Agents should never be free

You’ve lived this before. Free email, free social networks, free storage. There was always a price. And you paid it later, when it was too late to negotiate. Now there’s an AI agent that wants to act on your behalf. Are you sure you want it to be free?

The economics of free was never philanthropy. It was an asymmetric extraction strategy refined over two decades. Dropbox illustrates this with precision: of its more than 700 million registered users, barely 2.6% pay. But those 18 million subscribers generate over $2.5 billion in annual revenue ¹. The hundreds of millions who don’t pay aren’t a liability — they’re the product. Their usage patterns fuel the conversion engine, and their contact networks provide the virality that brings acquisition costs to zero. In social networks the exchange is even starker: your demographic data, your attention, and your social graph are auctioned to the highest bidder in real time. That’s the business model.

But in each of those cases, what you surrendered was passive: records, clicks, browsing history. What an agent surrenders is qualitatively different: it’s acts in the real world. When it prioritizes a vendor, executes a purchase, or dismisses an alternative, the consequence has already happened. There’s no undo button. And the bias isn’t accidental: industry analyses document that Gemini systematically favors products from Google’s ecosystem ², while Copilot prioritizes Azure architectures and GitHub repositories ³. OpenAI has already introduced sponsored recommendations in ChatGPT, reaching $100 million in annualized advertising revenue in just two months ⁴. This isn’t a programming flaw. It’s the reward function of whoever subsidizes the inference.

If you hired a human assistant and discovered that part of their salary was being paid by a vendor they negotiate with on your behalf, you’d call it a conflict of interest. It would probably be illegal. The SEC has already warned that it will examine how financial advisors use AI systems to ensure their automated recommendations comply with their fiduciary obligations ⁵ — a signal that regulators are beginning to treat AI as a vector for conflicts of interest, not just technical error. But when an AI agent does it, we call it a freemium model and accept the terms without reading them.

Agents are going to be the decisive interface of the next decade. But unlike a search engine or a social network, the agent doesn’t show you options to choose from. It chooses for you. Whoever controls that layer doesn’t need to convince you of anything. They just need to execute.

In the history of the internet, when you don’t pay for the product, you are the product. With agents, the equation gets worse: you’re no longer the product. You’re the transaction. Perhaps it’s time to demand from algorithms what we’ve always demanded from professionals: that whoever acts on your behalf works exclusively for you. And that has never been free.