.png)
🗺️ See where multilingual AI breaks, dialect by dialect
AI built on English defaults stumbles on honorifics, gendered grammar, formality, and more in other languages. Invisible's new Language Atlas shows where: a linguist-verified database covering 140+ languages and more than 1,700 dialects, each scored on 11 traits that break multilingual quality. Teams use it to pick dialects and check projects before launch.
🌮 Taco Bell's voice AI lifts drive-thru sales 1.5%
Taco Bell's drive-thru voice AI is live at about 900 locations, heading for 1,500 by year end. Where it runs, satisfaction is up 6% and sales about 1.5%, per parent Yum Brands. KFC's new kitchen system cut drive-thru times 8% and refunds 12% in early stores. McDonald's is testing similar voice ordering.
🏥 UTMB cuts specialist wait times 30% with AI triage
The University of Texas Medical Branch (UTMB) runs over 2,000 specialty referrals a month through an AI agent that reads the full chart and recommends a priority and pathway. In early results across four specialties, average waits for a first visit fell 30%, from 93 to 65 days. Those specialties also saw nearly 50% more new patients.
🛒 Amazon's shopping assistant users spend 50% more per order
More than 400 million customers used Amazon's AI shopping assistant in the past 12 months. U.S. customers who use it spend 50% more per order than those who don't, CEO Andy Jassy told investors. The assistant now handles more than 100 tasks, from gift registries and subscriptions to delivery changes and billing questions.
🏦 Westpac's AI halves the cost of a data migration
When moving its data warehouse, Australian bank Westpac used AI to convert old code, run tests and write documentation. The project is now expected to cost about A$35 million (US$25M), half its A$70 million budget, and finish more than two years early. Retiring the old system should save about A$15 million a year.
📋 Manulife's AI now instantly approves 58% of applications
Canadian insurer Manulife's underwriting engine now instantly approves 58% of eligible life insurance applications, nearly double its earlier rate. Across its broader mass-market business, 45% to 50% of applications get an instant decision. The insurer has also launched an AI assistant that helps advisors with life and critical illness products.
🔗 Chad Syverson: AI pays off when it crosses departments
Workers spend 37% of their time on coordination, like pulling data from several systems and reconciling numbers, finds a survey by Chicago Booth's Syverson and Yale's Seth Zimmerman. 76% want less of it, yet most AI tools work inside one system. At Uber, agents built across finance workflows cut one report from two days to 10 minutes.
🎯 Katryn Wright: AI that pushes back gets better results
At UK telecom giant BT Group, 2,400 employees were split into three groups for goal-setting: no AI, a default AI that leaned toward flattery, or a "challenger" AI built to push for more ambitious, specific goals. The challenger produced 3.5 times as many excellent goals as the other two. BT is now building it into goal-setting.
🧭 Katherine Kellogg: aim AI agents at growth, not just cuts
The automation versus augmentation debate misses the point, argue MIT Sloan's Kellogg and Rama Ramakrishnan. Agentic AI should be aimed at workflows that create growth, not only at cost savings. They rank candidate workflows on eight criteria, including strategic importance, technical fit, operational readiness, return on investment and whether the work can be reused elsewhere.
📉 Bart van Ark: keep investing while AI's payoff catches up
AI speeds up individual tasks, but those gains don't automatically reach the whole business, finds the University of Manchester's Productivity Institute. Retraining and reorganizing cost money up front, while the payback comes later. "The ambition gap lies between those individual efficiencies and sustained business results," says lead author van Ark.
🥇 Gemini 4 Carbon tops a leading AI leaderboard
Google's new frontier model, Gemini 4 Carbon, ranks first on Artificial Analysis's intelligence index with a score of 59. Introductory pricing is $2 per million input tokens and $10 per million output, rising to $4 and $20 later. Cyber defenders get access first, with developers and businesses to follow.
🧱 Mistral Large 4: a trillion-parameter model you can run yourself
Mistral has released the full weights of Large 4, its 1-trillion-parameter flagship, so companies can run it on their own infrastructure, with some commercial license restrictions. Through Mistral's API it costs $0.70 per million input tokens and $2.10 per million output. Its cost per benchmark task is 63% below the preview version's.
🛡️ GPT-6 Cyber: OpenAI's exploit hunter for approved security teams
OpenAI's GPT-6 Cyber, available to selected security customers, scored 48.5% on an exploit-research benchmark, up from 32.1% for its predecessor. It stayed within authorized targets in 99.93% of tests without safety prompts, and in every test once they were added. It costs $20 per million input tokens, double Anthropic's comparable model.
🕶️ Websites are now running ads only AI agents can see
Time magazine serves ads from Ally Bank and others on a text-only version of its site shown only to AI agents. Ten more publishers have followed, and ad-tech startup Mobian expects over two dozen by December. AI developers fear the ads could steer agents without users knowing. Perplexity calls them deceptive and now blocks them.
📨 UK councils are drowning in 200-page AI-written complaints
Complaints to UK local governments that once ran about two pages are now mostly ten times longer, says the head of Lawyers in Local Government, with the worst near 200 pages. Chatbots pull in case law "from Canada or somewhere," and sometimes invent it. "Half the emails we receive are some form of AI," says one MP's caseworker.
🐭 An AI agent injected gene therapy into a mouse's brain
After a veterinary surgeon opened the skull, an AI agent picked a gene therapy from 19 options, chose the brain region and coordinates, then steered the robotic needle itself. San Francisco startup Exin Therapeutics required human sign-off on the plan and kept an emergency stop button within reach. The mouse is recovering well.
🎭 AI fakes convinced a fan she was dating Ioan Gruffudd
A woman at a fan convention told the Fantastic Four actor she believed they had been in a relationship for over a year. A scammer posing as him on Telegram had sent her hundreds of AI-generated images and videos as proof. Gruffudd is now urging fans never to send money to anyone claiming to be him.
🏀 The Celtics' head coach runs his game plans past Claude
Boston Celtics coach Joe Mazzulla uses Anthropic's Claude to review his halftime and pregame routines. "It really helps you kind of test your processes that you think are efficient," he told a Boston sports radio station. He has company: one former NFL executive has estimated that 75% of NFL teams use AI in weekly prep.
