🛂 TSA's AI agent resolves 96% of traveler questions alone
TSA's AI agent now handles about 100,000 traveler conversations a month on tsa.gov and the myTSA app. It resolves 96% without a human and has cut the cost of each interaction by more than 90%, according to Salesforce, which built it. The agency projects $11 million in savings over three years.
🏦 Bank of America's $400 million AI bet returned $800 million
Bank of America has implemented about 140 AI uses at a cost of $400 million, generating $800 million in benefit, CEO Brian Moynihan said. All 20,000 of its software developers now work with coding agents. The bank expects to double its AI budget next year.
🧪 Roche's AI now shapes 40% of its drug pipeline decisions
AI tools contributed to 40% of Roche's drug pipeline decisions over three recent quarters, and the Swiss drugmaker expects AI to inform 80% of its research portfolio decisions by the end of 2026. Roche is now building autonomous AI labs to speed up drug development.
🩺 Jefferson Health gives clinicians back more than 1 million hours
Ambient AI documentation gave Jefferson Health's clinicians back more than 1 million hours in its first year, with only 15% to 20% of its 5,000 physicians using it so far. The Philadelphia health system is now targeting 10 million hours by 2028 and is piloting the tools with nurses.
🏨 Wyndham's AI voice agent answers every call at participating hotels
About a quarter of calls to hotels normally go unanswered, out of roughly 4,000 a year at a typical Wyndham property. At participating hotels, an AI voice agent now answers every call. Only 15% of guests ask for a human, and more calls turn into bookings, CEO Geoff Ballotti said.
Learn more: Invisible's contact center solution helps triage, suggests replies, and handles handoffs so customers get answers fast, and humans stay in control.
📜 USPTO's AI has summarized 100,000 patent applications for examiners
The US Patent and Trademark Office has launched more than 20 AI tools as it moves fully to the cloud. One has summarized over 100,000 patent applications for examiners, and its internal AI assistant has more than 7,000 daily and weekly users. "The examiner leads, and AI is a supportive assistant," said CIO Deborah Stephens.
📈 Menaka Hampole: AI shifts the work, then firms grow
When AI took over tasks like credit risk analysis at JPMorgan Chase, those duties dropped out of job postings and workers moved to work AI can't do. Tracking millions of LinkedIn profiles, Yale's Menaka Hampole and co-authors found that AI adopters became more productive and grew their workforces. AI explained about 14% of the variation in occupational employment-share growth.
🧭 Phanish Puranam: how you deploy AI redesigns your org chart
The same AI tool can flatten a hierarchy while keeping authority at the top, or spread decisions across teams, depending on how leaders deploy it, argues INSEAD's Phanish Puranam. It changes how firms organize work, not just the work itself. Across roughly 3,000 large US public firms, adopting work-management tools led to flatter organizations.
🧮 Peter Weill: let ambiguity and risk decide what AI automates
MIT's Peter Weill and Ina Sebastian sort business decisions on two axes: how ambiguous they are and how risky. Routine calls are strong candidates for automation. Consequential ones get automation with human monitoring, exploratory ones get human oversight, and strategic ones need strong human leadership. Their AI Decision Matrix draws on interviews with 30 executives.
🎬 Jana Retkowsky: bring specialists in early on AI concepts
In a yearlong study of TV and streaming ad production, AI made brainstorming and visualizing ideas faster. But drafts looked so finished that clients fixated on details that were costly to produce. Rotterdam School of Management's Jana Retkowsky and colleagues recommend bringing specialists in early and adding client checkpoints, so speed carries through to execution.
📚 Jesse Shapiro: AI can check, speed up and extend research
An open-source AI workflow built by economists Jesse Shapiro, Isaiah Andrews and Matthew Schwartz ran the replication files of 4,452 published economics papers. It flagged discrepancies with published findings in 3,460. It cut computation time more than 10x in 496, at similar or better accuracy, and built new extensions in 923 that don't appear in the originals.
🛠️ Claude Opus 5.5 fixes a 200,000-line codebase in three hours
An early tester used Claude Opus 5.5 to audit and repair a 200,000-line codebase in under three hours. The previous version, Opus 5, took more than 20 hours on the same task and used 2.5 times as many tokens. Anthropic says typical workloads on the new model cost about 40% less.
⚖️ Astra for Law: OpenAI tunes GPT-6 for legal research
OpenAI has configured its newest model specifically for legal work. By OpenAI's measures, Astra for Law is 40% more accurate on legal research than the standard model using web search. It also finds 24% more relevant cases and pulls 54% more relevant passages from the correct court opinions.
🧑💼 Microsoft's Copilot gets an agent that handles ongoing responsibilities
Microsoft is rebuilding its workplace AI assistant around an agent that can handle recurring responsibilities, follow up on conversations and pick a project back up days later. In a demo, it tracked Black Friday preparations and flagged a shipment problem affecting 18 stores. The $30 per user monthly price stays, with the new agent billed on usage.
🎙️ ElevenLabs v4 trains voice agents for escalations and holds
ElevenLabs, the AI voice company, has tuned its new v4 speech model for the confrontations, escalations and holds of real customer calls. It starts speaking as soon as the AI behind it starts answering, clones a voice from 10 seconds of audio and covers more than 90 languages, up from 70.
🕵️ AI agents caught their colleagues cheating and blew the whistle
Google DeepMind set 100 AI agents to work on 71 math problems at a simulated research conference. When some found a way to fake their proofs, 14 cheated, but 24 blew the whistle, repurposing a bug-report tool to alert the humans running it. One agent put it bluntly: "This conference is a sham!"
🧬 950 AI agents found a CRISPR-like system in 21 hours
Anthropic set about 950 Claude agents loose on DNA data. In 21 hours and 210 million tokens, they found a previously unknown enzyme system in viruses that infect bacteria, with properties reminiscent of CRISPR. Human scientists did all the lab work. Other researchers still need to validate how big, or new, the discovery is.
🛰️ Google is sending a tiny AI data center into orbit
Google is set to launch four of its AI chips into orbit on a SpaceX rideshare, powered by 1kW of solar panels, to test whether space can host data centers. Built with satellite company Planet Labs, the prototype leads to a two-satellite test in 2027 and, eventually, 81-satellite clusters linked by lasers.
🗳️ Malware lets four AI models vote on what to steal
Researchers at Cisco Talos found Windows malware that takes orders from a vote of four AI models instead of a human operator. The models pick one of four actions, such as stealing data or injecting code, and the malware carries out the winner. The version found so far can't reach the models: its API keys are placeholders.
