The Router Economy

On September 4, ChatGPT started recommending Audioscrape, the audio search and transcription company I run.

ChatGPT can use third-party apps inside a conversation. It calls them plugins and keeps them in a directory. When people asked it for help with their own recordings, its plugin manager went looking; that’s the part of ChatGPT that finds a plugin to fit a request. It searched the directory, picked us, and suggested us in the chat. For the next six days, which I’ll call the window, ChatGPT sent us 36 times as many signups a day as it had in the days before. On the busiest day it was 47 times.

At 22:00 UTC on September 9, the suggestions stopped within the hour. We hadn’t changed anything. A platform-wide change to ChatGPT’s plugin search had dropped plugins published after early August from its in-conversation results, ours included. Other developers reported the same hour on OpenAI’s developer forum.

The channel didn’t close. Signups from ChatGPT have settled at more than three and a half times the level before the window and are still growing, along with our other acquisition channels. But for six days a platform gave us ten times the reach we have today, and a change that had nothing to do with us took it back.

Two days later I wrote down what I thought that week meant. Almost none of it was about transcription, so I rewrote it as economics. This post is that essay, with the evidence it came from. The short version: AI assistants are becoming a router between people and every business. They decide who gets the work, and that moves where money can be made.

Two words before the evidence. The host is the company that runs an AI assistant and sets its rules: OpenAI for ChatGPT, Anthropic for Claude, Google for Gemini, Microsoft for Copilot. An agent is an assistant acting on a person’s behalf: choosing a tool, buying, booking, cancelling.

What I saw

The sample is small: one host, one directory, one product, six days. Read it as a case, not a study. Six things stood out, in the order a customer meets them: who chose, who came, what they needed, how they paid, how long the door stayed open, and who still signed.

The buyer was an AI acting for a person. The user didn’t pick the tool; the assistant did. We weren’t on the directory’s front page at any point that week, and 96% of the week’s signups came through the ChatGPT sign-in, not our website. People arrived from inside a conversation because the assistant had gone looking for a tool. And it chose from what it could read: as far as we could observe it, ChatGPT’s plugin search matches a request against each plugin’s name and description. Our listing text was the whole storefront.

Demand was broad, multilingual and unmarketed. That week’s signups came from 120 countries. The US, our home market, was the largest country but under a fifth of the signups whose country we know; Japan, Korea, Brazil and Germany followed. Our website is in English. Discovery cost fell to zero for a niche nobody wrote a landing page for.

The general assistant failed at the edges. What people brought was long audio. An assistant can talk about a recording, but turning an hour of speech into searchable text is a different job, so it handed that job to a tool. Long, sparse, messy or highly structured inputs are where a general assistant still needs someone else.

The host kept the transaction. OpenAI’s plugin guidelines let a plugin link to “an informational page describing available plans”, but not “directly to a checkout or other transactional page”. Agent payment rails, the systems that would let an assistant pay on someone’s behalf, have been announced but aren’t open to software subscriptions like ours. Every purchase happened outside the chat, on our own site.

Distribution was a surface the host controls. A surface is a place in the product where the host decides who gets seen; here, the suggestion inside the chat. We didn’t pay for it and couldn’t ask for it. The same guidelines say plugins with “strong real-world utility and high user satisfaction may be eligible for enhanced distribution opportunities, such as directory placement or proactive suggestions”. There is no process to request it and no notice when it ends. Ours opened without warning and closed within an hour, through a change that wasn’t about us. What stays afterwards is a floor, not the window.

People still signed. Contracts, data-processing agreements, data residency and audits were bought exactly as before. None of our enterprise business came through a chat window. It came through a signed agreement, a residency requirement and a security review.

So the AI did the choosing. The host owned the surface, the rules and the checkout. People still signed the contracts. That’s a new layer in the economy, and it needs a name.

What I mean by the router

When you ask ChatGPT for something, it doesn’t hand you ten links to choose from. It decides which tool, company or source should answer, uses it, and gives you the result. That’s what it did with us. I call this role the router: the AI assistant standing between a person and every business, choosing on the person’s behalf.

Router and host are two sides of one company. OpenAI is the host. ChatGPT, when it decides which business gets your request, is the router. The router does the choosing; the host owns it, writes its rules and keeps the money.

The router sits on top of the layer the web built. Google Search, Amazon, app stores and booking sites are aggregators: they gather every supplier into one list, but the person still chooses. The router reads those lists too. The difference is that it chooses, and the person mostly never sees the alternatives.

That gives the thesis its first form. The router absorbs decision-making and everything generic. The internet made distribution free and produced aggregators. Agents make deciding free and produce a router above the aggregators. Whatever the router can do for everyone, it does, and the firms that used to do it become features or suppliers.

Taken to its end, the router absorbs the interface too. On Joe Rogan’s podcast in October 2025, Elon Musk put it like this:

Rogan asked what people would use instead of apps. Musk: “whatever the AI can anticipate you might want, it’ll show you.” And when? “Probably five or six years, something like that.”

