Seven angles this week

Anthropic just posted numbers most companies would kill for — $16.8bn in a single quarter, customers spending nine times what they did a year ago — and the market called it a miss.

That tells you something important, and it has nothing to do with Anthropic.

It tells you that the AI trade now runs on expectations, not fundamentals. And the same week, the cost of insuring against tech defaults hit a record high, with data-center operators hit hardest.

Here is what that means if you run technology platforms: the compute underneath your shiny new AI features sits on a financial structure you do not control. You are no longer just betting on a vendor's product roadmap. You are betting on its funding cycle.

I treat a concentrated AI model dependency the way I treat any single point of failure in a core system. Documented exit path. A credible second source. Contracts that survive a bad quarter.

A hot vendor is not the same thing as a safe one. The firms that learn that distinction the hard way will be the ones who architected for the pitch deck instead of the ledger.

Model independence is not a philosophical stance. It is risk management.

What is your second source if your primary model provider has a rough 18 months?

There is a quiet clause being written into insurance renewals right now that could matter more to your AI strategy than any model benchmark.

The AI exclusion.

Marsh's CEO just warned that broad AI exclusions in policies will create real confusion and uncertainty. Translate that: your insurer may be deciding which of your AI deployments are effectively uninsurable, and most leadership teams have never read the language.

Think about what this means for an agentic workflow. You grant an AI system autonomy to act. It makes a consequential mistake. If your policy excludes AI-driven losses, that liability lands squarely back on your balance sheet.

Suddenly the architecture question — how much autonomy do we give this agent — is also an insurance question. They are the same conversation.

I have started treating coverage review as part of AI deployment, not an afterthought handled by a different department. Map which of your core processes now depend on AI. Then ask, plainly, whether a loss there is covered.

And push back on blanket exclusions. The nuanced, use-case-specific approach Marsh is calling for is the right one. A broad exclusion is the insurer protecting itself at your expense.

The AI safety conversation everyone is having happens in model labs. The one that will actually hit your P&L is happening in your renewal paperwork.

When did your team last read the AI language in your policies?

Your AI pilot looks great. That is exactly the problem.

McKinsey put numbers to something I have watched happen repeatedly: moving an AI tool from pilot to production can nearly quadruple its cost. The demo that wowed the steering committee becomes a very different business case once it runs at scale.

And there is a second trap. Analysis now runs on AI time. Judgment still runs on human time. You can generate a brilliant recommendation in seconds, then watch it sit for three weeks in review cycles, committees, and sign-offs.

Speeding up the analysis without redesigning the judgment flow around it is like buying a faster engine for a car stuck in traffic. You paid for horsepower you cannot use.

So I push two disciplines with every AI initiative.

First, map where human judgment genuinely adds insight versus where it is just a habit-formed gate. Remove the gates that add nothing. Protect the ones that do.

Second, cost every pilot at production scale before you approve it. Not the pilot price. The real price, with the 4x jump baked in. Do it before the project builds political momentum, because once everyone has celebrated the demo, killing it on economics is nearly impossible.

Adoption is a vanity metric. Value survives contact with production cost and human decision speed.

Where does judgment actually gate value in your organization — and have you ever measured it?

Too many companies are choosing an AI model the way they chose an ERP system in 2005. Pick a winner. Commit. Integrate deeply. Live with it for a decade.

That is the wrong mental model, and it will cost you.

Frontier models are not foundational infrastructure you marry. They are a fast-moving, rapidly commoditizing layer. There is an active price war. Capability leapfrogs every few months. Architecting your core platforms around one provider's API and prompt conventions is a bet against the pace of the entire field — a bet you will lose.

Model independence is the strategy that actually matters, and it is an architecture decision you make now, not a procurement decision you make later.

The move is unglamorous: build an abstraction layer between your business logic and whichever model answers the call. Your core systems should not know or care which provider is behind the interface. When a cheaper or better model appears — and it will — you switch in weeks, not quarters.

This is exactly the kind of work that gets skipped because it does not demo well. There is no flashy keynote moment for a well-designed interface. But it is the difference between optionality and lock-in.

I would rather be the firm that can swap models over a long weekend than the one that spent two years hardwiring itself to a vendor whose best quarter is already behind it.

Are your platforms wired to a model, or to an interface?

Apple built a smart-home device, boxed it up, and left it sitting in warehouses for the better part of a year — waiting for an AI that was not ready.

That is the real story of HomeView, and it is a warning for anyone putting AI into systems meant to last.

Apple's problem was never that its AI was bad. It was that a hardware roadmap became hostage to an AI capability the company could not ship on schedule. The product was designed from the start around a version of Siri that could recognize people and retrieve personal content. Without it, there was no product. So it slipped. And slipped. And finally arrived into a market where rival agents do not just answer questions — they act on them.

Here is the lesson I take into enterprise architecture.

Never couple an irreversible, long-horizon commitment to an AI capability that does not yet exist reliably. Thermostats, door locks, core platforms — these are things you install once and live with for years. If they only work when the AI finally catches up, you have built a product that is useless on arrival and dated by the time it functions.

Design for the AI you have today, with headroom to improve as it matures. Reversibility is the premium feature.

When you bet a ten-year platform decision on a model's promised roadmap, you inherit every single delay in that roadmap.

Are you building for the AI that exists, or the one in the keynote?

While everyone watched the frontier AI labs this week, the more telling story in insurance was quieter: CCC Intelligent Solutions, the platform that runs a huge share of auto claims decisioning, agreed to go private in a 6.1 billion dollar deal.

Same week, a report showed auto total-loss frequency climbing to 23.5% of claims.

Put those together and you see where the real AI value in P&C actually lives. Not in chatbots. In the unglamorous infrastructure that processes millions of structured decisions every single day — claims estimation, total-loss calls, repair networks.

Private capital does not pay a premium to take a company private so it can keep running the status quo. GTCR and Elliott see room to rebuild these systems with AI at the center, away from quarterly earnings scrutiny. That is a signal worth reading.

For technology leaders in insurance, this is the frontier that matters. The decisioning platforms are where AI compounds, because every point of accuracy and every day of speed flows straight to the loss ratio. Rising total-loss frequency only raises the stakes on getting those calls right.

The flashy use cases get the headlines. The claims and underwriting infrastructure gets the returns.

My rule of thumb lately: watch where private capital concentrates. It is a pretty good map of where the durable AI value sits — and it is pointing straight at the core systems, not the demos.

Where is your organization placing its AI bets — the frontier, or the foundation?

The common fear about AI is that it commoditizes human skill. I think the truth is almost the exact opposite.

AI produces what one guest on Stratechery this week called the greatest average. Competent. Middlebrow. Predictable. That is genuinely useful for the 95% of work that is routine. It is nearly worthless for the 5% that actually differentiates you.

And here is the part most people miss: the value of what you get out of AI depends entirely on the expertise you bring to it. Ask a generic question, get a generic answer. Ask the question only a deep expert would know to ask, and the output can be extraordinary. Same model. Completely different value. The scarce ingredient is the human on the keyboard.

That reframes how I think about talent and organization design.

Automate the average aggressively. It is cheap now, and it is only getting cheaper. Stop paying premium rates for competent-and-predictable.

Then concentrate your scarcest resources — your best people, your attention, your budget — on the judgment and taste that cannot be reduced to a probability distribution. The things, as one investor put it, that cannot be disintermediated.

Your competitive edge will not come from having AI. Everyone will have it by next year. It will come from the quality of the human judgment you point it with.

So the real question for leaders is not how much AI can you deploy. It is: who are the people whose taste you are amplifying — and are you protecting them?