The file formats that make AI cheaper and greener

Every time you hand AI a Word document or ask it for a PowerPoint deck, it does a lot more work than your words need. That work is measured in tokens. Tokens eat into your usage limits, show up on your bill and draw on the electricity and water that data centres run on.

Andy Crestodina at Orbit Media recently tested this across nine common formats, and the gap between the lightest and the heaviest was big. It's worth a read. Below is how I'd put it into practice in a small business.

Why some files are heavier than others

A .docx or .pptx file is a zipped bundle of code wrapped around your words: styles, fonts, layouts, positions on the page. To read one, the AI has to run a tool to unpack it. To make one, it writes code that builds it. PDFs and images often go through a separate, slower visual process.

Plain text formats skip all of that. The AI reads and writes them directly.

Orbit's rule of thumb is a good one. If you can open the file in Notepad or TextEdit, it's light. If you need Word, PowerPoint or Excel to open it, it's heavy.

The light formats to reach for

• Markdown (.md) for documents and drafts. It's plain text with a # for headings and a - for bullets, and it's the format AI works in best.
• CSV for tables, lists and data exports.
• HTML for reports and anything that needs to look polished. Everyone has a browser, so it opens on any device.
• The chat box itself for short things. If it's a paragraph or a small table, paste it in rather than attaching a file.

The heavy ones, and what to use instead

Word is the format most people ask for, and it's heavy both ways. Draft in Markdown, then paste the final version into Word and tidy the styling yourself.

PowerPoint is the heaviest of the lot. If the slides are the deliverable, fair enough, but ask for them once at the very end. If a readable page would do, ask for HTML.

For Excel, use CSV unless you need formulas, filters or several sheets.

PDFs are the one I see most often in Claude Projects and Custom GPTs. Convert each one to Markdown before you upload it, and ask the AI what can be trimmed while you're at it. You do that work once, and every future chat in that project is cheaper.

Where the cost really adds up: editing

Ask for a small change to a heavy file and the AI will often rebuild the whole thing. Five small tweaks can mean five full rebuilds (a bit like reprinting the whole brochure to fix one typo).

A few habits fix most of it:

• Agree the content in the chat or in Markdown before you ask for any file.
• Put all your changes in one prompt instead of five.
• Ask for the final format once, when the content is locked.
• Make the last small tweak yourself in Word or PowerPoint. It's quicker anyway.

Keeping it in proportion

Your prompts are small next to the energy it takes to train the models, and small next to flights or diet in your own footprint. That's a fair point.

But these habits cost nothing. They make your usage limits last longer, they speed things up and they add up across a team. They also pair well with telling your AI to be less wordy in its replies, which saves tokens in the other direction.

Where to start

Pick one habit this week. The easiest is to stop attaching Word documents and paste the text in, or save it as Markdown first.

If you'd like help making your team's AI use leaner and more responsible, take the free AI Check or get in touch for a chat.