Claude cuts a drug maker's record checks by 89%, and Apple's office apps open to outside AI agents

Before a batch of a new drug can be given to patients in a trial, a quality team reads its whole production record and checks every step against the approved recipe. At Enveda, a US drug developer, four reviewers needed almost 26 hours for four records: 1,061 scanned pages, mostly handwritten. With Claude reading first, the same review took under three hours, including the time people spent judging what it raised.

Claude found all 11 problems that the reviewers had found, plus seven real ones that they had missed. One pass over all four records cost under $40, at about 3.5 cents a page.

The design is as useful as the score. Claude only comments: it cannot change the record, and every batch is still approved by Enveda's quality team. Because people still make the decision, the tool stayed outside the formal validation that regulated production software must pass, so the hours came back without a rebuilt process.

Most companies have checking work of the same shape: a document compared, line by line, with the version it should match. For example, a procurement team that reviews 30 supplier contracts a month against its standard terms might spend three hours on each. If AI marks every clause that differs and a buyer decides which ones matter, each review takes about 30 minutes, and the team gets back 75 hours a month, about two working weeks.

Since October 2025, technicians at Rockwell Automation's factory in Singapore have used the assistant to fix faults on hundreds of machines. Engineers with 20 to 30 years on the shop floor wrote down the fixes they knew, and the assistant sorts them by symptom, cause and action. A new technician types the error into a tablet and sees the likeliest causes first, instead of searching a manual or looking for a senior colleague.

The model underneath is GPT-5.4-mini, one of OpenAI's smaller models. The assistant's value came from the hours that veterans spent writing down what they know.

Gmail and Outlook users could already link their mail to ChatGPT, and now iCloud users can too. ChatGPT can search and read that mail and send messages, and it can add and change calendar events.

It connects with an Apple app-specific password: a separate password that you create on Apple's account website and give to ChatGPT instead of your real one. Deleting it cuts ChatGPT off and leaves your main password alone.

Connect the calendar first. A meeting at the wrong time is easy to spot and move; an email that went to a customer cannot be taken back.

When an auditor questions a spreadsheet figure, they can open the file and recalculate it, because the same formula always gives the same result. An AI model does not work that way. The same question asked twice can get two different answers, and the model behind a tool changes from one version to the next. A bad-debt estimate that AI helped prepare in March cannot be reproduced at the year-end audit.

So the evidence has to be created at the time: what the AI was given, what it produced, and who reviewed and approved it. For a finance team, that can be three extra columns beside each figure that AI helped produce: which tool, who checked it, and what they changed.

On Macs running macOS 27, the apps support MCP, a common standard that lets an AI assistant read a document, edit it and carry out tasks inside another program. Apple is also adding its own assistant, built on Apple and OpenAI models, for paying subscribers.

So the AI that edits a company's slides and spreadsheets no longer has to be the one Apple chose. A company that already pays for an assistant that supports MCP can point it at those files. That puts Apple's paid assistant in competition, inside Apple's own apps, with the AI its customers already have.