Mara ReyesMore than 90% of Claude Cowork usage is unrelated to software development. The agent escaped its presumed category before its interface finished escaping the desktop.
The chatbox made intention the user’s job
The first interface answered a specific question: how should a person access a general-purpose model? The answer was a destination, organized around a text field and a sequence of turns. Users brought the intention, supplied the context, waited for the response, and decided what to carry back into the rest of their work.
That structure made the model legible. It also created the boundary that the next phase would erase. OpenAI could let ChatGPT users adjust tone and formatting because the interaction still happened inside a recognizable assistant. The system became more personal, but it remained somewhere users had to go.
The movement beyond that box is easy to misread as a series of feature releases. It is a change in function. AI is no longer organized primarily as a destination; it is being placed inside the interfaces where attention and work already occur.
The answer was not a better chatbot. It was less chatbot.
Every new surface removes a deliberate step
OpenAI’s new voice architecture makes simultaneous listening and speaking the design. The model is no longer limited to a clean sequence in which one side finishes before the other begins. Voice becomes an ongoing surface rather than a different input method for the same box.
Meta is making the same structural move from the opposite direction. Image generation is entering its social apps, placing creation inside an existing flow instead of requiring a separate visit to an AI tool. The company is also reportedly testing glasses capable of continuous sensing, though that effort remains a report from sources rather than an announced product launch.
Anthropic’s Cowork expansion closes another gap. The capability is moving from the desktop app to web and mobile in beta for Max subscribers, extending an agentic work surface across devices. Its adoption has already outrun the category most closely associated with advanced AI tools.
That figure is not merely evidence of broader tasks. It shows software development losing control of the system’s revealed purpose. As Cowork crosses web and mobile, the interface follows the user instead of asking the user to return to it.
Meta’s acquisition of Manus points in the same direction. Manus talent is joining Meta to deliver agents across Meta products, including Meta AI, while the Manus service continues to operate and sell separately. Generation supplies content; sensing supplies context; agents supply action. Distributed across existing products, those functions form a loop rather than a collection of tools.
The surface stays stable while the intelligence changes
There is a second reversal inside the first. As AI moves beneath familiar interfaces, the model can gain reach while losing ownership of the user relationship. The social app, work application, phone, or glasses remain visible; the intelligence underneath becomes a component that the owner of the surface can select, replace, or price against alternatives.
In Excel and Outlook, Microsoft is reportedly beginning to replace OpenAI and Anthropic models with its own MAI models to reduce AI costs. The applications remain Excel and Outlook. The model layer changes underneath them.
The outside models have not stopped being useful. But usefulness inside a destination and usefulness inside someone else’s surface produce different economics. Once the interface owns distribution and context, model substitution becomes an operating decision. The intelligence is more ambient precisely because its identity can become less important to the person using it.
This is the treacherous turn built into the chatbot phase. The destination created direct relationships between model companies and users, but it also trained users to expect the capability everywhere. Meeting that expectation distributes the product into surfaces controlled by other companies. The strategy succeeds until ubiquity makes the original destination optional.
Ambient intelligence requires visible boundaries
The strongest counterweight appears on Meta’s glasses. Alongside reports of continuous-sensing tests, Meta says an update will disable the camera when the glasses detect tampering with the privacy LED. The tests remain unconfirmed; the shutdown rule is announced. Together they expose the bargain: greater sensing requires a boundary other people can see and trust.
Continuous sensing should not be confused with permission to sense without limits. An ambient system can recede from the wearer’s attention only while its boundaries remain legible to everyone else. The model can leave the chat window, but the camera still needs a light, the light needs protection against tampering, and the sensor needs to shut down when that protection fails. Intelligence becomes less visible; accountability has to become more concrete.
The ambient layer still has an address
“Ambient” describes the experience at the edge, not the structure supporting it. The layer that appears weightless across glasses, feeds, and assistants rests on infrastructure substantial enough to reorganize corporate financing.
Meta also moved $30 billion of debt for building AI data centers off its balance sheet through special-purpose vehicles. More data-center capacity makes it possible to place inference across more surfaces; more surfaces create demand for more capacity. Each side reinforces the other until the chatbot is no longer the product’s natural container.
The 90% figure now looks less like a change in audience than a change in container. The chatbot’s box has been turned inside out: capability spreads across the surfaces people already use, while its physical limits reappear at the privacy light on one end and the data center on the other.