ChatGPT answers with working calculators, and Claude starts editing inside Google Docs

Ask ChatGPT whether peak season is cheaper with overtime or agency temps, and the answer can arrive as a small tool with boxes for hours, wages and rates that recalculates every time you change one. ChatGPT already had interactive charts and ready-made modules for more than 70 maths and science topics. GPT-6 builds a tool for whatever question you bring, and decides by itself when an answer should be one. It is live on paid plans and reaches free accounts from today.

A written answer hides its assumptions inside one verdict, while a calculator turns each of them into a box that you can change.

For example, a warehouse needs 600 extra hours over the peak. Overtime at $30 an hour, time and a half on a $20 wage, costs $18,000. Agency temps at $26 an hour look cheaper, until you enter that new pickers work at about three quarters of a trained picker's speed. The same work then takes 800 temp hours and costs $20,800. Drag the agency rate down, and the answer flips at $22.50 an hour. That figure is what the manager takes into the call with the agency, and a paragraph would never have shown it.

A calculator looks just as confident when its formula is wrong. Before relying on one, enter last year's peak, where you know what the overtime really cost, and check that it gives the same figure.

A sidebar opens next to the document, sheet or deck, reads what you have selected and makes the change in place. It works on every paid plan. Claude already worked this way inside Excel, Word and PowerPoint.

For a company on Google Workspace, Gemini came built in, so using another assistant usually meant copying text between browser tabs. Now a finance team that prefers Claude can use it in Sheets while everyone else stays with Gemini. The assistant becomes a choice that each team makes, separate from the office software that the company chose.

In Microsoft's demonstration, Copilot found tax documents scattered across the Downloads and Documents folders, then renamed and zipped them with models running on the laptop. The features reach Copilot+ PCs in the coming months, and Copilot will still use the cloud when a task needs a larger model.

The documents that staff most want help filing, such as contracts, payslips and supplier invoices, are often the ones that company rules keep away from cloud AI. Laptops without the Copilot+ label will not get the features, so for many companies the next laptop purchase becomes an AI decision as well.

By 2013, doctors at two large hospitals were overriding 88% of the drug allergy warnings in their records software. More than half of the overrides gave the same reason: the patient had taken the drug before. Each needless alert taught doctors that alerts were needless, and in the end even warnings about a known, life-threatening allergy were overridden about three times in four.

An AI tool that checks invoices or orders can fall into the same pattern, however accurate it is. The measure that counts is how often its flags are dismissed, because every dismissal teaches users to dismiss the next one.

Most contact centre AI still stops when the conversation ends, and anything left open goes back to the customer or into a human queue. Dialog, part of Cisco's Webex AI Agent platform, coordinates with staff and company systems until the case is resolved, for example chasing a refund through billing and telling the customer when it arrives.

That turns one answer into a long chain of steps, and errors multiply along a chain. An agent that gets each step right 95% of the time finishes a 20-step case without a mistake only 36% of the time. After the chat, nobody is watching, so a wrong step shows up when the customer calls back. The number of steps that an agent takes without a person checking decides its success rate more than the quality of any single step.