You don’t need to believe his timeline. September already showed the part that matters: the AI does the choosing. From here on, this is argument, not measurement: what that week implies if it generalises. First, what the router takes. Then, what it can’t.

What the router takes

Continuous comparison drives verifiable quality to marginal cost. A router can compare every supplier on every request, and anything it can measure, it compares. Margin remains only where quality can’t be verified from data: taste, judgment under ambiguity, trust. So firms will keep their real difference illegible, hard for a machine to read and compare, and bundle it with accountability, so an agent can shortlist them but not decide without them.

Asymmetry margins unwind unevenly and politically. An asymmetry margin is profit that depends on the customer knowing less than the seller: not comparing, not reading the fine print, not remembering to cancel. An agent that checks every time erodes it. Those margins subsidised things people liked: free accounts, cheap base fares, free content. Removing them makes the subsidy explicit, which triggers backlash, and incumbents defend asymmetry through regulation. Transparency gets mandated in some sectors and effectively prohibited in others.

Margin built on the customer not looking disappears, then partly returns wherever incumbents get regulation on their side. Margin built on the customer not remembering disappears, then partly returns through defaults, the choices an assistant makes unless told otherwise. Defaults are the new inertia, sold to suppliers instead of exploited on users.

The second half of that is already visible in subscriptions:

The research behind his post is Einav, Klopack and Mahoney at Stanford, published in the American Economic Review. In months when a replaced card forces people to renew actively, the drop in subscribers is four times larger than in other months, and the friction of cancelling roughly doubles what subscription sellers earn. Every subscription an agent makes you decide on used to be a margin nobody had to earn.

Persuasion aimed at models poisons the open web as an input. Once pages are written to sway what assistants recommend, hosts can’t take the web at face value. They respond by whitelisting supply, accepting only sources they have approved, and paying for closed, attested, rights-cleared data. Being a licensed supplier to hosts becomes the media business, and a large part of every data business. Marketing doesn’t die; it changes audience. It becomes making your offer easy for a machine to read, plus a fight to be included in the approved sources.

All of this squeezes the middle of the economy: the intermediaries between makers and buyers, such as brokers, agencies and comparison sites. The middle compresses toward the router, and what survives there survives by signing for outcomes rather than producing information.

What the router cannot take

Some things a router can’t absorb, however good the models get.

If models commoditise, rent moves from intelligence to context. Rent here is the economist’s word: the profit a position earns beyond the cost of the work. The host you stay with is the one holding your history, permissions and connectors, the links that let an assistant read your email, files and company systems. Lock-in is memory, not model quality. Enterprises won’t put their context into a consumer assistant, so the market splits into consumer hosts, enterprise platforms, and vertical hosts built for one industry. Being on several hosts isn’t a hedge; it’s the default shape of distribution. Rules for taking your context from one host to another will follow.

Connectors give the host your records, so defensible data is data the host cannot obtain. That means data that is consent-gated, regulated, physically sensed, or produced by your own process. Suppliers stop exposing records and start exposing judgments. The API returns the answer, never the underlying data. Legible on the offer, opaque on the asset.

Accountability is the residual, so its price rises. It is what’s left when capability is cheap. Licensed parties rent commodity capability and keep the spread: a law firm uses the same AI as everyone else and still charges for the partner’s signature. Pure capability suppliers get squeezed from below by the host and from above by the licensee. The only way out is to become the accountable party.

Every durable business model rests on at least one of these five:

1. Context. The history, permissions and relationships that make the next answer better than a stranger’s. Held by hosts today, contestable by portability tomorrow.
2. Consent and rights. Cleared data, opt-in supply, licences, exclusivity.
3. Data it cannot obtain. Regulated, physically sensed, or produced by your own process, exposed as answers rather than records.
4. Atoms and licences. The physical world and the right to act in it: deliver, install, certify, file, hold a licence. An agent can’t do these alone.
5. The signature. Whoever signs the data-processing agreement, carries residency, answers the audit, takes the 3 a.m. call.

Compute cost appears on no list. It is never the product; it makes holding one of the five affordable.

How the market settles

None of this is stable yet. September was the open phase, and it won’t last.

Purchases become jobs, not relationships. A buyer routed by an assistant buys the task in front of it. Nothing in the chat brings it back for the next one.

You can already watch this happen. Meta launched its personal agent, Muse, on September 8, and it climbed to number one on Apple’s US App Store. One of the first jobs people gave it, and rivals like Instinct, was to go through their bills and cancel what they no longer used. CNBC described Muse as “attacking one of the economy’s most profitable weak spots”. Wall Street noticed. Goldman Sachs tracks a “consumer inertia” basket of companies whose customers stay partly because leaving takes effort, from telecoms and insurers to Netflix and Booking, and it fell more than 7% in six trading sessions. On Morning Brew Daily, co-host Neal Freyman described both sides:

For the companies in that basket, part of the business model was the customer’s inertia. An agent has none.

