of the day.

Somebody on your team is probably doing it right now. A very long chat, a PDF or two attached, already maxing out their limit. And it’s costing you.

In another one of my courses. Someone told the group she regularly blows through her tokens on Claude, so she's now super careful about how she writes her prompts.

This is the wrong approach.

Today I will show you how to reduce 90% of your Claude usage, whilst getting the same results.

NOTE – The same switches can be used for any AI, today's examples just use Claude.

​

​

Don't cut your team's AI use. Cut the waste.

If your Anthropic bill keeps going up every few months, don't start by asking your team to use less. Ask them to waste less.

It's the same as turning off the lights before you leave the office. There are ways to spend less money and keep the same results.

And as we move to usage-based pricing, this matters more every month.

Claude remembers your whole conversation: your messages, its replies, your uploaded files, your instructions and your preferences.

After every new prompt, it reads everything again from the beginning. Only then does it answer. That's where most of your tokens go.

How much waste? I’d say Claude wastes almost 90% of your tokens, and only 1% goes into the answer you asked for.

Treat that as an illustration though, since Anthropic keeps changing things. An example of this is the recent release of Sonnet 5.5, which Anthropic claims is faster, and uses less tokens.
​
The fix? My 5 Smart Switches:

​

​

The 5 Smart Switches

Switch 1: Make Sonnet everyone's default. Opus has to earn its seat.

Choose the model before you type a word. Here's how to pick:

• Sonnet: your everyday default. Use it for writing, analysis, coding, and Excel and PowerPoint work. It gives results very close to Opus, for almost half the price.
• Haiku: for quick, simple jobs. Writing an email or a memo and nothing else? Haiku is enough. No need to pay for more.
• Opus: for the hard stuff. Deep, complex analysis and big multi-part tasks. It takes longer and burns far more tokens, so save it for work that really needs it.
• Fable: ignore it for now. It's mainly for people using Claude Code. You don't need it in Claude chat, Cowork, or Claude in Excel and PowerPoint.

Try Sonnet first. If it doesn't perform, that's when you pay for Opus. You decide when a task needs this extra intelligence.

Super important – the model you pick decides how fast your tokens go. Choose it before you type your first prompt, every time.

Switch 2: Turn your PDFs into markdown before you upload them.

Say you have a report, a document or a presentation as a PDF. You give it to Claude and start asking questions.

Claude doesn't only read the words. It parses the whole PDF: the structure, the format, the images, the fonts, the headlines, the metadata. That's why it's heavy.

Quick maths: one PDF page can use 1,500 to 3,000 tokens. The same content in markdown can use fewer than 200.

Three ways to convert, straight away:

• Ask Claude to convert the file, once.
• Use the free markdown converter from Microsoft.
• Paste the text into a Google Doc or Word document, then save it as a markdown file.

Then upload that version instead of the PDF.

Tip – Start with the files you use all the time.

Switch 3: When Claude gets it wrong, edit your prompt. Don't argue with the answer.

Claude misunderstood you. So you reply, "No, that's not what I meant." We all do this.

But look at what you just added to the chat: your original prompt, Claude's wrong answer, and your correction. That's three pieces of context, and Claude re-reads all three on every future turn. The wrong answer is still in there too, pushing later answers in the wrong direction.

So do this instead:

1. Scroll back to your original message.
2. Click edit and fix the wording.
3. Hit enter.

Claude deletes its previous reply and writes a new one.

If a chat started well, then drifted a few messages in?

1. Go back to the last message where everything was on track.
2. Edit it with your improvement.
3. Send it again.

You keep the useful context. You lose the noise. This one saves a lot of tokens.

Switch 4: New topic, new chat. And when a chat gets long, summarize it and start fresh.

Say you write the financial commentary for your monthly report in one chat. Then, in the same chat, you start a memo on revenue recognition.

The two topics have nothing to do with each other. But every time you prompt about the memo, Claude still takes into account the whole commentary conversation with it. That's a total waste of tokens, and I see it in my trainings all the time. If it's you, that's fine. Just start a new chat.

Same topic, though? Say you do the board presentation after the commentary. Staying in the chat makes sense, because it already has the context. But if the chat gets really long, or the answers get slower or worse, don't push on.

Two rules:

1. New topic, new chat.
2. Long chat and falling quality? Ask Claude for a summary, choose what to keep, and start a new chat with only that. You'll probably get better results.

Not sure how to ask Claude for the summary? Here's the prompt I use.

Use it as it is, or add more detail about what you want Claude to remember. Either way, you keep the important knowledge and drop the old messages and the back and forth.

Switch 5: Put the files you reuse in a Project.

Do you use the same files every week? Do you have to upload them every time? No.

Uploading the same file into a new chat every time is the expensive way.

Inside a project, Claude can load the relevant parts of your files instead of reading whole documents on every turn. That's much more efficient than a fresh upload in a new chat.

Use markdown files in your project too. Switch 2 and Switch 5 work on top of each other.

How much you save depends on the file and how often you use it.

​

​

The One Thing to Remember

Don't use less AI. Make AI re-read less.

1. Right model.
2. Lighter files.
3. Edit, don't argue.
4. New topic, new chat.
5. Files in a Project.

When people are scared of hitting the limit, they stop using AI.

And there’s no way to prove AI ROI if people don’t use it 😉

​

Best,
​
Your AI Finance Expert,

– Nicolas
​
​

P.S. – The best way to learn from me directly is to join my weekly masterclasses. 60 mins + Q&A where I'll get all your AI in Finance questions answered. Join me here.​

P.P.S. – Want to see all 5 switches applied live? Here's the full walkthrough → How To Save 97% of Your Claude Tokens As a CFO​

​