Why AI hasn't made insurance more efficient

Every year, someone tells the insurance industry it's on the verge of getting more efficient. New technology, new promise, same story. I've been hearing a version of it since I started my career, first as a broker, now on the technology side of the business.

So, here's a question worth asking honestly: if all that promised efficiency were real, why hasn't it shown up in the numbers?

A recent McKinsey analysis of the industry's economics makes the case. Labor productivity across claims, servicing and policy issuance has improved by 14% in P&C and 24% in life insurance. And yet the average P&C expense ratio has hovered between 27% and 32% since 2005, essentially unchanged in two decades, despite tech improvements. On a broader measure of cost efficiency, McKinsey found insurance cost ratios, measured as SG&A relative to revenue, are 10% higher globally than they were in 2005.

Productivity went up but costs went up more.

That's not a story about technology failing to deliver but one about technology getting layered onto processes that were never actually fixed. Every new tool made an existing step faster and none of them asked whether that step needed to exist in its current form at all. IT spend, compliance overhead, and the sheer complexity of bolting modern tools onto decades-old operating models absorbed the gains before they ever reached the bottom line.

I'll offer an opinion that may be unpopular in rooms full of vendors. I think most leadership teams in this industry are considerably more cautious about betting on AI than their public messaging suggests. Pilots are getting funded and proofs of concept are being announced. But the structural bets, the ones that actually change how a placement gets built or an underwriting decision gets made, are happening at a fraction of the pace and the caution is rational. Getting a big, visible bet wrong in a business built on trust and capital is expensive in ways that go beyond the balance sheet.

We need to remember that caution has a cost too. Here's the distinction I think gets lost. Speed was never actually the prize. If you take a broken process and make it faster, you arrive at the same limited outcome sooner. That's useful, but it's not transformative. The real unlock is removing the friction sitting underneath the process altogether: The rekeying of the same submission data across five systems, the manual reformatting of a document that should have been usable the moment it arrived, and the hours an underwriter spends normalizing data before they can even begin evaluating the risk in front of them.

Remove that friction, rather than simply speeding past it, and something different happens. Underwriters gain the capacity to see more of the risk in their pipeline, not just process the risk they already had faster. That's not a productivity story but a capacity story and the only kind of story that would move the ratios. And yes, there are much larger and deeper opportunities well beyond what I'm discussing here. But getting the basics right is the real ticket-punching step, the one that lets everyone confidently invest in structural change, new and innovative models, or both.

So, the next time a vendor tells you their AI strategy is working because things are faster, there's one question worth asking back: are the ratios actually moving? If the answer is no, the friction is still there. You've just found a more expensive way to live with it.