Six angles this week

The most important insurance article this week wasn't about a storm or a lawsuit. It was about plumbing.

A viewpoint piece described the 'unbundled' insurer — where the roles between a policyholder and the capital standing behind their policy keep splitting apart. Retail agent, wholesale broker, underwriter, carrier, reinsurance panel, and now AI-driven functions in between.

Most people read that as a distribution story. I read it as an architecture story.

When a value chain fragments, every function quietly becomes a service — with its own interface, its own data contract, its own way of failing. AI speeds this up because it makes each piece cheap enough to run on its own. But it also multiplies the number of seams. And seams are where systems break.

Here's the uncomfortable part: most core platforms were built assuming an integrated carrier. One underwriting engine. One policy system. One claims flow. That assumption is dissolving underneath us.

The firms that win won't be the ones that resist fragmentation. They'll be the ones that fragment on purpose — exposing clean internal interfaces so they can recombine functions faster than anyone stitching together third parties.

The ones that fragment by accident inherit something worse: a system where no single person can explain how a policy actually gets priced.

Before you chase AI features, ask a harder question — do you actually know where your seams are?

Most AI programs are winning on paper and losing in reality.

McKinsey made a point this week that I keep seeing play out: AI's easiest wins are misleading leaders. Summarize this email. Draft that memo. Clean up these notes. All real. All measurable. All deceptive.

Here's why. Those wins improve how fast an individual works. They do almost nothing to change how work actually flows through the organization. And in most companies, the bottleneck was never the individual — it was the process they're trapped inside.

So you end up with a scoreboard full of pilots and time-saved anecdotes, and a business that performs exactly as it did before. Leaders mistake a pile of local speedups for a transformation.

The real value is somewhere far less fun to talk about: redesigning the process itself so an agent can own a whole step end-to-end, and redrawing accountability so a human still owns the outcome.

That's slow. It's political. It means someone's job description changes. Which is exactly why most programs avoid it and settle for tools instead.

My advice to anyone running an AI effort: stop counting tools. Pick one process — underwriting intake, a claims step, a reconciliation — and redesign it around agents from the ground up. One real redesign teaches you more than twenty bolted-on features.

Find the constraint first. Then spend.

This week OpenAI released 722 mathematical manuscripts from an unreleased model. Problems that stumped the best human minds for decades fell in an afternoon of compute.

The reaction split into two camps. One is amazed. The other worries we're heading toward a world with more answers and less understanding — proofs that are verified correct but that no human actually comprehends.

That sounds like a philosophy seminar. It isn't. It's a direct warning for anyone deploying AI in a regulated business.

Here's the translation. A model can produce an output that is verifiably correct and completely unexplainable. In mathematics, that's an interesting tension. In insurance, it's a liability.

Picture a pricing adjustment or a claims denial that is statistically sound and impossible to explain in human terms. Now picture explaining it to a regulator. Or a court. Or your own audit team. 'The proof checks out' is not an answer when a customer has been denied coverage.

The maths story is a preview of a choice every technology leader will face: do you insist your systems explain themselves, or do you accept answers you can't defend?

I'd argue interpretability has to be a first-class requirement — captured reasoning, decision lineage, human-legible rationale — even when it costs you some accuracy.

The firms that treat explanation as optional will ship genuinely impressive models they cannot stand behind the first time someone pushes back.

Answers are cheap now. Defensible answers are the real asset.

Nine South Korean banks got hit in a wave of cyberattacks this week. The detail that matters: experts say AI is lowering the bar for criminals who don't have deep technical skills.

We usually comfort ourselves with the idea that AI arms the defenders too. True. But it misses the real shift.

AI collapses an asymmetry that used to protect us. Sophisticated phishing, convincing social engineering, credible intrusion — these once required genuine expertise. Now they're cheap, fast, and available to anyone. The pool of capable attackers didn't grow a little. It grew by orders of magnitude.

Most security budgets were built assuming a certain class of adversary. That assumption just broke.

Two consequences land at once.

Inside your own walls, the perimeter is no longer the defense. Identity and verification are. As Experian put it this week: trust no one, verify everyone. That's not a slogan, it's an architecture decision — strong identity, zero implicit trust between systems.

And if your firm underwrites cyber, there's a quieter problem. Your loss models are backward-looking at the precise moment frequency and severity are being rewritten. Pricing yesterday's threat model is how you end up badly mispriced.

The practical move for technology and underwriting leaders is the same instinct from two angles: stop trusting defaults. Stress-test your defenses — and your loss assumptions — against AI-enabled attack economics, not historical patterns.

The attackers already upgraded. The question is whether your models did.

A federal lawsuit this week accuses McDonald's of using AI pricing tools to fix menu prices. It's the latest in a growing pattern: companies accused of using algorithms to quietly collude.

Most AI risk conversations fixate on bias and hallucination. This is a different and more mundane danger, and it's heading straight for pricing-heavy industries.

Here's the mechanism. Algorithmic pricing can produce collusion-like outcomes without anyone intending to collude. Several companies buy the same pricing engine. They feed it overlapping market data. The models converge. Prices move together. No smoke-filled room required.

That matters because antitrust law traditionally turned on intent. With AI, the test may shift toward outcomes. And 'we didn't mean to' is a weak defense when the result looks identical to coordination.

Now look at insurance. Pricing is algorithmic. Increasingly model-driven. Often built on shared industry data and common vendors. That's precisely the recipe these lawsuits are circling.

You don't need a conspiracy to catch a claim. You need prices that drift together because everyone licensed the same model.

The practical defense isn't exotic. Document that your pricing logic is genuinely independent. Keep decision inputs auditable. Be wary of black-box vendors whose models you can't inspect — because you may have to explain them under oath.

And retire the phrase 'the algorithm decided.' In a courtroom, that's where your problem begins, not where it ends.

A London strategist warned this week that the AI trade could cause the worst market crash since 2008. The usual bull-and-bear fight broke out underneath the headline.

I think that debate misses the question an operator should actually be asking.

Here's a quieter detail from the same week: AI revenue is running around $276 billion annualized — but a striking share of it comes from a small slice of heavy users. One newsletter put it bluntly: the 1% pays AI's bills.

That concentration, not the valuation, is the real risk for anyone building on top of these vendors.

Because whether or not the stock market corrects, the business underneath your critical AI supplier might be thinner than it looks. If the economics depend on a narrow base of power users and the subsidy era ends, two things happen. Some of today's vendors disappear. And the survivors raise prices.

So the bubble question I care about isn't 'will the market crash.' It's 'will my supplier still exist in three years, and at what price.'

The defense is architectural. Put your AI capability behind an abstraction layer so you can swap providers. Avoid building deep dependencies on one model's quirks. Negotiate assuming prices go up, not down. And treat today's AI pricing as promotional — because it is.

Build so that a vendor failing is an inconvenience you route around, not an outage that takes you down with them.

Bet on the capability. Don't bet the company on any single provider.