The format of business is changing

Something happened while we were building our own operating system. We noticed that the connective tissue between every agent, every workflow, every piece of automated intelligence was the same thing: markdown.

Not JSON. Not HTML. Not a proprietary file format. Plain, readable, structured markdown.

Our proposals start as markdown. Meeting notes get synthesized into markdown. SOWs, internal docs, client briefs — all markdown. This article that you are reading? If you were to look at the files on our server, you would find one for this article that is written in, you guessed it, markdown. Not because we mandated it, but because it's what works when both humans and machines need to read, write, and act on the same content.

That realization changed how we think about everything we publish, including this website.

Why Markdown matters now

Markdown has been around since 2004. It wasn't designed for AI. But it turns out the properties that make it good for humans — lightweight, readable without rendering, structurally expressive — are exactly the properties that make it efficient for machines.

It's human-readable. You don't need a browser, an app, or a special viewer. Open a markdown file in any text editor and you can read it immediately. In fact, let us show you. Here's the entire syntax you need to know:

# This is a heading
## This is a smaller heading

This is a paragraph. You just write.

**This is bold** and *this is italic*.

- This is a list item
- This is another one

[This is a link](https://www.buildfm.com)

> This is a quote.

That's it. Congratulations — you just learned markdown. No classes to take, no software to install, no spec to memorize. Everything above is readable before it's rendered, and it maps cleanly to the structure of any document you'd ever write. This is why it's winning.

It's machine-readable. LLMs process markdown natively. When an AI agent needs to understand a document, markdown preserves the semantic structure — headings, emphasis, links, lists — without the overhead of HTML tags, CSS classes, or JavaScript event handlers.

It's token-efficient. This is where it gets interesting from a cost and environmental perspective. Consider a typical webpage. The HTML source might be 50KB, but the actual content — the words, the meaning — might only be 5KB. The rest is markup, styling, scripts, and structural noise. When an LLM processes that HTML, it's spending tokens on <div class="flex items-center justify-between px-4 py-3 sm:px-6"> instead of the sentence inside it.

Markdown strips that ratio down dramatically. You get the structure and meaning with a fraction of the tokens.

The environmental math

This isn't just an efficiency argument. It's an environmental one.

Every token an LLM processes requires compute. Compute requires energy. When millions of AI agents crawl the web to gather information, they're processing vast amounts of HTML, JavaScript, and CSS that contains no useful information. It's the digital equivalent of shipping products in boxes ten times their size.

If a webpage's meaningful content is 10% of its HTML source, then 90% of the tokens spent processing it are waste. Multiply that across every AI-powered search, every agent gathering context, every RAG pipeline pulling from the web — the energy cost of that inefficiency is staggering.

Markdown doesn't solve climate change. But when the format of information exchange can reduce token consumption by an order of magnitude, it's worth paying attention to.

What we built

We put this philosophy into practice on our own site. In the header of every page, next to the menu, there's a switch with two options: Human and Machine. You are reading the human version of this article right now.

Click Machine, and the page transforms. Instead of the rendered website, you see the raw markdown representation of everything on the page — headings, paragraphs, links, team bios, all of it. Clean, structured, portable. You can copy it, paste it into any tool, feed it to any LLM, or just read it. The switch remembers your choice as you move from page to page, so you can browse the whole site as a machine would.

This isn't a gimmick. It's a statement about how we think content should work: every page should be legible to both audiences.

How it works: Every page on the site has a markdown twin. Add .md to any URL and you get it; this article lives at /resource-center/markdown-is-eating-the-world.md. When you switch to Machine, the page simply loads that twin, the same markdown an agent would get. For articles like this one, it's the file the article is actually written in. For every other page, our server converts the rendered page into markdown with Turndown, stripping away navigation, scripts, and decorative elements. Nothing extra loads until you flip the switch.

The part we're most excited about

Here's where it gets really interesting.

We didn't stop at a client-side toggle. We built content negotiation directly into our server. Any HTTP client — including an LLM, an agent, or a simple curl command — can request any page on our site as markdown:

curl -H 'Accept: text/markdown' https://www.buildfm.com/resource-center/markdown-is-eating-the-world

That's it. No special API. No API key. The same URL that serves a human-readable webpage will return clean, token-efficient markdown to any machine that asks for it. Standard HTTP content negotiation — the way the web was designed to work. (Tools that can't set headers can use the .md version of the URL instead.)

Every page on our site speaks both languages. A browser gets HTML. An agent gets markdown. Same URL, same content, different format for different audiences.

To close the loop, we added instructions to our llms.txt file — the emerging convention for communicating with AI crawlers — telling any LLM that visits our site how to get markdown: add .md to any URL, or send the Accept: text/markdown header. The file already describes who we are and what we do. Now it also tells machines how to read us efficiently. Any agent that checks llms.txt before crawling will know it can get clean markdown from every page, skipping the expensive HTML-to-text conversion entirely.

We think this is what the web should look like when machines are first-class consumers of information. Not a separate "API" bolted onto the side. Not a scraping-friendly sitemap. Just proper content negotiation — the same pattern the web has supported since HTTP/1.1, finally being used for something it was always capable of — combined with a simple, discoverable instruction file that tells machines the best way to consume your content.

What this means

We're not suggesting everyone rebuild their website to serve markdown. But we are suggesting that the format of business communication is shifting, and the organizations that recognize it early will have an advantage.

When your content is markdown-native:

  • Your AI agents work better. They spend tokens on meaning, not markup.
  • Your workflows compose. Markdown flows cleanly between tools, systems, and people without conversion.
  • Your content is portable. No vendor lock-in, no proprietary formats, no rendering dependencies.
  • Your environmental footprint shrinks. Less compute per interaction, across every automated process.

Marc Andreessen famously said software is eating the world. Two decades later, we'd argue the format is narrowing. Markdown is becoming the lingua franca of the human-machine interface — not because anyone decided it should be, but because it's the simplest format that both audiences can read.

At FM, we build software that works for humans and machines alike. If you're thinking about how AI fits into your content, your operations, or your product — let's talk.