The Claims Paradox

Walk into any major carrier’s claims operation and you’ll see a contradiction. On one screen, automated decisioning models are quietly resolving thousands of low-complexity claims a day with no human in the loop. On the next desk, an experienced adjuster is keying repeated context from one screen to another, hand-rolling notes that an LLM could have drafted in two seconds, and toggling between seven systems to assemble a picture of a single claim.

Both are real. Both are happening at the same company. This is the claims paradox.

How we got here

Claims technology has been accreted, not architected. Every system you find in a modern claims operation was, at the time it was built, the right answer to a specific question: how do we capture FNOL faster, how do we standardize estimating, how do we detect fraud, how do we route assignments. The questions were asked one at a time, and the answers were built one at a time. The result is a collage, not a stack.

The collage worked because the binding agent was the human adjuster. They were the integration layer. The translation layer. The judgment layer. The thing that made the seams invisible to the customer.

That worked well enough for thirty years. It does not scale into the next ten.

What’s actually changing

Three things are happening at once, and the interaction effects matter more than any one of them.

The shape of a claim — structured fields, document-heavy, decision-graph-driven — is unusually well-suited to the strengths of modern model architectures.

One. Foundation models are now good enough at reading documents, summarizing context, and reasoning across structured and unstructured data that the cost of “understanding” a claim has fallen by an order of magnitude. The bottleneck is no longer comprehension. It’s orchestration.

Two. The data spine in front of carriers — telematics, IoT, photos, scene reconstruction, parts catalogs — has matured to the point where the inputs are richer than any one human can hold in their head. The bottleneck has shifted from data scarcity to data fusion.

Three. Customer expectations have decoupled from carrier capabilities. The customer expects the experience of their last good app — Apple, Amazon, Uber. The carrier delivers the experience of its last system migration. The gap is structural and growing.

The operator opportunity

When all three of those things are true, the right move is not to add another tool. It is to build an operator — a piece of software that absorbs the integration, translation, and judgment layers that the human used to perform, and gives the human a higher-leverage seat.

The right metaphor is not “AI assistant.” It is air traffic controller. The human doesn’t fly the planes. They direct the system that flies the planes, and they intervene when the model is uncertain, when the stakes are unusual, or when judgment is what’s actually being asked for.

That requires three things most claims software has historically gotten wrong:

1. A canonical model of the claim. Not five overlapping models in five systems. One.
2. A decision-grade UI. Built for the cadence and density of real claims work, not for screenshots in sales decks.
3. An honest treatment of uncertainty. When the model doesn’t know, it has to say so, in a way the human can act on.

The carriers that figure out all three first will redefine the unit economics of claims. The ones that don’t will pay a slow, compounding tax.

What I’m watching

Three things I’m tracking through 2026:

• Carrier-internal builds vs. platform partnerships. Most carriers don’t have the engineering bench to do this alone. Most platforms don’t have the claims fluency to do it without the carrier. The interesting work is at the seam.
• Regulatory clarity around model decisions. State-level guidance is catching up. The carriers that move first will get to shape the precedent.
• The reskilling of the adjuster role. The job changes. The people in the job have to change with it. The carriers that treat this as a software problem will fail the human side. The ones that treat it as a workforce problem will fail the software side. It’s both.

There is a version of the next decade where claims becomes the most operationally elegant function in financial services. There is a version where it stays a collage. The choice is being made right now, in the room where the next platform decision gets approved.

I know which version I’m building toward.