AI Roosevelt made it clear: don't ever call him Teddy

INSIDE INVISIBLE 

⚖️ Thomson Reuters: 2.5x more tasks, annotation framework in under 2 weeks

Thomson Reuters needed AI that could read high-stakes professional documents the way a trained human would. That meant building data structures and an annotation process that taught the model what mattered across complex document layouts like tables, graphs, and form fields.

See how Invisible built the solution that drove a 2.5x jump in task throughput in less than two weeks.

“AI has to do more than generate plausible answers. It has to understand the underlying structure and meaning of complex professional documents.” Joel Hron, Chief Technology Officer at Thomson Reuters

DOWN TO BUSINESS

💳 Nubank's AI agents now perform on par with human support

Nubank's AI agents, deployed across 100 million users, lifted satisfaction scores 37 points in card-delivery status inquiries and self-service resolution rates 29 points, with similar gains in debt management and credit-limit support. Human agents now step in only for complex requests.

Learn more: Invisible's contact center solution helps triage, suggests replies, and handles handoffs so customers get answers fast, and humans stay in control.

🏥 NHS ends the 8am phone rush with AI triage

An AI triage tool built into the NHS App cut phone queues by 29% at a Sussex practice serving 23,000 patients, without replacing clinical judgment. NHS England is expanding the tool to 200,000 patients within a year, reaching all app users by 2028. A related ambient-notetaking trial freed nearly 25% more direct patient time.

📠 Duke Health turns fax backlog into $285,000 in savings

Duke University Health System's cardiology division now automates more than 5,000 faxes a month, freeing up $285,450 in annual staff capacity. The system reinvested the freed time directly into patient care rather than cutting positions. Referral processing that once consumed hours of manual sorting now runs largely unattended.

⚡ Eversource Energy sees outages before they happen

Eversource Energy's AI-driven outage-prediction platform, which merges weather, grid sensor, and vegetation data, prevented 40,000 customer outages within its first two months. Eversource says the platform now flags risk before storms hit, not after.

🏨 Hyatt's group sales teams got 20% more productive with AI

Hyatt's group sales teams have become roughly 20% more productive since deploying AI tools, according to J.P. Morgan analysts tracking the hotel industry's AI rollout. Wyndham reported similar gains, with AI-powered call centers cutting labor costs for franchisees. Analysts call 2026 the first year AI investment in hospitality is expected to show up in earnings, not just pilots.

FROM THE EDGE

🌱 Shlomo Benartzi: AI's real payoff is growth, not efficiency

Behavioral economist Shlomo Benartzi argues in Harvard Business Review that companies fixated on AI-driven cost-cutting are leaving the bigger opportunity on the table. Efficiency gains are capped by what a company already spends; growth gains are not, since AI can reach new customers at near-zero marginal cost. Chasing savings first trades a small win for a much larger one.

🌐 Azeem Azhar: the AI economy is in its early innings

Azeem Azhar's bottom-up accounting of AI spending, built to avoid double-counting customer demand, finds the AI economy already covering its own infrastructure bill, not running at a loss as skeptics claim. The report calls the current scale "small enough to be early." What happens next depends on how fast demand grows as prices fall.

📊 Alexander Bick: AI adoption is bigger than official numbers show

Federal Reserve Bank of St. Louis research by economist Alexander Bick and coauthors finds that measured AI adoption swings wildly based on how firms are asked. A narrow "production use" question yields roughly 10% adoption; asking about use "for any purpose" pushes the figure closer to 45%. The gap suggests enterprise AI use is far more widespread than surveys imply.

HOT MODEL NEWS

🌶️ OpenAI and Broadcom: spicy AI chip cuts inference costs 50%

OpenAI and Broadcom unveiled a custom AI chip, codenamed Jalapeño, that Broadcom's CEO says cuts inference costs roughly 50% compared to standard GPUs while matching Nvidia's current flagship on performance. The chip is set for gigawatt-scale deployment by year end. Microsoft has reportedly committed to buying roughly 40% of the first production run.

🤖 GPT-5.6 ships restricted to 20 government-approved partners 

OpenAI's GPT-5.6 family, branded Sol, Terra, and Luna, is rolling out to roughly 20 government-approved partners under new federal review requirements, with broader availability expected by mid-to-late July. Sol scored 96.7% on OpenAI's internal cybersecurity test, crossing the threshold that triggered the restriction. Enterprise buyers should expect gated access to be the new normal for frontier releases.

🧠 Claude Sonnet 5: Anthropic's new default for enterprise workloads

Claude Sonnet 5 replaces Anthropic's previous mid-tier model as the default for high-volume enterprise workloads, pairing frontier-level coding and reasoning performance with lower cost and latency than top-tier models. Anthropic positions it as the model most enterprises will run in production, reserving larger models for the hardest tasks. Enterprise adopters report faster agent workflows at a fraction of the cost.

🔒 Anthropic: distillation attempts are pushing labs toward tighter API limits

Anthropic says an operation linked to Alibaba's Qwen lab ran 28.8 million exchanges with Claude over six weeks, using roughly 25,000 fraudulent accounts, to extract its coding and reasoning capabilities. Alibaba denies training on proprietary outputs. Expect frontier labs to respond to incidents like this with tighter API monitoring and stricter rate limits going forward.

🔍 Gemini 3.5 Pro: a 2 million-token context window is coming

Google's Gemini 3.5 Pro is in enterprise preview ahead of a broader rollout, headlined by a 2 million-token context window, the largest of any major frontier model, plus a new "Deep Think" reasoning mode. Full benchmarks and pricing are expected at general availability. It is currently the only major frontier flagship not subject to a government access restriction.

PLOT TWIST

🎙️ Theodore Roosevelt is back, and he has opinions

A new AI avatar of Theodore Roosevelt greets visitors at his North Dakota presidential library, insisting he never went by "Teddy" in private and cracking that today's senators could stand to speak the truth too. Trained on hundreds of thousands of archival documents, it happily discusses his life and legacy but refuses to weigh in on modern politics.

💭 Meta's AI decodes thoughts into text, no surgery needed

Meta's Brain2Qwerty research project decodes typed sentences from non-invasive brain scans, without any surgery or implant, using headsets on consenting volunteers. The system reached 61% average word accuracy, and 78% for its best participant. Meta released the code and dataset publicly as part of a broader effort to help people with speech or motor impairments communicate.

🚁 Two hikers went missing. An AI drone found them fast

Two hikers went missing in Australia’s Kosciuszko National Park, but a thermal-camera drone with real-time AI image analysis found them within five hours, not the days a search might otherwise take. They'd wandered about 500 meters off the trail. Fire and Rescue NSW called it a first for the agency.

🐦 A zebra finch translator just won $100,000

Zebra finches have been chirping in code all along, and Julie Elie finally cracked it. The UC Berkeley scientist used machine learning to decode 11 distinct calls and their meanings after more than a decade of recordings, winning the $100,000 Coller-Dolittle Prize. A $10 million bonus awaits whoever proves genuine two-way conversation with an animal, still unclaimed.