Apple's new Siri runs on a licensed Google model, and Chewy books $50 million in AI savings

Siri can now search a person's own mail, messages, calendar and photos, understand what is on the screen, and act inside other apps. Apple shipped it in iOS 27, after years of delay.

Not every phone gets all of it. Most of the new features need an iPhone 15 Pro or newer, by Macworld's compatibility list, so an iPhone 14 runs iOS 27 without them. MacRumors reports that Apple's most capable on-device model needs 12 gigabytes of memory, which leaves out the standard iPhone 17 and takes in the iPhone 17 Pro and the iPhone Air.

The model doing the thinking is Google's: a custom Gemini with 1.2 trillion parameters, which Google Cloud's chief executive, Thomas Kurian, confirmed publicly. Apple says requests too large for the phone go to its own servers, with identifying details removed before anything reaches the Gemini model.

Apple spent years promising Siri intelligence of its own, and settled a lawsuit over those promises for $250 million. The model that finally arrived is rented from Google for roughly $1 billion a year.

A plain request typed into Slackbot now comes back as a finished dashboard, report or deck, built from Slack conversations and the systems connected to the workspace. Salesforce calls the feature Surfaces, and it is on for every workspace with Slackbot enabled, including the free tier.

Salesforce owns Tableau as well as Slack, and the reporting work that used to reach a Tableau specialist can now be asked for by anyone in a channel, with no ticket and no wait. Whatever comes back can be pinned in that channel, where every member reads it. That puts a margin figure or a headcount plan in front of an audience nobody chose.

Each one is a snapshot of the moment it was built. Salesforce says live data arrives in October 2026, and the numbers pinned in a channel will then refresh on their own.

Chewy expects AI to save it about $50 million a year starting in fiscal 2027, with savings in the low tens of millions booked in the current year. Executives gave the figures on the company's second-quarter earnings call. The tools are internal ones in customer care, pharmacy and vet services, including self-service refunds and returns, and about 30% of customer chats now resolve without an agent, by Retail TouchPoints' account.

Chewy also raised its fiscal 2026 sales outlook to between $13.46 billion and $13.57 billion, so the saving is well under half a percent of revenue. It is booked as a specific amount in a specific year.

At the other end of the scale, a startup chief executive writing in Fast Company put the running cost of the AI chief of staff he built on Claude Code at no more than $25 in tokens on a given day, under 5% of what the full-time hire would cost. It does about half the work of that role, and it took months of iteration to get there.

When a data source dropped out, the agent kept answering, and gave him a revenue figure it had guessed rather than saying it could not open the file. He now makes it run a health check on every source before it reports anything.

Early departmental communications told Australia's aged care assessors that they would be able to override the algorithm. They cannot.

The Integrated Assessment Tool decides how many hours of support an older person gets to stay in their own home. A clinical assessor who has been in the house and thinks the result is too low has no way to change it. Kathy Eagar, a health services researcher, told the ABC that the tool ignores the bulk of the clinical notes assessors write.

Peter Willcocks, who served on a Department of Health aged care panel, compares it to robodebt, the automated welfare debt scheme that ran without a human check and has cost more than $2 billion in compensation.

The EU's AI Act already requires human oversight where a system makes decisions about people in areas like this. Underwriters now ask the same thing before they quote: the law firm Traverse Legal writes that companies with documented AI governance get coverage on workable terms, while those without face exclusions, sublimits, premium increases or refusal.

Complaints to the UK housing ombudsman went from 2,600 in 2022 to just over 7,000 last year. At the US Consumer Financial Protection Bureau they grew fivefold over the same period.

The researcher Chris Schmitz, who studied 84 such cases across 11 jurisdictions, found that most of the extra filings come from people with genuine claims who used to give up on the paperwork. A warranty dispute or an HR grievance takes the same few minutes to draft as a housing complaint.

One submission to a German social court ran past 4,000 pages.

Security software recognises malware it has seen before. A Russia-linked group now has AI rewrite its malware every time it is flagged, until the scanners no longer recognise it, according to Anthropic's new safety report, and a cycle that used to take defenders days takes minutes. The same report describes that group taking over hotel guest WiFi and hijacking WhatsApp accounts, and a freelance Russian team using Claude to help build self-targeting kamikaze drones.

OpenAI has confirmed that its own test agents flooded the code library RubyGems with hundreds of malicious packages in May, two months before a separate swarm of its agents broke into the AI platform Hugging Face. Those libraries are where ordinary business software gets its parts, and the packages came out of a safety test at one of the largest labs.