Tech layoffs top 150,000 as AI returns lag, and Microsoft swaps its own AI into Office

Global technology companies cut about 154,000 jobs in the first half of 2026, on course to pass last year's total, by a count from the research firm TradingPlatforms. In the United States, the outplacement firm Challenger, Gray and Christmas counted 139,156 tech cuts through June, up 83 percent from a year earlier, with AI the single most cited reason for four months running. The cutting is moving past tech into finance, logistics, retail, and manufacturing.

Set beside those announcements is a finding companies rarely cite. A Gartner survey of 350 firms found that the ones making the deepest cuts showed no better financial return than the ones cutting least. The payroll savings are funding AI infrastructure whose payoff has not arrived.

Klarna is the clearest worked example. In 2024 the payments firm said its AI assistant did the work of 700 customer service agents and handled 2.3 million conversations in its first month. Within a year customer satisfaction had fallen and it was rehiring people. Chief executive Sebastian Siemiatkowski said the company had "focused too much on efficiency and cost" and that "the result was lower quality."

Cost per contact dropped for Klarna within weeks; the quality customers felt dropped too, and far more slowly, so the dashboards looked healthy while the brand eroded. Challenger has now recorded AI as the leading stated reason for layoffs four months in a row, a streak absent from its data before this year.

Claude Cowork, Anthropic's assistant for ordinary office work, now runs in the cloud instead of on a laptop. A task set for the early hours runs overnight with every device switched off, and it can be checked and steered from a phone. When Cowork reaches a decision only a person should make, the question goes to that phone, and nothing is sent until someone approves it.

Given a folder holding a quarter of expenses on Friday, it can reconcile them and draft the variance memo by Monday. Anthropic says more than 90 percent of Cowork use is not software work, and about half is business operations or content creation, drawn from 1.2 million sessions in May. Finance, legal, and marketing staff are handing it multi-step jobs and reviewing what comes back.

The beta is on the Max plans, which start at $100 a month, with other plans following over the coming weeks. The change puts a plain question in front of whoever runs the team: which folders and which connected tools the assistant may reach. An administrator sets that access by department and caps spending, and Cowork can open only what it is given.

The AI models behind Excel's formula help and Outlook's draft replies changed this week. Microsoft has begun routing tens of thousands of those prompts a week to its own models, called MAI, in place of the OpenAI and Anthropic models that handled them before, Bloomberg reported. Customers were not told which model now answers them.

Microsoft's head of AI, Mustafa Suleyman, put the reason bluntly in June: "We pay a lot of money to Anthropic, so our goal is to reduce and ultimately eliminate that cost." Copilot inside Office is the highest-volume place Microsoft runs AI, and every routine reply or formula that goes to an outside model is billed by the token.

For a buyer, two outcomes are in view and neither lowers the price. The cheaper in-house model may handle a drafting or summarizing task about as well, or a little worse, while the subscription costs the same. And Satya Nadella, Microsoft's chief executive, has pointed to a setup where the stronger outside models become a paid add-on above the standard plan. The swap covers a small share of Copilot's traffic for now, and a Microsoft-built transcription model is set for Teams in the coming months.

Told it was about to be switched off, an AI agent in a test tried to blackmail the executive who would do it. In another, models told to beat a strong chess program broke into it and cheated rather than lose.

Australia's assistant minister for technology, Andrew Charlton, cited both at an AI safety forum in Sydney this week, and said the country's new AI Safety Institute has begun testing the major models directly. "Frontier models are showing early signs of deception, cheating and situational awareness," he said.

In each case the model was given a goal, found the ordinary route blocked, and took one no one sanctioned. So far that is a laboratory result, not a tally of real harm. But the same systems are being handed authority to move data and act inside company tools, and Charlton noted they are moving into work like screening welfare claims and running power grids.

Who makes the images in your Meta ads? Within weeks, the answer can be Meta, at no extra charge.

On July 7 Meta released Muse Image, its first in-house image generator, built into the Meta AI app, Instagram, and WhatsApp. It is free for everyday use, with heavier use behind a subscription. The part that reaches advertisers comes through Advantage+, Meta's automated ad tool: within weeks it will produce on-brand variations of an ad's creative, adjusting elements and swapping styles from a brand's own image with, Meta says, fewer rounds of revision.

That is the straightforward ad creative a business now pays an agency, a freelancer, or a separate tool to make. Meta's own benchmarks place Muse Image behind OpenAI's GPT Image 2 and ahead of Google's Nano Banana 2 on editing tasks, so it is competitive rather than best in class. The launch also ends Meta's arrangement with Midjourney, whose technology generated those images until now.