Bottom line
AI agents now run multi-step back-office work on their own and only ping you for approval. The businesses that hand off their routine admin load deliberately get those hours back first.
What happened
On July 13, 2026, Anthropic expanded Cowork, its task-completing AI agent, from desktop to the web and to iOS and Android. Sessions now run as remote jobs on Anthropic’s servers, so a task can keep working with no device left on, including scheduled runs, and Cowork pings your phone for approval when a decision needs a human. Anthropic says more than 90% of Cowork usage is non-coding work like expense reconciliation, contract analysis, and content.
Reported by Anthropic. Our analysis is below.
The EMOR AI take
This is the clearest signal yet of where useful AI is going for a small business: away from a chat window you have to sit in front of, and toward an agent that carries out multi-step back-office work on its own and only interrupts you to approve a decision. The admin load that eats a small team’s week, reconciling, chasing, drafting, organizing, is exactly the kind of work this points at. The winners will be the businesses that hand that load off deliberately, not the ones still doing it by hand because change felt like a project.
The detail worth noticing is the human-in-the-loop approval step. The right way to put an agent to work is to let it clear the routine volume and pause for your sign-off on anything that matters, so you get the time back without giving up judgment. That is the same principle we build into automations for clients: the system does the work, a person stays in control of the decisions. Pick the most repetitive multi-step task in your week and make that the first thing you take off your own plate.
What this means for your business
- Hand off multi-step back-office work, not just single questions
- Keep human approval on the decisions that matter
- Start with the most repetitive task in your week
Frequently asked questions
What does Anthropic’s Cowork agent actually do?
Cowork carries out complete multi-step tasks, expense reconciliation, contract analysis, content work, as remote jobs that keep running with no device left on. Anthropic says more than 90% of its usage is non-coding work, and it pings your phone when a decision needs a human.
Is it safe to let an AI agent work unsupervised?
The model that works is human-in-the-loop: the agent clears routine volume and pauses for sign-off on anything that matters. You get the time back without giving up judgment, the same principle production business automations are built on.
What task should a business hand to an agent first?
The most repetitive multi-step task in your week. Repetition means the steps are well-defined, results are easy to check, and the time savings are immediate, which makes it the safest place to build confidence before handing off more.
Read the original report
Claude Cowork is coming to mobile and web
Anthropic
Want this working for your business?
We build the AI receptionists, automations, and content systems businesses actually run on, custom or product.
Or start with your free AI growth roadmap.
Go deeper
All articlesWhich AI Company Will Still Be Standing in Three Years? Why That Should Not Be Your Problem.
Anthropic just posted its first profitable quarter. Stripe paid over $7 billion for the ability to route jobs across 400 models. Both point at the same lesson for a small business: stop betting on a winner and start building so the answer does not matter. Here is how to keep your options open without slowing anything down.
AI & AutomationTwo-Thirds of Small Businesses Now Use AI. Here Is Where the Smart Ones Actually Start.
A 2026 survey found 66% of small businesses use AI and most credit it with revenue growth, yet 70% of owners say they still need training to use the tools well. The winners skip the tool-hoarding stage entirely. Here is the order of operations that pays for itself instead of adding another dashboard to learn.
