The middleman always wins

AI was supposed to cut out the middleman. That was the pitch. Intelligent systems that connect you directly to what you need, no gatekeepers, no rent-seekers, no unnecessary layers between you and the thing you're looking for. Instead, we got a new generation of middlemen. They just happen to run on GPT.

The disintermediation promise

The original vision was elegant. AI would collapse the distance between producers and consumers. Why hire a travel agent when an AI can plan your trip? Why pay a recruiter when an AI can match candidates to roles? Why use a search engine when an AI can just give you the answer? This idea has deep roots. Every major technology wave has promised disintermediation, the removal of intermediaries from a value chain. The internet was supposed to do it. E-commerce was supposed to do it. Blockchain was definitely supposed to do it. And now AI carries the same torch. But there's a pattern that keeps repeating, and it's worth paying attention to.

Aggregation theory never left

In 2015, Ben Thompson introduced Aggregation Theory, a framework for understanding how internet-era companies win. The core idea is simple: in a world of zero distribution costs, whoever controls the user interface controls the value. Suppliers become commoditized. The platform that aggregates demand wins. Google didn't create web content, it organized access to it. Netflix doesn't own most of the content it streams, it owns the experience of finding and watching it. Spotify doesn't make music, it makes the interface between listeners and songs frictionless. The pattern is always the same. The aggregator sits between supply and demand, makes itself indispensable to both sides, and captures the margin in between. AI hasn't broken this pattern. It has accelerated it.

Every AI startup is becoming a middleman

Look at the current landscape. Thousands of companies have raised money to build products that sit between foundation models and end users. They don't own the intelligence. They rent it from OpenAI, Anthropic, Google, or Meta. What they own is the wrapper, the interface, the workflow, the orchestration. A "ChatGPT for lawyers" startup doesn't build the model. It takes an existing LLM, adds a legal-specific prompt layer, connects it to some document sources, and charges a subscription. An "AI-powered" customer service tool does the same thing with a support knowledge base. An AI writing assistant does it with a text editor. As Perplexity CEO Aravind Srinivas has pointed out, "Everything is a wrapper. OpenAI is a wrapper around Nvidia and Azure. Netflix is a wrapper around AWS. Salesforce is an Oracle database wrapper valued at $320 billion." The dismissal of wrappers misses the point. The wrapper is the business. It always has been. Even the infrastructure layer is generating new intermediaries. The Model Context Protocol (MCP), introduced by Anthropic and now governed by the Linux Foundation, has become the de facto standard for connecting AI agents to external tools. MCP servers are middleware, a new connective layer that sits between models and the systems they need to interact with. Every MCP server is, in essence, a middleman. The AI orchestration market is exploding precisely because someone needs to coordinate all these moving parts. And coordination is just another word for intermediation.

Why this keeps happening

The answer is uncomfortable but consistent: distribution beats product. It doesn't matter if a user could replicate the same output by calling the OpenAI API directly for a fraction of the cost. Most people won't. They'll pay for the convenience of a polished interface, a curated workflow, and the feeling that someone else has done the hard thinking for them. Convenience beats raw capability, every time. That's why Expedia thrived even though you could call airlines directly. It's why Google Flights is eating into Expedia's share even though the same flight data is publicly available. It's why AI wrappers survive even when the underlying model is a commodity. The layer that makes things easy is the layer that captures value. This is not a bug in markets. It's a feature.

The middleman just changes form

Travel is the clearest historical example of this cycle. In the 1970s and 1980s, booking a flight meant calling a travel agent. They had access to airline reservation systems, understood fare rules, and handled the complexity on your behalf. You paid for their expertise, sometimes directly, sometimes through commissions baked into the ticket price. Then the internet arrived. Online travel agencies like Expedia, Travelocity, and Booking.com promised to cut out the travel agent. And they did, sort of. But they didn't eliminate the middleman. They became the middleman. They aggregated supply from airlines, hotels, and car rental companies, presented it through a searchable interface, and took a commission on every booking. Then Google entered the picture. Google Flights, Google Hotels, and Google's travel search features promised to disintermediate the OTAs. Expedia's own CEO acknowledged that "Google tried to disintermediate us." And Google is doing it again with AI Mode, building a closed loop from travel inspiration to booking that threatens to cut off the OTAs' primary traffic source. The travel agent didn't disappear. The role just migrated up the stack. From a person behind a desk, to a website, to a search engine, to an AI assistant. At every stage, someone new sits in the middle.

The Strait of Hormuz principle

There's a physical analogy that makes this pattern even clearer. The Strait of Hormuz is a narrow waterway between Iran and Oman, barely 33 kilometers wide at its narrowest point. According to the U.S. Energy Information Administration, it is the world's most important oil transit chokepoint, with roughly one-fifth of global oil supply passing through it daily. The strait doesn't produce oil. It doesn't refine it. It doesn't consume it. It just happens to sit between the producers and the consumers, and that geographic position gives it enormous strategic leverage. Digital middlemen operate on the same principle. They position themselves at the chokepoint of a value chain, the narrow passage through which supply and demand must flow, and extract value from the traffic. AI infrastructure is creating new chokepoints everywhere. The model API is a chokepoint. The orchestration layer is a chokepoint. The user-facing interface is a chokepoint. Whoever controls the narrowest passage between the model and the user captures the most value.

For builders: don't fight the pattern

If the history of technology teaches anything, it's that trying to eliminate middlemen usually just creates new ones. The more productive question is: which middleman position can you own? The companies that are building durable businesses in AI aren't the ones trying to disintermediate everything. They're the ones embracing their role as orchestrators. They integrate deeply with existing systems, accumulate proprietary context, build workflow-specific interfaces, and make themselves hard to replace. The best "AI wrappers" aren't thin prompt layers that can be replicated in an afternoon. They have deep integrations with CRMs, ERPs, compliance systems, and industry-specific tools. They accumulate customer data, usage patterns, and domain knowledge. Over time, that context becomes a moat. As one venture analysis put it, if you have deep integrations, domain-specific drudgery automation, and accumulated context, "you are not a wrapper. You are an AI-first company that happens to use foundation models."

The only two escape routes

If you don't want to be at the mercy of a middleman, there are really only two options. The first is to own the end user relationship. If users come to you directly, no intermediary can extract a toll. This is why Apple, Google, and Amazon invest so heavily in owning the consumer touchpoint, whether it's a phone, a browser, a speaker, or a storefront. The closer you are to the user, the harder it is for someone to wedge themselves between you. The second is to own the model itself. If you control the intelligence layer, everyone building on top of you is paying rent. This is the position OpenAI, Anthropic, Google DeepMind, and Meta are competing for. It's an expensive game, requiring billions in compute and talent, but it's the ultimate leverage position in the AI stack. Everyone else is playing the middleman game, whether they admit it or not.

The pattern is the point

It's tempting to be cynical about this. Another layer of rent-seekers, another generation of companies that own nothing but charge for everything. But intermediation isn't a market failure. It's a market function. Middlemen exist because they solve a real problem: they reduce friction, aggregate options, manage complexity, and save people time. The travel agent existed because booking was complicated. Expedia existed because comparing options was tedious. Google Flights exists because even comparison shopping takes effort. AI middlemen exist for the same reason. Most people don't want to manage API keys, craft system prompts, and orchestrate multi-step tool calls. They want someone to package that complexity into something usable. And they're willing to pay for it. The middleman always wins because the middleman serves a real need. The form changes with every technology wave, from travel desks to websites to algorithms to AI agents, but the function persists. Someone will always sit in the middle. The only question is whether it's you.

References