Organizational Scaffolding

Highlights

Top Insights

In a two-year field study, both a healthcare organization and a law firm provided secure AI environments, training, and encouragement. Yet the healthcare organization reached 141 organization-wide AI solutions, while the law firm ended up with only three; more than 80% of its participating domain experts eventually dropped out. Researchers found no major difference in technology access, AI readiness, regulation, or problem suitability.

The main differentiator was organizational scaffolding that sustains experimentation: ongoing enablement, shared evaluation and learning, recognition and incentives.

Source: Why some organizations turn AI experiments into business value while others quietly fail  (MIT Sloan School of Management)

Top News

1. DeepSeek released V4.1-Flash with native vision and a one-million-token context window.
2. OpenAI launched GPT-Image-2.5 Flare and Sunburst, improving image fidelity and precision editing.
3. JD.com open-sourced JoyAI-EchoWM, an interactive world model that generates synchronized video.
4. OpenAI launched the Agents API in public beta with managed long-running sessions.
5. OpenAI added a Data agent to ChatGPT Work that investigates connected company data.
6. Alibaba launched QoderWake 1.0, enabling one-line creation of AI workers.
7. Salesforce expanded Agentforce with job-ready agents for sales, service, commerce, and workforce tasks.

Additional Insights

1. Building confidence for 2027  (IDEO)
Planning for 2027 should focus less on predicting an uncertain future and more on making deliberate choices with confidence. Its Planning Canvas offers a simple three-step process: reflect on what has changed and what you’ve learned, define the impact you want to create in the near future, and then decide what to double down on, reform, stop, or pursue as a new bet. The tool is designed to help individuals and teams enter annual planning with a clearer point of view, shift resources away from outdated priorities, and have more productive leadership conversations about where to focus.

2. From Prototype to Production – Developing Enterprise AI Scaling Practices  (California Management Review)
Enterprise AI succeeds not by scaling prototypes faster, but by building the organizational, technical, and governance capabilities needed for production from the start. Leading adopters first encourage broad, bottom-up experimentation, then concentrate resources on a small set of high-value, repeatable use cases rather than proliferating isolated agents; they establish trust through stakeholder-designed workflows, phased “crawl-walk-run” deployments, clear performance metrics, human oversight, and continuous monitoring; and they treat high-quality, connected data plus risk-based governance as foundational infrastructure developed alongside AI rather than prerequisites to finish first. At scale, governance should be modular and proportional to risk, with greater autonomy for low-risk applications and stronger controls for consequential ones. Finally, successful firms move away from purely centralized AI teams toward tiered, hub-and-spoke operating models, where a central hub owns strategy, infrastructure, governance, and major enterprise projects while embedded domain teams own specialized agents and ongoing adoption. The broader takeaway is that scaling AI is ultimately an operating-model transformation: value comes from coordinating use-case selection, trustworthy system design, data and governance, resource allocation, and organizational change as one integrated capability.

3. Formalizing Fermat’s Last Theorem  (Anthropic Research)

Anthropic reports that Claude produced the first complete, computer-checked formalization of Fermat’s Last Theorem in Lean, working largely autonomously for 11 days and generating roughly 13 million lines of code and proofs for about 30,000 intermediate theorems. The key breakthrough was not new mathematics—the formalization follows a streamlined version of Wiles’s proof—but dramatically faster **verification**: Lean checks every logical step, addressing the growing challenge of validating increasingly complex human- and AI-generated mathematics. Anthropic credits a multi-agent setup plus Prove2Me, which organized dependencies between theorems, enabled parallel work, improved search/reuse, and prevented agents from losing track of project state. The broader implication is that AI may make large-scale formalization of existing mathematics practical, helping uncover errors, reduce the burden on human referees, and provide a trustworthy verification layer for future AI-generated results; Anthropic expects formal proofs to increasingly accompany human-readable mathematical papers rather than replace them.

4. The Einstein test: what happens when AI tries to rediscover relativity?  (Nature)

The article explores an “Einstein test” for AI: train language models only on knowledge available before landmark discoveries and see whether they can independently rediscover ideas such as relativity, quantum mechanics or other scientific breakthroughs. Early experiments suggest today’s models can sometimes produce intriguing hints or recombine existing concepts in novel ways, but they struggle with the kind of abductive reasoning, sparse-data inference and coherent world-model building that underlies major scientific paradigm shifts; for example, models trained on planetary data often generate different approximate laws rather than inferring Newtonian gravity. Researchers also face practical problems such as historical training data contaminated with later knowledge, poor digitization and limited data volume. The broader conclusion is that current AI seems stronger at prediction, verification and incremental scientific work than at Einstein-level conceptual leaps: it may generate many plausible hypotheses, but distinguishing a profound new theory from a convincing-sounding wrong one remains a central challenge. Still, successes in mathematics show that models can occasionally form genuinely useful new abstractions, suggesting that future systems built around stronger world models and reasoning mechanisms could move closer to transformative scientific discovery. 

5. How AI is rewriting the decisions that leaders need to make  (World Economic Forum)

The article’s core message is that AI is changing leadership not just by improving answers, but by influencing how problems are framed before decisions are made. As organizations increasingly use AI upstream—for gathering information, identifying opportunities and defining strategic options—leaders face two risks: frame compression, where fast AI-generated interpretations reduce internal debate and exploration, and frame convergence, where organizations using similar models develop similar assumptions and blind spots. The article argues that simply keeping a “human in the loop” or writing better prompts is insufficient if humans are only reviewing choices that AI has already defined. The emerging leadership advantage is therefore the ability to retain control of the framing loop: questioning the problem before seeking answers, deliberately exploring alternative interpretations, challenging assumptions, and treating AI’s first plausible framing as a hypothesis rather than a conclusion. In this view, preserving an organization’s ability to define its own strategic agenda becomes a form of organizational sovereignty. 

Innovation Radar