Website restrictions expose AI agents’ access problem

Delta Air Lines blocks AI agents from booking flights, while Yelp offers access through a paid data licensing programme. As The Verge’s report on websites restricting AI agents explains, the services are taking different approaches to the same permission problem: an assistant can complete an online task only if the website lets it in.

The report examines access issues involving agents from OpenAI, Meta, xAI and other companies, across services such as Amazon, JPMorgan Chase and Charles Schwab. The banks pointed to security concerns. An agent’s technical ability is only one part of the equation. Planning a booking and completing it are still different things.

Better reasoning cannot resolve a website’s objection to automated access. An assistant may understand a request perfectly yet still be unable to act on it. That makes the conditions behind a product demonstration matter. Which services allow access, and what happens when it is refused? The answers shape how far its results can reasonably be generalised.

The reasons for restrictions also need examining separately. Yelp’s licensing programme turns access into a commercial negotiation, while the banks’ stated concerns are about security. Amazon has taken legal action against third-party agents and objected to Meta’s Muse assistant accessing its site. In each case, it is worth asking what would need to change for access to be accepted.

A failed task does not prove there has been a deliberate ban. General anti-bot systems may catch AI agents along with other automated traffic. Cloudflare has introduced measures against certain AI crawlers, citing infrastructure strain and disruption. Crawling for information is different from completing a transaction, so evidence about one cannot automatically account for failure in the other.

Negotiated access provides another route. Meta has partnered with retailers including Walmart and Gap, while Shopify has opened its storefront network to agents. Those arrangements raise a practical question about the choices available to an assistant. If access relies on commercial agreements, any assessment of usefulness must take account of services beyond them.

For education and academic work, the key principle is disclosure. An assistant that cannot inspect a source should say so. These restrictions concern commercial websites, yet inaccessible material cannot support a claim that it has been checked, however plausible the resulting prose may sound.

Progress would need evidence that tasks can be completed across clearly identified services, with failures accounted for. Websites would also need safeguards that do not impose excessive verification demands on ordinary visitors. Authorised tasks would need to work reliably, while people need a clear account of what the assistant attempted and what remained unfinished.

← Back to Blog