AI assistants are leaving the desk

Anthropic says more than 90% of sampled Claude Cowork use was not software development. TechCrunch reports the sample covered 1.2 million anonymized Cowork sessions from more than 600,000 organizations during the last two weeks of May. The largest category was business process work at 33.4%. Content creation and copywriting followed at 16.4%. Software development was 8.7%.

That tracks with the work people actually complain about. Not the clean work in the job description. The surrounding mess.

Reconciling the quarter’s spend. Turning a folder of contracts into a renewals tracker. Building tomorrow’s client deck from call notes and pipeline data. Pulling scattered updates into a report. Making the status note nobody wants to write but everyone needs before the meeting.

These are not glamorous AI agent demos. They are the chores that make a week feel eaten.

That is why mobile and web access matters. The user does not always have a clean two-hour block at the desk. They check the shape of work between calls, on a train, while waiting for a bag, or after closing the laptop because life kept moving.

A bad version of this product category is easy to imagine.

The assistant keeps running, then the phone becomes a stream of foggy updates: working on it, found something, need your input, almost done, review this, approve that. The user technically has an AI helper. Emotionally, they have a smaller inbox.

Jun’s test is simpler: when I open the phone, can I tell what happened in five seconds?

Not a transcript. Not a full reasoning trail. A plain review surface. Finished and ready to use. Needs your decision. Drafted but not sent. Blocked because access or context was missing. Safe to ignore until later.

Anthropic’s announcement says decisions still come back to the user and that nothing ships until review and approval. Good. But approval is only calm when the user can see the edge of the action.

If the assistant drafted a client email, show the source threads it used, the sentence that might be risky, and the exact send state: unsent, scheduled, or waiting. If it built a deck, show what changed since the last version and what it could not verify. If it reconciled a spreadsheet, show the unresolved rows and whether it edited the source file or only made a copy.

OpenAI’s guidance for ChatGPT apps makes a related point from the developer side: do not port the whole product into chat. Pick the useful capabilities. Let the system know, do, or show something the base conversation could not. For normal users, that becomes a product promise: do not make me learn your architecture to trust your help.

If you are testing an AI assistant for ordinary work, ignore the biggest demo first. Try one repeated chore. Ask it to turn a folder, thread, or messy note pile into one usable artifact. Then check the handoff, not just the output.

Can you see what it read? Can you see what it changed? Can you tell what is waiting on you? Can you stop, redirect, or leave it in draft mode without hunting through settings? Can someone else on the team understand the result without asking you to explain the whole run?

That is the difference between an AI assistant that gives time back and one that relocates the work to review. The next useful AI agent may not look like a robot coworker. It may look like a boring card on your phone that says: here is what changed, here is what still needs you, and here is what I left alone.