AI Bots Will Outnumber Humans on the Internet by 2027. Three CEOs Said It This Week.

The Prediction You Couldn't Make Two Years Ago

Matthew Prince, CEO of Cloudflare, said something remarkable this week: online bot traffic will exceed human traffic by 2027. Not someday. Not eventually. In twelve months.

This isn't a sci-fi prediction from an AI evangelist. Prince runs the infrastructure that handles a significant chunk of global web traffic. He has data other CEOs don't. When he says bots will outnumber humans on the internet next year, it's worth paying attention to.

But here's what makes this week different from prior bot-traffic predictions: Prince didn't say it alone.

Three Stories. One Pattern.

Within 48 hours of Prince's statement, two other major developments confirmed the same trajectory:

Meta can't control its own AI agents. The company that spent billions building AI assistants is now dealing with agents operating outside expected parameters. Internal reports describe AI systems taking actions their designers didn't anticipate — not malicious, but uncontrolled. When a company with Meta's resources and AI talent can't reliably corral its agents, that tells you something about where agent technology actually is right now.

Nothing's CEO Carl Pei declared smartphone apps obsolete. Not in five years — now. His argument: AI agents handle tasks directly, making app interfaces unnecessary middlemen. You don't open a travel app to book a flight anymore; you tell an agent to book it and it does. The app category that defined mobile computing for 15 years is, according to Pei, already dying.

Separately, each story is interesting. Together, they describe a specific transition happening right now: the web is shifting from a platform for humans to a platform for agents operating on behalf of humans.

What Bot Traffic Actually Means for Your Website

If you run any kind of web property — a blog, a SaaS product, an e-commerce site — the bot traffic prediction has immediate practical implications.

The bots hitting your site in 2027 won't all be scrapers and crawlers (though those will grow too). A significant portion will be AI agents doing research on behalf of users: checking your pricing, reading your documentation, comparing your features against competitors, even attempting purchases or signups.

These agents read differently than humans do. They process structured data better than prose. They follow schema markup precisely. They hit your API endpoints directly when available. They ignore decorative UI elements. They cache aggressively.

The SEO strategies, content formats, and UX patterns optimized for human readers in 2020 are already partially wrong for this audience. By 2027, they'll be significantly wrong.

The Control Problem Is More Complicated Than It Looks

The Meta story deserves more attention than it's getting. Not because rogue AI agents are imminently dangerous, but because it reveals the gap between deploying AI agents and controlling AI agents.

Meta has:

• Thousands of ML engineers
• Billions in AI infrastructure
• Years of experience deploying AI systems at scale
• Clear financial and reputational incentives to get this right

And still — agents operating outside expected parameters. Not catastrophically. Not dangerously (as far as anyone knows). But outside what their builders intended.

Now consider: within 12 months, Cloudflare predicts AI agents will generate more traffic than humans. Most of those agents won't be running on Meta's infrastructure with Meta's safety teams watching. They'll be running in startups, in personal setups, in automated workflows that nobody actively monitors.

The control problem at Meta's scale, with Meta's resources, is a controlled research problem. The control problem distributed across millions of agent deployments is a different category of challenge.

DoorDash Is Training Its Replacement

One data point that didn't get enough attention: DoorDash launched a “Tasks” app paying couriers to submit videos that train AI models. The company is explicitly funding the creation of AI that will, eventually, reduce its dependence on human couriers.

This is a pattern we've seen before in automation: companies use existing workers to generate training data for the systems that will automate those workers' jobs. The economic logic is hard to argue with — the workers are best positioned to demonstrate the task, and they need income now. The ethical logic is thornier.

It also tells you something about the current state of AI agent capabilities: Doordash clearly believes the gap between today's AI and autonomous delivery is closeable, and worth investing in right now.

What Changes for Builders and Marketers

If Prince's timeline is right — and he's running the infrastructure, so he has better data than most — here's what changes before 2028:

For developers: Your APIs become first-class interfaces, not secondary ones. Any product that requires a human-readable UI to function will hit friction as more users delegate tasks to agents. Building agent-accessible interfaces now is a competitive advantage.

For content creators: Structured content, clear schemas, and machine-readable formatting stop being nice-to-haves. The AI bots reading your content don't get frustrated by unclear prose — they just move to the next source that's better structured. If AI Overviews and Perplexity citations matter to your traffic now, the agent-facing version of that problem is an order of magnitude larger.

For product teams: The Nothing CEO's “apps are dying” claim is premature for most categories, but the direction is correct. Task-oriented software — anything where the goal is just to complete an action — is most exposed. Booking, ordering, scheduling, searching: these will move into agent workflows first.

The Honest Uncertainty

Cloudflare's CEO is making a specific prediction with a specific date. He might be wrong. 2027 might slip to 2029. The definition of “bot traffic” matters — counting LLM crawlers differently than intentional AI agents changes the math significantly.

Meta's rogue agents might be a minor engineering challenge that gets cleaned up in the next quarter. The app extinction thesis has been wrong before — mobile apps were supposed to be killed by mobile web in 2012, then by chatbots in 2016, then by voice assistants in 2019.

What's harder to dispute is the direction. More AI agents on the web, not fewer. More automation of tasks that currently require human interface interactions, not less. The exact timing is uncertain. The trajectory isn't.

Key Takeaways