It is a small dialog box, and it looks like nothing at all. In the last week of August, Anthropic shipped its own browser inside its desktop application, and it offers to bring your logins over, site by site, from Chrome, Edge or Firefox. Banking, mail and single sign-on stay out unless you explicitly say otherwise. Anyone reading that thinks convenience. One click less, no second sign-in, finally seamless.
What is handed over in that dialog box is not a cookie. It is the authority to appear on the internet as me. And this is where something breaks that has carried the structure of the web for thirty years, the assumption that at the other end of a session sits a human being who looks, clicks, remembers and sees advertising.
The number everyone quotes is already the old one
In the spring the news went around that machines had overtaken humans on the web. On June 3, 2026, Cloudflare Radar showed 57.5 percent of the requests for HTML content coming from automated systems, and the Imperva Bad Bot Report published this April puts 2025 at 53 percent using a slightly different basket. That was celebrated as the tipping point, and it is only the prelude.
Those figures describe machines that identify themselves as machines. A crawler has a signature, an IP range, a behaviour you can recognise and lock out. More than a million Cloudflare customers did exactly that, and in the five months after Cloudflare made blocking the default on July 1, 2025, 416 billion AI bot requests were stopped. The market has learned how to deal with visible machines.
The digital colleague inside the built-in browser appears in none of those statistics. It arrives through my login, in my session, from my machine, carrying my usage profile. To the provider on the other side this is not bot traffic, this is me. The line between human and machine is not disappearing because there are more machines. It is disappearing because from now on they carry my identity papers.
No provider can lock out a customer who sends a machine
Anyone who believes this can be contained by contract or by technology should read the Ninth Circuit decision of August 4, 2026 in the matter between Amazon and Perplexity. The court held that where an assistant accesses a site using the user's own credentials, it is the user who accesses the site with the help of an AI actor, and not the AI company. The assistant, however advanced, is a tool and not a person for the purposes of the statute. American computer fraud law therefore does not reach the provider of the assistant. What remains are contract and terms of service.
Legally that is a footnote. Strategically it is an earthquake. A provider can lock out machines, which is a technical problem with technical answers. What a provider cannot do is lock out paying customers because those customers delegate their work. Anyone who tries is litigating against their own revenue. Spotify, every streaming service, every portal, every line-of-business application now faces a choice that is not a choice, either accept the machine-represented customer or lose the customer. This is why every loud defensive move of the past two years has been aimed at crawlers. Against the logged-in stand-in there is no clean instrument.
The browser was never a product, it was a toll booth
This is the real fracture, and it is larger than a product announcement. The browser never made money by displaying pages. It made money because every search, every piece of research, every buying impulse passed through a narrow gate where attention was sold and behaviour was measured. Searching, browsing, informing yourself, comparing, those were sessions, and sessions were inventory.
When I move those activities into the assistant, my need for information does not fall. What falls is the number of moments in which somebody can sell me something. The direction is already visible in the access data. For the first week of August 2025 Cloudflare counted fifty thousand pages fetched by Anthropic's crawler for every single visit that crawler sent back, against roughly nine hundred for OpenAI's and one hundred and eighteen for Perplexity's. Readings of the same dashboard through 2026 show that distance narrowing and nowhere near closing. Content is consumed, visits do not come back. The built-in browser is the consequence of that, one step further along, because now it is not the public part of the web being drained but the part behind my password.
And it does not stop at research and commerce. As soon as the virtual colleague processes video and audio, it takes over the news and entertainment layer too, at precisely the point where those business models earn their money, in the recommendation, the playlist, the next suggestion. A service whose value hangs on the session does not lose its users. It loses the session, which is the same event with a friendlier balance sheet in year one.
The interface is still priced as substance in the cycle that turns it into a liability
For owners and funds this is where it becomes concrete. Two asset classes sit directly across this development, and in both of them the decisive item is still being booked as an asset.
The first is everything financed by attention. Those valuations rest on reach made of sessions, and those sessions are moving into a context where nobody can place a format. The user base stays stable for a while, the monetisable interaction does not.
The second is enterprise software, and there it becomes structural. The value of a vendor has been calculated for twenty years out of two things, the installed base and the interface that locked that base in. User interfaces were differentiation, habit, switching cost, pricing power. A per-seat licence counts humans in front of screens. The moment the digital teammate treats the interface merely as a protocol to reach the function behind it, the differentiator becomes an interchangeable connector. Anyone valuing a software portfolio today still prices the interface and the seat count as substance. In this cycle both are the liability, because both are exactly the quantity that disappears first when the customer hands work to machines and does not buy a second account for them.
The blind spot sits in the organisation, not in the technology
The uncomfortable part is that most organisations cannot even measure this. Their analytics knows two categories, human and bot. For the third one, the human under machine representation, there is no field. The access log carries my name, the security report carries my name, and the usage figure that goes to the board says nothing any more about who is actually working in there.
The problem, as always, is not the code. It is the organisation that has grown around it. It has an owner for privacy, an owner for security, an owner for licences, and not a single owner for the question of what the application is supposed to be when the thing on the other side is no longer a person.
That question is the decision now on the table, and it is still open. Whoever defines what their application offers a machine, an outcome instead of a screen, a described entry point instead of a defensive wall, a price for effect instead of a price for seats, keeps access to their customers even when those customers stop showing up in person. That choice is available to everyone today, and it costs no acquisition, only clarity about your own product.
Whoever waits will find in two or three years that their market never left them, it simply stopped attending in person. And they did not see it coming, because their own name was in every log.
Gerhard Kürner is CEO of 506.ai, the European platform for Service-as-a-Software and agentic engineering, and author of Work After AI. Around 1,000 conversations with boards, owners, and PE funds across DACH and Europe.