Delta, Chase and Amazon are shutting out AI agents, and every task you delegate now depends on the other side's permission

The Verge reported on 17 October that websites are closing their doors to AI agents. Delta Air Lines is stopping agents from booking tickets for users. JPMorgan Chase and Charles Schwab confirmed they block agents from accessing accounts, and both cite security. Yelp told TechCrunch that it only lets in agents that pay for its data licensing programme. Users say Cars.com and Amtrak, which used to work with their agents, no longer do. Amazon said the "continued access" of Meta's Muse assistant violates its site policies, and users report that xAI's Grok, Instinct's agent and OpenAI's Dots cannot use Amazon either.

Access is now negotiated site by site

OpenAI, Meta, Google, Instinct and others sold agents as a pair of hands that could do anything you could do online. The Verge's reporting shows that access is now decided one site at a time, and every site sets its own terms. Yelp charges. The banks refuse. Amazon goes to court: it sued Perplexity over Perplexity's agentic shopping feature, a judge ordered Perplexity to stop, an appeals court overturned that order in August, and Amazon has since filed an amended complaint.

Other companies are making deals. Meta chose not to fight Amazon and partnered with Walmart, Gap, Sephora, Best Buy and others so Muse can shop on their sites. Shopify has opened its network of storefronts to agents.

The record shows the motive. Perplexity argued that Amazon wants human eyeballs in front of its ads. Meta put it like this: "Turning away a personal agent means turning away the customer behind it." Whoever owns the customer relationship decides whether an agent counts as a customer.

A partnership does not guarantee access either. The Verge cites Reddit reports that Google is blocking some accounts that use Muse for "suspicious activity", and that some users still cannot check out on Walmart despite the Meta deal. Chris Canal, chief executive of the AI safety evaluator EquiStamp, told The Verge that anti-bot systems built to stop scams and hacks catch agents too: "These AI agents trigger those same safeguards."

What this does to delegated work

Canal draws the commercial conclusion: "If there are companies that are actively working to make that not happen, there are just fewer useful things the agents can do."

For an operator, the point is narrower and more useful. Handing a task to an agent only delegates it if the site on the other end accepts the agent. How reliable your agent is now depends on other companies' access policies, which you do not control and which can change without notice. Cars.com and Amtrak went from working to unreachable, according to the users The Verge cites.

The framing still holds: AI does the routine work. You do the thinking. In this case the thinking includes knowing which channel is sanctioned and whether a failed booking was caught or simply missed. Canal expects the web to get more "annoying", with human-verification steps spreading to sites as ordinary as a grocery shop. That makes the person who can route around a closed door more valuable.

Your Next Move

1. Map every workflow you hand to an agent against the sites it touches. For each one, note whether there is a sanctioned route: a retailer partnership like Meta's, a Shopify storefront or a paid licensing programme like Yelp's. Anything that touches an airline or a bank that has said it blocks agents goes back to a person or an official channel today.

2. Add a check for silent failure. Sites that worked last month can stop working without telling you. Set a weekly review of agent-completed tasks, and require a confirmation number or receipt before you count a task as done.

3. Re-sort your own task list with access in mind. A task an agent could do in principle but cannot reach in practice is still your job. The method for sorting a week's work into what a system can absorb and what stays with you now needs one more column: is the agent allowed in?