United Locked Out AI Agents. The Labs Are Selling the Gates.

Scott Kirby didn't dress it up. United's chief executive spent this week explaining why the airline will not let third-party AI agents book its flights — credential-sharing, security rules that require a human to acknowledge the fine print an agent cannot — and then he said the part he actually meant.

To be honest, I don't want to create another middleman. — Scott Kirby

A day earlier, OpenAI — whose ChatGPT agent is among those United shut out — started selling a service called Business Hub, where companies claim verified profiles, feed live pricing and inventory, and surface as sponsored agents ^[1]. That is an identity desk and a listings page: a middleman's job, offered to the middlemen, by the company the airline just refused to let in. The door-closing is everywhere this fall, and it is fast. Goldman Sachs is rationing its frontier models to the employees who need them, mostly for cyber work; JPMorgan capped what engineers can spend running Claude Code at $2,000 a month ^[2]. Apple — which sells no agent product at all — added agents to the risk factors in its annual filing, warning they could change code, send messages, or finalize purchases nobody meant them to ^[3]. The Financial Stability Board told banks to keep a human in the loop ^[4]. Amazon blocked Meta's Muse agent when it refused to stop crawling the site ^[5]. And across more than 1,500 organizations, ninety-four percent of websites cannot accept an agent connection at all, while a quarter actively block the agents that do identify themselves ^[6]. Now look at what the labs everyone is locking out have been shipping, and the blockade reads like a spec sheet being filled. Every demand the gatekeepers make — prove who you are, let a human approve, meter the spend — is a product the labs already have on the shelf. Take identity. Dun & Bradstreet has proposed a "Know Your Agent" registry: a persistent, verifiable link between every agent and the business accountable for it ^[7]. Okta, whose login software already sits in most corporate doors, plans to re-price its identity business from per-user to per-agent — billing each autonomous worker the way it bills each human one — and its chief executive put the prize plainly ^[8].

The agentic, the security and governance and visibility for agentic, being the control layer for that, could be bigger than all of identity is today. — Todd McKinnon

Anthropic — whose Claude is the agent JPMorgan capped — has been selling the oversight side since February, through private plugin marketplaces that let a company build its own shelf of approved agents and distribute them internally, giving the CIO curation as a product ^[9]. Since last fall, Google has been signing American Express, PayPal, and more than 60 other companies onto its Agent2Agent and AP2 protocols — the rails that let one agent talk to another and settle a payment, the plumbing a toll booth needs to sit on ^[10]. And OpenAI, which cannot get its own travel agent past airline blocks, is selling travel sites the integration path around those blocks, turning chat into bookable outcomes with MakeMyTrip and ixigo ^[11]. OpenAI cannot walk its agent through United's door. What it shipped instead is the identity desk the door's owner wants. Underneath the salesmanship is arithmetic that does not bend. Anthropic's IPO filing, made public this week, shows $895.4 billion in purchase commitments and lease agreements against $16.3 billion in first-half revenue, plus a $650 million operating loss over the same stretch ^[12]. A company that must multiply its revenue roughly sixty times over cannot afford to stand on the outside of the doors its agents need to pass through. Dario Amodei, Anthropic's chief executive, stated the stakes in his own filing.

If my revenue is not $1 trillion, if it’s even $800 billion, there’s no force on Earth, there’s no hedge on Earth, that could stop me from going bankrupt. — Dario Amodei

None of this means every gate is bolted. Shopify opened its checkout to ChatGPT and to Meta's Muse, and when Amazon blocked Muse, shoppers and sales quietly moved to Shopify merchants instead ^[13][14]. A hotel-booking site called BusinessHotels built an API that lets any of the three big chatbots check live rates and book a room ^[15]. Italy's regulator forced Meta to reopen WhatsApp to rival chatbots ^[16]. And while American banks cap agents, UAE banks are pulling them in ^[17]. The alignment is real; the uniformity is not. The finished hardware, though, is metering a market that barely exists. The busiest agent-payment rail to date — x402, running on Coinbase's Base network — has moved something like 165 million transactions worth roughly $50 million in total ^[18], and 87% of those trades were between a hundredth of a cent and a dime ^[19]. Ninety-four percent of websites still cannot accept an agent at all ^[6]. And by the only survey that has asked, six percent of consumers would let an AI complete a purchase on its own ^[20]. Six percent. The gates are built, the identity desks staffed, the toll rails standardized — for a customer who, so far, mostly isn't ready to walk through.