As AI loses its novelty, leaders are starting to ask of it the questions they already ask of their software vendors, and some of the answers mean rethinking security, program management and even insurance. This issue starts with insurance, then covers vendor rules and prices worth checking before year-end.
Most AI risk is neither clearly insured nor excluded - Eight US states sued OpenAI over ChatGPT's safety, and the FTC sent document demands to OpenAI and Anthropic. These are allegations and inquiries, and no court has ruled. Meanwhile the broker Aon estimates that more than 90% of AI-related risk is "silent": it sits under policies that neither clearly cover nor exclude it, so whether a claim gets paid is more likely to be disputed. Insurers are moving in opposite directions, with ISO introducing optional generative AI exclusions and Munich Re selling cover for errors by AI models. Before your liability, cyber and directors-and-officers policies renew, ask your broker in writing whether each one covers losses caused by the AI tools and agents you use.
Claude's usage rules change on November 12 - Anthropic's updated policy treats repeated, pointless cruelty to Claude as a rule breach in extreme cases, and states its surveillance rules more plainly: anyone building on Claude may not use it to track people without their consent or to build surveillance tools. Anthropic says enforcement hasn't changed. Ask your HR and security software vendors whether their products use Claude and whether any feature tracks employees or customers without consent.
Gemini 4 Argon doubles in price on December 1 - Google's newest model moves from $2 and $10 to $4 and $20 per million tokens of input and output on its standard paid tier. If a team is testing it, budget at the new price.
Borrowing to build AI is getting more expensive - The cost of insuring CoreWeave's debt against default rose above 10% a year, and Anthropic offered Blackstone 10% interest on part of its chip financing, the Financial Times reported. Providers that borrow heavily may pass some of that on through prices and usage limits, so before signing a multi-year AI contract, ask which prices and limits are fixed for the term.
Also in this issue:
- News Bits - Anthropic replaced Git after AI-written code changes grew more than tenfold in four months; VentureBeat surveys show more companies keeping a person reviewing the changes agents make to live systems; Andrej Karpathy estimates only about 20 million people use frontier models for demanding work.
- Interstitials / Overheard - A chat with Claude that goes from code review to damage control.
- Meanwhile... - A Toronto surgeon's AI shows green "safe to cut" and red "don't" zones during gallbladder surgery, and the surgeons still keep the final call.
- What I'm Consuming - Peter Yang on an agentic engineering workflow, The Economist's Callum Williams on why AI hasn't changed the economy yet, Satya Nadella on treating models as insider risks, Harvey's policy engine for the tools its agents use, the FT on AI tokenomics, WIRED on AI and science, and MIT Sloan on competing with your client's AI.
- After Hours - Michael Connelly's The Black Echo.