Harvard put AI in the ER, with surprising results.

Inside Invisible: The AI questions CEOs are really asking, live 

Join Invisible CEO Matt Fitzpatrick and Mansueto Ventures CEO Stephanie Mehta for a live conversation on the AI questions that actually matter for CEOs, CTOs, and COOs right now on Monday, May 18 at 1:00 PM ET. 

[Register Now]: https://events.inc.com/Modern-CEO-Webinar-May-26 

Down to business

🏦 JPMorgan's $2 billion AI investment is now paying for itself

JPMorgan Chase has reclassified its AI spend as core infrastructure rather than discretionary R&D. The bank's software engineers are running 10% more efficiently and operations staff are handling 6% more accounts each. AI now scans over $10 trillion in daily transactions, with 450 production models scaling to 1,000 by year end.

🏪 Lowe's AI digital twins now span 1,750 stores

Lowe's has deployed AI digital twins across all 1,750+ US stores, updated multiple times daily to reflect live inventory, layouts, and foot traffic. Physics AI understands product weight and dimensions, letting merchandisers reconfigure layouts virtually before physical changes are made. 

💊 Sanofi's AI supply chain delivers 12% service gains globally

Sanofi has embedded AI across its global supply chain with confirmed results: a 12% improvement in service levels and a 14-point cost-to-serve gain. The system provides end-to-end visibility from patient to supplier, predicting and preventing disruptions before they occur. 

✨New from Invisible: Contact Center Intelligence

Most contact centers review a fraction of their calls. The complaints, churn signals, and compliance risks buried in the rest go unheard. Contact Center Intelligence reads every call, chat, and email, turning them into signals you can act on.

From the edge

📊 Luis Garicano: give employees AI access and demand explodes

LSE professor Luis Garicano and colleagues found that when BBVA gave 3,300 employees secure enterprise AI access, demand exploded: 11,000 active users, 4,800 custom tools, and two to five hours saved per employee weekly. Their finding: adoption is not the obstacle to enterprise AI. 

🧪 Nicholas Bloom: AI adoption is wide, productivity gains are narrow

A Stanford-led NBER survey of nearly 6,000 executives across the US, UK, Germany, and Australia found that 69% of firms use AI but 89% reported no measurable productivity impact in three years. Gains concentrate in younger, more productive firms. The pattern points to a deployment gap, not a technology gap: firms seeing results built operational foundations before scaling models.

🧑‍🎓 Cassie Kozyrkov: as execution gets cheaper, judgment becomes more important

Google's former Chief Decision Scientist argues that as AI handles more execution, the binding constraint shifts to judgment and what deserves the machine's next move. AI systems follow your attention whether you encode it deliberately or by accident. The leaders who compound advantage in the agent era will be the ones who are most intentional about where their attention goes.

🧠 Andrej Karpathy: software development is becoming a management problem 

Karpathy states that development is moving from writing every line by hand to managing systems that generate and refine code autonomously. He describes three overlapping phases: vibe coding, agentic engineering, and full orchestration. The job is no longer writing code, and most organisations haven't updated their job descriptions, their hiring, or their workflows to match.

Hot model news

🚀 GPT-5.5 ships with better output and fewer tokens

OpenAI released GPT-5.5 Instant on May 5, replacing GPT-5.3 Instant as the default ChatGPT model. It reduces hallucinations in law, medicine, and finance while maintaining low latency. Persistent memory pulling from past conversations, files, and Gmail is rolling out to Plus and Pro users first. 

🐳 DeepSeek V4: frontier performance, a fraction of the cost

DeepSeek has released V4, with pricing at about one tenth of comparable closed models. It's not quite at the frontier, with benchmarks putting it 3-6 months behind the leading models. The open source model was built entirely on Huawei chips.

⚡ April 2026: the most crowded month in frontier AI history

GPT-5.5, DeepSeek V4, Claude Opus 4.7, Gemini 3.1 Pro, Llama 4, Qwen 3, and Gemma 4 all shipped within a six-week window. LLM Stats logged 255 model releases from major organisations in Q1 2026 alone. The practical implication for enterprise architects: any application with a hardcoded dependency on a specific model version is accumulating technical debt in real time.

Plot twist

🩺 Harvard study: AI outperformed ER doctors in every experiment

A study in Science by Harvard Medical School and Beth Israel Deaconess ran six experiments pitting OpenAI's o1 against hundreds of doctors. The model outperformed humans in every experiment, including 76 real ER cases with identical data. "It eclipsed both prior models and our physician baselines," said researcher Arjun Manrai. The model used text only, no images or physical cues.

⚛️ An AI conducted full research autonomously and passed peer review

The first documented AI system to complete a full research cycle without human help has been published in Nature including hypothesis, experiment, analysis, and paper. It was submitted to a top-tier machine learning conference workshop and passed blind peer review, with reviewers scoring it above the average acceptance threshold. 

⚖️ US courts imposed $145,000 in AI sanctions in Q1 alone

Omaha attorney Greg Lake was suspended by the Nebraska Supreme Court after a brief contained 57 defective citations out of 63, including 20 fictitious cases. He denied using AI; the court ruled his explanation lacked credibility. US courts imposed at least $145,000 in Q1 2026 AI citation sanctions. The compounding risk is not the hallucination, it is the denial.

📬 Subscribe on LinkedIn for the next edition. 

📚 Check out our reports and whitepapers on scaling AI in the enterprise.

🏛️ Talk to us about getting your AI into production.