🎻 Unified patron records across three systems, in under three weeks
Orchestra of St. Luke's, which plays Carnegie Hall and Lincoln Center, had its donor, patron, and ticketing data in three disconnected systems, leading to duplicate and fragmented patron profiles. For an organization where donor relationships are the core of the business model, that gap carried real consequences.
In under three weeks, Invisible unified the data with AI-matching of records and staff reviewing merges. Identity match accuracy rose 10% and drove donations from previously unreachable patrons. Learn more about the success from President & Executive Director, James Roe:
🛒 Walmart's AI shopping assistant lifts order values 40%
Shoppers who use Walmart's AI shopping assistant spend more: it is driving 40% higher average order values, Walmart U.S. CEO David Guggina said. Weekly engagement with the assistant grew 60% quarter over quarter. Searches that start from a photo instead of typed text convert 57% better.
🏦 TD Bank cuts mortgage processing from 15 hours to minutes
TD's first agentic AI application now handles every new personal mortgage. Average processing time has fallen from 15 hours to about three minutes. The bank reports $195 million in AI value in nine months of its fiscal year, against a $200 million target, lifting it to 10th in Evident's global ranking of banks' AI maturity.
✈️ Booking Holdings' AI keeps cutting customer service cost per booking
AI voice agents now answer most eligible inbound traveler calls at Booking Holdings. Customer service cost per booking is falling at a double-digit rate as bookings grow, and customer satisfaction has risen. Its roughly 9,000 engineers also get about 30% more code into production with AI.
🫀 Houston Methodist adds 7% surgical volume without new operating rooms
Houston Methodist deployed image-based AI that automatically tracks key events across its 15-room cardiothoracic operating suite. Over the next 10 months, the hospital was able to see an additional 25 cases each month, a statistically significant 7% gain measured against a synthetic control. Faster turnovers let more procedures fit into the same operating-room hours.
Learn more: Invisible's computer vision solution turns video footage into structured data, tracking timing, execution, and adherence in real time, with your data staying under your control.
🗺️ Rembrand Koning: map your workflows before you deploy AI
Koning and INSEAD's Hyunjin Kim gave 515 startups the same AI tools and training. Those also shown how AI-native companies redesign workflows found 44% more AI use cases, completed 12% more tasks and were 18% more likely to land paying customers. They earned 1.9 times the revenue and sought 39.5% less outside capital.
⚖️ Arvind Karunakaran: sell AI as enrichment, not productivity
Two divisions of one law firm got the same AI tool. One manager pitched faster drafting. The other asked paralegals which tasks bored them, set aside weekly time to experiment and offered training. Paralegals in that division used the tool 58% more and experimented 70% more, then moved into legal research and case analysis.
🥊 Monideepa Tarafdar: make AI argue with you
Knowledge workers who asked AI to challenge their ideas did better work, according to Tarafdar's UMass Amherst team, which studied 45 professionals. One marketer had AI play customers who would dislike an accounting certificate program and learned buyers wanted cases from more industries. An attorney used it to surface obscure legal loopholes.
💻 Alex Zhang: AI coding tools differ mostly on cost
Most leading AI coding tools break down and run work the same way, argues MIT researcher Alex Zhang, so the real difference is cost. His recursive language models point to where the gains are: trained only on short tasks, they carry the same strategy to tasks many times longer.
🧠 Muse Wonder: Meta's newest model undercuts Claude Opus per task
Meta's new frontier model matches Claude Opus 5.5's list price of $4 per million input tokens, but costs $4.52 per task to Opus's $5.98 on Artificial Analysis benchmarks, where it scores 52 to Opus's 58. Meta's consumer Muse app, running an earlier model, hit an estimated 5 million downloads in 22 days. ChatGPT took 56 days.
☀️ GPT-6.1 Sol: near-flagship performance at one-fifth the price
OpenAI's new model scored 52 on the Artificial Analysis index, one point behind flagship GPT-6 Astra, at one-fifth the token price. Running the benchmark cost $0.72 per task versus $3.26. OpenAI also launched always-on agents, each with its own cloud computer and access to more than 4,000 apps.
✅ Decision models: AI that answers yes or no for pennies
OpenAI, Microsoft, Perplexity and Cloudflare have launched models that return a choice, score or yes/no with a probability attached, instead of paragraphs. OpenAI's charges $0.10 per million input tokens and nothing for output. Perplexity's costs $0.017 per 1,000 decisions. TypeSafe, the startup that opened the category in September, has served about 1.05 billion requests.
📬 Gemini Agent: an AI coworker with its own inbox
Google's new enterprise agent can get its own email address, calendar and file storage, and works across Google Workspace, Microsoft 365 and Slack. It picks whichever approved model fits each task on cost and performance, including Anthropic's Claude. Admins set spending limits that pause the agent when budgets run out. It is in private preview.
🎮 AI clears World of Warcraft's first zone without seeing it
Given one instruction, an OpenAI GPT-6 Astra agent created an Orc character and cleared the starting zone in 40 minutes, reportedly without dying. It never saw a rendered frame. It read the game server's network traffic, pulled quest data from server files and wrote its own route-planning code, all on a private test server.
💊 Utah lets AI write acne prescriptions for $4.99 a month
Utah has approved what is believed to be the first US pilot letting an AI issue initial prescriptions. Nolla Health's dermatology app picks from a short list of approved treatments, starting at $4.99 a month. Doctors approve every prescription for the first 100 patients, then shift to after-the-fact review if the AI hits 95% agreement.
🏎️ Red Bull's AI told Verstappen to give back a position
Red Bull Racing's Oracle-built AI agent started as a Formula 1 rulebook search tool and now recommends race-day actions. Last year at Monza, Max Verstappen cut a corner while defending, which risks a penalty. The agent advised giving the place back, against the team's instincts. The team complied, and Verstappen passed his rival without penalty.
🐈 Kazakhstan enlists outdoor cats with cameras to find lost pets
A Kazakh Telegram service uses AI to match photos of sighted animals against missing-pet reports, and its organizers say it has reunited 407 pets with their owners. Now it is fitting free-roaming cats with tiny cameras and GPS trackers in neighborhoods with the most missing-pet reports.
