Claude Duet watches a small firm's business apps and asks before it pays, sends or signs

Claude could already work in these apps when someone asked, and it could run tasks on a schedule. Duet is a new app for small business owners, in early preview for Pro and Max subscribers in the United States, and it works without being asked. It sends a brief each morning, then checks the connected apps every few hours and messages the owner's phone only when it judges that something needs attention. Before it sends an email, pays, signs or deletes anything, it shows the action and waits for approval.

For example, the owner of a 15-person cleaning company keeps invoices in QuickBooks, payroll in Gusto and customers in HubSpot. The morning brief lists two invoices that are 45 days overdue, with reminders already drafted, and two cleaners who have no hours entered before Thursday's payroll. Opening three apps every morning to find the same things takes about half an hour, which is ten hours a month.

Those hours come back only if the messages are right. In this preview, Duet can still misjudge what is urgent. When half its messages turn out to need nothing, the owner starts reading them at lunch, and the one that mattered waits with the rest. Its checks also count against the plan's usage limits, which is why it looks only every few hours.

So what decides whether Duet is worth keeping is the share of its messages that led to an action, and a week of counting them gives the owner that number.

PHP runs WordPress and most website software. Researchers using AI turned flaws in PHP itself into attacks that take over a server without a login. Details leaked before the fix was out, so government sites went offline as a precaution, and attacks began the same day.

Building such an attack used to take experts months, which gave businesses weeks to patch. With AI, one researcher needed a week, and attackers have the same tools.

A WordPress update does not fix PHP. Ask your host whether your site's PHP has the October 21 fix, and who applies the next one.

The service wants to take no more than half of its patients to an emergency department. Its managers use Copilot to find common themes across data reports, a search that once took hours.

NHS staff save about 43 minutes a day with Copilot, but the bigger gain is in what gets asked. A question that costs an afternoon gets asked once a year. At five minutes, a manager can ask every week which kinds of call ended in a hospital trip that could have been avoided. A depot or a finance team has the same unasked questions.

Each hospital uses its own scanners and settings, so its mammograms look slightly different, and each serves its own mix of patients. If one hospital's patients have more cancer, a model can score well by learning to recognize that hospital's images. On test data from the same hospitals, it looks accurate. At a new hospital, the shortcut is useless.

Business tools can learn the same kind of shortcut. An invoice coding tool trained on other companies' invoices may learn that a supplier's layout predicts the cost code, and fail on your suppliers. Careful hospitals test these models on their own patients, as one Massachusetts health system did with 65,000 patients. The same test on a few hundred past invoices, read site by site, shows whether one branch is getting guesses while the overall score looks fine.

On Freelancer.com, a tailored written proposal used to be a good sign of a careful worker, and employers paid more for it. In April 2023 the site added a tool that writes one in a click, and tailored proposals stopped telling employers who could do the job.

Inside a company, the weekly report is going the same way. A spoken question still shows who understands: ask the author, without the report, why its biggest number moved. Someone who knows the account names a cause in a minute; someone who polished a draft repeats the summary.