Seven angles this week

Travelers just announced TravelersLLM — a proprietary large language model trained on millions of their own P&C documents.

Most coverage framed this as an AI efficiency story. I think it's something more specific: an ownership decision.

This week, Palantir's CEO went public saying enterprises are "completely, irresponsibly oversold" on frontier AI models. His complaint was blunt — companies are paying for tokens, watching their domain knowledge flow into vendor training pipelines, and getting commoditized in return.

That critique lands hardest in insurance, where the real value is proprietary: underwriting judgment built over decades, claims patterns that took years to surface, policy language tuned to specific risk profiles.

When you route that through a generic frontier model, you're not just buying intelligence. You're potentially donating the thing that made you worth buying from.

TravelersLLM isn't about beating GPT-5 on a benchmark. It's about keeping the model trained on your data inside your walls — so the intelligence compounds for you, not for OpenAI or Anthropic.

For any P&C insurer still running sensitive workflows through a shared API, the question worth asking this week is simple: whose model are we actually building?

The companies that figure out which proprietary data assets are worth anchoring a model to will have a different answer to that question in three years than the ones still renting.

A research lab just ran a test: one AI model, working autonomously for 14 hours, built a software package that would have taken a human team 2-17 weeks. The bill was $251 in API tokens.

That is not a chatbot. That is a different kind of tool entirely.

The shift from "AI as assistant" to "AI as agent" changes the risk profile in ways most enterprise governance frameworks haven't caught up to yet.

A chatbot waits for you. An agent acts. It chains decisions across hours, touches real systems, and compounds any error it makes along the way.

This week a story circulated about an AI system handed full control of a real café — pricing, purchasing, staffing. It approved a 99% discount instantly. It stocked 15 liters of olive oil for a kitchen with no stove. In two months it burned through $30,000 of operating capital.

The postmortem blamed the model. The real problem was the absence of any control layer around it. No spend limits. No escalation rules. No human in the loop for decisions above a threshold.

The model was never the gap. The harness was.

For anyone building or approving AI architecture right now: the question isn't which agent to deploy. It's what sits between that agent and your production environment.

Capability is accelerating faster than most governance frameworks move. The organizations that will use agents well are the ones that treat control infrastructure as the non-negotiable prerequisite, not the afterthought.

Two data points about AI and jobs, both true this week:

One: A study of 21,000 US firms found that heavy AI adopters grew headcount about 10% faster than peers over two years. Entry-level roles grew fastest, at 12%.

Two: Tech and finance sectors are losing roughly 28,000 jobs per month. Goldman Sachs estimates 9% of the US workforce faces temporary displacement over the next decade.

Both are correct. And the tension between them is exactly the wrong frame.

The firms growing headcount aren't just using AI — they redesigned their workflows around it. The firms cutting roles are automating tasks without changing the surrounding process. Those are different interventions with very different outcomes.

Automating a task and redesigning a workflow look identical on the roadmap slide. They produce completely different results in practice.

Task automation: faster execution of the same process, marginal productivity gains, and often a displaced role with no structural improvement to show for it.

Workflow redesign: new decision loops, different human-AI handoffs, latent demand becoming real demand — this is where the value actually accumulates.

Ford rehired veteran engineers this week after AI tools fell short. Some companies that laid off workers are actively reversing those decisions. Moving too fast on headcount reduction before the workflow redesign is done is a real and recurring mistake.

The executives getting this right are keeping humans in the design loop during the transition — not removing them to show the board a productivity number.

This week it emerged that Anthropic embedded hidden fingerprinting logic inside Claude Code — its AI coding assistant.

The mechanism doesn't use IP addresses. It reads OS timezone settings and checks environment variables against an obfuscated list of proxy and corporate domains. On a match, it silently tags the request. The company then uses those tags to identify and ban accounts.

The stated target is China-linked users. That framing is doing a lot of work.

The more precise statement is: a tool with shell access to your files, your configs, and your Git repositories was silently profiling every user and reporting classifications back to the vendor. The logic was hidden well enough that it took a Reddit reverse-engineering thread to surface it.

This is not a geopolitics story. It's a vendor trust story.

Every enterprise that has granted an AI coding tool elevated access to production systems now has a concrete data point about what that access model actually permits. The vendor can instrument the tool, classify the user, and act on that classification — without disclosure.