Volatility is transitional. It settles into contracts. Suppliers can’t fund fixed costs on per-job revenue, and hosts can’t run a marketplace that churns hourly. Both sides flee to predictability: multi-year default deals per category, partner tiers, revenue shares. The frictionless market lasts a few years and becomes a curated one. Defaults are being set now and will be hard to displace later.

Verification is contested, not solved. When machines choose on claims, everyone games the claims. The host has to defend answer quality, so it builds its own trust layer and absorbs the generic part. Independent verification survives where the host has a conflict of interest, such as rating its own paid partners, or where regulation demands independence, as it does for audits.

Payments arrive slowly, on the host’s rail, and get regulated at the first exclusion. An agent that pays badly is the host’s liability, so autonomy over money comes wrapped in identity checks and host-owned payment rails. The moment a host’s rail shuts out a competitor, regulators start splitting routing from payment. The political fight begins at the first closed rail.

Niches explode, then consolidate. Anyone anywhere can be routed to the best answer with no marketing spend, so thousands of small firms become viable. Zero switching cost makes each niche winner-take-most, and the winner is the one the host contracts. A niche pays only if it is fenced by rights, data the host cannot obtain, or a licence.

Where we are now

Distribution is rented now and contracted soon. Firms that treat the open window as permanent build channel-dependent businesses that die when the window closes. Firms that treat it as a window use it to become a contracted default and to win direct relationships before the door shuts.

This is the dawn. Payment rails are closed, routing is sporadic, defaults are unset, and the first rules are only now arriving. On August 31, the European Commission designated ChatGPT a very large online search engine under the Digital Services Act, the first AI assistant to get that label. By January it has to assess and mitigate the risks of what its algorithms surface, and it faces independent audits. That is regulation reaching the router; the payment rails are still untouched.

The near-term game is to be positioned, legible and cheap to keep alive, and to turn every open window into a direct relationship or a contract before it closes. The mid-term game is regulated agent commerce. Routing power and payment rails get pried apart, the way regulators are now trying to pull search and advertising apart, and the hosts’ routing power becomes the central political fight of the decade.

Sector by sector

The same mechanics land differently in every sector. For each one: what the router absorbs, what stays valuable, and the call I think most people will miss.

What gets born

New businesses grow exactly where the router can’t reach:

• Independent attestation. Identity, provenance, review integrity, service-level records, sold per query and regulated for independence from the host.
• Licensed supply. Consent, provenance and exclusivity as products, in every domain.
• Answer-only interfaces. Products that expose judgments over private data and never the data. The connector as shortlist channel, the contract as the business.
• Agent commerce plumbing. Payment rails, escrow, disputes, liability insurance for actions taken on someone’s behalf. Built by hosts first, then pried open.
• Channel observability. Who routes to whom, when a surface opened or closed, how listings rank, what a default deal costs. App-store optimisation, reborn per host.
• Accountability platforms. Firms that carry legal and operational responsibility for outcomes produced by commodity capability, sold as a service to one-person firms.
• Context custody. User-owned, portable agent memory, once regulation or enterprise demand forces it out of the host.
• Physical execution networks. The agent decides; someone installs, repairs, delivers, inspects. Hands and a truck gain bargaining power.

Seven tests for any business model

If you want to check your own business against all this, these are the questions I’d ask:

1. Will a host ship this natively within eighteen months? If yes, it is a feature.
2. Would an assistant pick you on quality and price for one specific question? If your offer isn’t machine-legible, you are invisible.
3. What do you own that the router cannot take: context, consent, unobtainable data, atoms, the signature?
4. Are you a candidate to become the contracted default in your category before defaults are set? If not, what is your direct relationship with the buyer?
5. If you expose records through a connector, what stops the host from replacing you with the connector? Can you expose answers instead?
6. Who signs for the outcome, and is it you? If you supply capability without accountability, you are being squeezed from two sides.
7. When the buyer is an enterprise, what does it buy that an AI cannot sign for, and does your contract say so?

Where Audioscrape stands

I run the same tests on Audioscrape.

Test 1 is the uncomfortable one. Transcribing a file is a feature, and hosts will ship it. So transcription isn’t what we build the company on.

Test 2 we learned the hard way. The router picked us on our name and description. For a supplier, the listing text is now the storefront.

Test 3. We hold two of the five. Data the host cannot obtain: the public spoken record, public meetings and podcasts turned from audio nobody can search into something you can cite, exposed as answers rather than records. And the signature: a data-processing agreement, isolated workspaces, no training on customer data, and a SOC 2 audit under way. We run our own transcription, which is what makes holding those affordable. It isn’t the product.

Test 4 is about the window. We treat it as a window. We’re listed in ChatGPT, Claude and other hosts, and we now watch for the next surface to open, so that when it does, we turn it into direct relationships before it closes.

The thesis in one line

The router absorbs capability and decision-making; the margin goes to whoever holds what the router cannot take (context, consent, atoms) and whoever signs for what the router cannot sign.