Paul Ford’s view is direct: browsers remain extraordinarily powerful application software, and the cost of accessing that capability appears to be falling every day. This means the value boundary of AI products increasingly depends not on how much content their models can generate, but on whether they can connect to the real world at low cost.
What this is
A browser is, in essence, a unified gateway: users enter their needs, and websites and services return information or complete actions. In the past, people had to open e-commerce, mapping, office, and government websites separately. AI agents—software that can autonomously plan and invoke tools to complete tasks—are now trying to place all these entry points behind natural-language interfaces.
But the more unified the gateway becomes, the more complex the chain of responsibility becomes. Website redesigns, login verification, permission settings, and anti-automation mechanisms can all cause an automated action to fail. Therefore, falling “browser capability costs” include not only lower invocation fees, but also the engineering cost of connecting legacy, nonstandard systems to the web.
Industry view
We believe browsers remain the most realistic gateway to AI services in the short term because they naturally cover large numbers of websites and software products already in use, eliminating the need to develop a separate interface for every service. Models handle understanding and planning, while browsers handle navigation and execution. This division of labor could significantly shorten the time it takes for products to move from demos to practical use.
The risks are equally clear: prompt injection on a page may trick an AI into performing the wrong action; CAPTCHAs and risk controls can block automation; and page scripts may read sensitive information. If companies delegate high-privilege actions such as payments, price changes, or file deletion to agents, model judgment alone is not enough. They also need permission isolation, confirmation mechanisms, and operational audit trails. Another counterargument is that platforms may restrict automated access, in which case authorization and compliance requirements could offset part of the browser gateway’s cost advantage.
Impact on regular people
For enterprise IT: Agents may become a new operational layer across systems, but companies should first expose low-risk, reversible workflows and avoid giving AI direct control of high-privilege accounts.
For individual work: Collecting information across websites, filling out forms, and organizing data will take less time, while repetitive operations may decline. Decision-making, review, and accountability still require human confirmation.
For the consumer market: Users may become more accustomed to saying “book this for me” or “compare prices for me,” but price, terms, and responsibility for after-sales service will not disappear. Instead, platforms will need to explain their service rules more clearly.