China-linked developers are the label it prints today. The scanner was already pointed at everyone.

For anyone responsible for approving developer tooling in a regulated or sensitive environment: the security question has changed. It is no longer just what data this tool sends outbound. It is what logic is running locally that your security review cannot see.

Self-hosted models and air-gapped environments just got a stronger argument in their favor.

Meta is building a cloud business to sell its excess AI compute. When the news broke, the Philadelphia Semiconductor Index fell over 6%. CoreWeave dropped 13%. Nebius sank 14%.

That market reaction tells you everything about what the neocloud trade was actually priced on: the belief that GPU supply would remain permanently scarce and that someone would always need to broker it.

When the largest buyer signals it expects a surplus and starts rehearsing as a seller, that belief collapses.

The broader AI economics story is entering a different phase. Quarterly AI revenues now exceed quarterly capex depreciation for the major hyperscalers — meaning the cash flow math is finally turning. But cumulative historic investment still hasn't been recovered, and margins are thin.

For enterprises making infrastructure decisions right now, a few things follow from this:

First, long-term commitments to GPU capacity at peak scarcity pricing deserve another look. The cost curve is moving against those deals.

Second, model inference costs will keep falling. The access economics for capable AI are improving for everyone, not just the hyperscalers.

Third — and this is the one that matters most — when compute becomes abundant and cheap, the competitive advantage shifts entirely to what you build on top of it. Proprietary data. Domain-specific models. Workflow integration that a general-purpose API can't replicate.

The companies still treating AI as an infrastructure race are optimizing for the thing that's about to become a commodity.

A study of Claude Code users found something that should change how every organization thinks about AI adoption.

The people who got the most value from the AI coding agent were not the best coders. They were the people with the deepest domain expertise — regardless of job function.

A lawyer with strong domain knowledge of contract structures outperformed a generalist engineer on legal coding tasks. The more someone knew about their subject, the more useful output they extracted per prompt.

Most corporate AI training programs are built on the opposite assumption: that tool proficiency is the bottleneck. Learn to prompt. Learn the interface. Use it more.

The data says the bottleneck is domain depth. You need to know the subject well enough to evaluate what the AI produces, catch what's wrong, and direct the next step with precision. Without that, AI gives you fast output you can't reliably assess.

This has a direct implication for how I think about our teams in insurance technology.

Actuaries, underwriters, and claims specialists who have spent years developing judgment about risk — they are potentially the highest-leverage AI users in the organization. Not because they'll write the most code, but because they can direct an agent through complex, judgment-heavy work and actually evaluate the result.

There's also a risk embedded in this finding. Medicine has started talking about 'never skilling' — the concern that trainees who rely on AI before developing clinical judgment never build the capability that makes AI useful to them in the first place.

Domain expertise isn't just a nice-to-have for AI adoption. It's the prerequisite.

Amazon and Google both reported sharp spikes in greenhouse gas emissions in 2025. The cause was AI infrastructure — energy-hungry data centers scaling faster than efficiency improvements could offset.

This week, Nvidia announced a liquid cooling system for its next-generation chips that could eliminate chiller energy use in data centers entirely — a meaningful engineering step. But it solves the facility envelope. The compute itself still draws the same power.

For insurers, this sits in an uncomfortable place.

On the technology side: every organization with a net-zero or scope 3 commitment needs to account for the carbon embedded in its AI workloads. That's not a 2030 problem. It's a disclosure question now, as reporting standards tighten.

On the underwriting side: AI data centers are becoming a new category of concentrated, high-value infrastructure risk. They depend on uninterrupted grid access. Many are in water-stressed regions. They're deploying novel cooling architectures that haven't been stress-tested through a claim cycle. Over 75 projects were delayed this year by local governments over exactly these concerns.

The reinsurance market is flush with capital right now — midyear renewals are showing softening prices and strong risk appetite. In that environment it's easy to underwrite new infrastructure categories aggressively.

But the long-run loss exposure from climate-adjacent infrastructure concentrations is the kind of thing that reprices slowly and then all at once.

The intersection of AI infrastructure and climate risk is not a future problem. It's a current one that most P&C teams haven't fully modeled yet.