Latest
This week's notes.
Assume the Injection Will Work
OpenAI's security team told insurers at Guidewire Connections to assume prompt injection will succeed — and build deterministic controls around the agent anyway. Claims is where the attack surface is the workflow itself.
Nº 014 · Oct 30, 2026The $100 Billion Is Not a Headcount Cut
Zywave prices AI in distribution at more than $100 billion a year — not a layoff number, a rebuild number. The broker's economics are being replaced, not the broker.
Nº 013 · Oct 30, 2026The Examiner Already Has the Questions
The NAIC's AI Risk Evaluation Supplement v6.0 is the questionnaire an examiner will bring into your building. The inventory, the mandate questions, and the proxy question are already written down.
Nº 012 · Oct 30, 2026Who Carries the Token Risk?
IBM will pass token consumption into the client's bill from November. Tokens behave like claims — frequency times severity — and no carrier has reserved for them yet.
Nº 011 · Oct 28, 2026The Core Just Shipped the Agents
Guidewire shipped agents inside the core and paid $1.7 billion for an agentic underwriting platform. The question is what you still own once your core vendor ships them.
Nº 010 · Oct 27, 2026The Policyholder Is Now an Agent
Delaware wants to let an AI agent run an entire company — hold assets, sign contracts, sue and be sued — with the backers shielded from liability. The policy becomes the accountability, not the backstop behind it.
Nº 009 · Oct 26, 2026The Return Is in the Rebuild
AI is booming and productivity isn't. The token bill is $12.50 per employee per month — a rounding error. The return comes from rebuilding the journey, not buying the next tool.
Nº 008 · Oct 24, 2026Who Pays When the Agent Goes Rogue?
The insurance market just split on AI risk: specialty carriers writing affirmative AI cover while the largest carriers exclude generative AI from general liability. Both halves are rational. Only one half is learning.
Nº 007 · Oct 23, 2026The Wrapper Is the Moat
A 21-year-old founder says services businesses win the AI decade. In insurance he's right — with one correction: the service wrapper isn't packaging. It's accountability, and accountability doesn't commoditise.
Nº 006 · Oct 23, 2026Access Is Not Authority
Chubb's CEO wants certification standards for AI agents; a market architect wants verifiable agent mandates. Insurance's oldest discipline — governing delegated authority — is about to be extended to software.
Nº 005 · Oct 22, 2026The Other AI Exposure
A vendor tagged $45.92 billion of AI exposure inside life insurers' investment portfolios — 70% of it in 20 names. Carriers govern AI in operations and price it in underwriting. Nobody had inventoried the asset side.
Nº 004 · Oct 21, 2026From Speed to Decisions: What Berkley's AI Numbers Actually Say
A carrier CEO put numbers on AI — 20%+ intake efficiency, more than half of claims settling at $5,000 or less, and AI priced as an exposure. The next mountain isn't speed. It's judgment.
Nº 003 · Oct 20, 2026The Two Decisions Every Carrier Has to Make About AI Agents
Cover the risks AI agents create, and serve the customers who arrive through them — both decisions turn on the same missing piece: authority. Drift is deciding for most carriers.
Nº 002 · Oct 19, 2026The Exclusion Reflex
Verisk is weighing agentic-AI exclusions while 90% of AI risk sits silent in the book. The carriers that underwrite it on purpose get a growth line; the rest get the dispute line.
Nº 001 · Oct 18, 2026Why I'm writing every day
The insurance AI conversation is dominated by vendors. This notebook is the operator's counterweight — one note a day, from the other side of the table.
The creed
Three rules for this notebook.
No vendor pitches.
Nothing here sells software or services. If a vendor is mentioned, it's because carriers need to know something — not because anyone paid.
No hype cycles.
Every note has to survive the question: what does this change for a carrier on Monday morning? If nothing, it doesn't get published.
No invented numbers.
Every figure is sourced or labeled as judgment. If I don't know something, I'll say so — in writing, the next day if I get it wrong.
The book
The Insurance AI Operating System
Carriers don't need more AI pilots. They need the operating system: AI that survives contact with production, with every consequential decision named, owned, and defensible. I'm writing the book in public — the essays are its first draft.
Follow the bookGet the daily note.
One short essay, every morning. Operating models, governance, economics — the insurance AI conversation without the vendor filter.
Free. Unsubscribe anytime. Your address is never shared.