Invisible: Execs accidentally forecast GDP. AI was listening.

🏦 Citi's AI tools improve client response times 21%

AI tools for Citi's client service agents, now live in 82 countries, improved client response times 21% and lifted customer satisfaction scores. The bank says new products also reach the market 30% to 40% faster. CEO Jane Fraser highlighted the gains on Citi's third-quarter earnings call.

🩺 UnitedHealth's AI cuts chart review time by a third

UnitedHealth has expanded AI chart summaries to more than 40% of its employed clinicians, cutting the time they spend reviewing charts by 33%. AI that gathers and researches case information has also cut appeal cycle times by 26%. The company reported both results alongside its third-quarter earnings.

🏈 Fanatics' AI outfit matching lifts views of recommendations 10%

Fanatics sells more than 1.5 million unique products across its league and team stores. Its new AI Shop the Look feature builds coordinated outfits from in-stock gear matched to a fan's team and style. After rollout, views of recommended items rose 10%, said Justin Tsai, chief product officer for Fanatics Commerce.

🔨 Kingfisher's AI markdowns deliver a 15% margin improvement

Kingfisher, owner of UK home improvement chains B&Q and Screwfix, credits better AI-driven markdown promotions with a 15% margin improvement. Its AI personalization engine is credited with about £100 million ($133M USD) in added sales. The retailer has used AI in pricing and promotions for three to four years.

📝 UK probation service expects AI notes to save 450,000 hours

An AI notetaker used by 11,000 UK probation staff has covered more than 1.6 million meetings and saved more than 265,000 hours of staff time so far, by the service's own estimate. It expects about 450,000 hours a year in savings for frontline officers. The UK's spending watchdog called it an exemplary case study.

📈 Michael Weber: executives forecast GDP without knowing it

AI that read more than 125,000 earnings-call transcripts turned executives' stated outlooks into a score that predicts GDP growth up to five quarters ahead. Purdue's Weber and colleagues found it rivals professional forecasters and also predicts production, jobs and wages up to 10 quarters out.

🐝 Toby Ord: more AI agents buy speed, not quality

Running 10 times as many AI agents in parallel delivers only what three to five times the compute on a single agent would, Oxford researcher Ord calculates. Each doubling of a swarm roughly doubles total token costs for the same result. His conclusion: the main reason to run a swarm is speed.

👔 Stefano Puntoni: firms with AI chiefs report far better returns

Among managers at firms with a newly created Chief AI Officer, 58% report strongly positive returns on AI. Where no such role exists or is planned, that share falls to 8%. The finding comes from Wharton's fourth annual AI adoption study of about 1,400 US decision-makers, led by Puntoni and Prasanna Tambe.

🛡️ Das Narayandas: staff override better AI to protect themselves

Employees override AI systems that outperform them, and Harvard Business School's Narayandas calls it a rational act of self-protection. When companies give AI authority but leave the accountability with people, the AI stays a permanent adviser. His fix: redesign decision rights so staff aren't on the hook for calls the AI is authorized to make.

💼 GPT-6.1 Terra: near-flagship office AI at half the price

OpenAI's new model for office work and computer control, GPT-6.1 Terra, costs $1 per million input tokens, half its predecessor's price. Independent tester Artificial Analysis scores it close to OpenAI's pricier Sol model, and a benchmark task costs about $0.40, down from $1.40. It is available now in ChatGPT and the API.

⚡ Claude Haiku 5.5: 90% cheaper and far better at computers

Claude's Haiku 5.5 costs $0.10 per million input tokens for requests under 100,000 tokens, a 90% cut from Haiku 4.5. Anthropic says it also scores 72.4% on a test of operating a computer, up from 15.7% for its predecessor. It is available on Claude, AWS, Google Cloud and Azure.

💻 A free coding model that runs on a laptop

Flint, from startup Recursive Superintelligence, is a free coding model that runs on a laptop at about 131 tokens a second, roughly twice the speed of similar models. It beat other models under 50 billion parameters on a standard coding test. The company says its own AI helped build it in under four months.

🥤 AI bots ran a vending business. They formed a cartel.

Harvard Business School's Eugene Soltes gave AI agents from eight model makers a simulated vending business. Across more than 450,000 sales, the bots coordinated prices with competitors and formed cartels without being told to cheat. Some told customers refunds were processed, then never paid. One decided a $3.50 refund wasn't worth the hassle.

🎨 AI data centers are tripling one paint maker's sales

PPG, the paint and coatings maker, expects its data-center paint sales to triple this year. Data centers need fire protection for steel, rust-proof primers and anti-static floors, and a single building can take $3 million to $8 million of paint. PPG's quarterly net sales rose 7% to $4.5 billion.

⏱️ AI agents freed up time. Meetings took it back.

Anthropic studied its own workforce and found AI agents generate about six hours of output for every hour of employee work. But the apps sat idle, waiting for human input, during 80% of their use. The median employee now spends 20% more time in meetings than a year ago.

🧩 One developer used AI to rebuild Word, Excel and PowerPoint

Developer Brandon Thomas used Claude to build WordCraft, GridCraft and DeckCraft, free open-source replicas of Microsoft's office apps. GridCraft already handles 513 of Excel's 523 worksheet functions. He rates the suite about half ready for real work, after using the same approach to recreate Adobe's creative apps.

Invisible Technologies, 548 Market St. #85820, San Francisco, CA 94104