AI in P&C Insurance Claims: Why the System Around the Model Matters More Than the Model

Model releases come fast. Every few months a more capable one lands, and claims leaders across property and casualty, in personal and commercial lines, get asked the same question. Does this change everything?
Usually it doesn't. It's worth knowing why before you sign anything.
We learned this building ClaimLens. The model is the easy part. What makes AI trustworthy on a real claim file is the system built around it.
On its own, a strong model will sometimes invent a policy limit that does not exist, lose track of a long file, or miss a contradiction on page 2,000. Put the same model inside a well-built AI claims processing system and it catches those problems, and it shows you where each answer came from.
In P&C claims, and especially on complex liability files where medical, legal, policy and damages questions all overlap, that system is what you are paying for.
What "the system around the model" means, in plain terms
Think of an AI model as a very talented new chef. Trained on everything, fast, and genuinely capable.
Now put that chef in an empty room. No recipes, no labeled ingredients, nobody checking the food, no record of what was served. You would get some great dishes and the occasional disaster, and you wouldn't know which was coming.
A restaurant works because of everything around the chef. Good ingredients. Tested recipes. A kitchen that tracks every order. Someone who checks each plate before it goes out.
People in AI sometimes call this the scaffolding. Whatever you call it, it is what makes a capable model dependable on real work. In claims, where a wrong number or a made-up citation has legal and financial consequences, that is where the value is.
Hold on to the chef. The rest of this piece is about the kitchen.
The five parts of any AI agent
The chef analogy gives you the intuition. Here is the engineering version.
If the chef analogy is the intuition, here is the engineering version. Jensen Huang has described an autonomous AI agent, as opposed to a simple chatbot, using a "worker in a workshop" analogy with five distinct parts. It is a clean way to see where the real work happens, so we'll borrow it and then show where claims value actually sits.
The model is the brain. The underlying LLM. It reads, reasons, plans and decides. This is the part everyone talks about, and the part that changes every few months.
The harness is the body. The code wrapper around the brain that orchestrates the actual work: it understands intent, observes context, holds short- and long-term memory, manages the lifecycle of a task, and decides when to take an action.
The tools are the instruments. The programs the agent uses to get things done, such as retrievers, classifiers, calculators, databases and code execution.
The sandbox is the workshop. A sealed, secure runtime where the agent can work and verify outcomes without breaking anything around it.
The knowledge is the data. The specific enterprise data, files, guidelines and context the agent needs to do the job accurately.
Here is the point for a claims leader. Only the first part, the model, is a commodity you buy off the shelf and swap when something better arrives. The other four are what a vendor builds, or what you end up building yourself. Of those four, the harness decides whether AI is reliable on a real claim file.
The model sets the ceiling. the harness, the tools, the sandbox and your knowledge set the floor. On a bodily injury file running to thousands of pages, the floor is what you live on.
So when you sit down with a claims AI vendor, skip "which model do you use?" and ask "show me your harness."
Where general chatbots fall short on a claim file
A general-purpose chatbot, the kind anyone can sign up for, is a powerful brain with almost no harness around it. On real claim files we see three problems over and over, and all three are harness problems.
It loses track of details
Claim files get big: demand packages, IME reports, depositions, repair estimates, thousands of pages of medical bills. As they grow, the details that decide the claim get fuzzy. The exact deductible. A specific exclusion. The actual date of loss. That is a serious problem for bodily injury claims in auto and general liability, and for medical bill review, where the right answer depends on cross-referencing dozens of documents.
It is a memory failure, which makes it the harness's job.
ClaimLens by DocLens keeps the whole file in view and brings the relevant evidence back in for every question, so a fact on page 2,000 is as available as one on page 2.
It guesses, and sounds sure about it
Ask a general chatbot a narrow coverage question and you will often get an answer that reads well and is sometimes wrong. In claims, that is exposure. The legal profession has already seen where this leads. A growing number of lawyers have failed to verify work they delegated to AI. Many of the errors are hallucinated citations to cases or legislation that don't exist, which may come from the model's tendency to produce a plausible-sounding answer instead of acknowledging a gap in its knowledge. AI has also inaccurately attributed legal principles to real cases. Lawyers who fail to verify AI-generated content have faced fines, sanctions and disciplinary action.
For carriers the implication is simple. An ungrounded chatbot adds risk. A tool that grounds and cites its answers reduces it, and the same grounding is what makes AI-assisted claims fraud detection something you can defend.
Grounding is a harness-and-knowledge job. It is the difference between answering from the model's training memory and answering from this claim's record.
It can't show where an answer came from
"Trust me" doesn't hold up in a bad-faith dispute. A general tool gives you an answer but can't point to the page it came from. ClaimLens attaches a page and line citation to every extracted value. Click it and the source text opens.
A bigger model doesn't fix any of these. Better structure does.
Going deep: what our harness actually does
This is the part most vendors skip, so it is worth being specific. The ClaimLens harness is the code around the model that turns a raw claim file into grounded, traceable, decision-ready information. It maps directly onto the way we think about claims, Extract → Evaluate → Execute. Better information leads to better decisions, which lead to better claim outcomes. Walk it through in order.
1. It understands what's in the file (Extract). Before a model answers anything, the harness classifies what is actually in the claim. That means demand packages, court filings, IME reports, medical and pharmacy bills, police reports, depositions, policy declarations, ACORD and ISO forms. It handles handwriting, multi-page tables, poor scans and files that run to thousands of pages without a person sorting them first. This is in-repo document classification, built for insurance, not a generic PDF reader.
2. It retrieves evidence from the claim's own record (Extract → Evaluate). The harness runs grounded ("zero-hallucination") retrieval over the documents in this claim. Every answer ties back to the file in front of you, never to what the model happened to learn on the public internet during training. This is the single most important thing the harness does, because it is what makes an answer defensible. It is also the fix for the guessing problem.
3. It holds durable context across the life of the claim (the memory job). Complex liability claims unfold over months and thousands of pages. The harness keeps facts persistent across many turns and across the whole file, so a detail established early doesn't evaporate later. This is the "long-term memory" element in the workshop map, applied to a single claim, and it is the fix for losing track of details.
4. It scores confidence and knows what it doesn't know (Evaluate). Each extracted value carries a confidence score, so adjusters can see at a glance what to trust and what to check. The harness doesn't pretend to certainty it doesn't have.
5. It routes exceptions to a human instead of guessing (Execute). Missing or low-confidence data goes to a person as an explained exception (here is the gap, here is why) instead of being passed through quietly. People make the decisions. The harness gets them there faster, with better information in front of them. Human-in-the-loop is designed into the workflow from the start, rather than bolted on at the end.
6. It enforces an insurance domain layer on top of the model. Insurance ontologies, medical taxonomies, state regulatory frameworks and risk-signal libraries encode how claims actually work and enforce source traceability. A general model can't copy this layer, because it comes from claims expertise rather than the public internet.
7. It runs on swappable models, which is why a new release helps you. Our harness talks to models through Amazon Bedrock, which offers access to 100+ foundation models, and it is model-agnostic by design. The underlying model can be swapped as better ones arrive, while the harness, the domain layer, the grounding and the governance stay put. The ceiling goes up; the floor you built doesn't have to be rebuilt.
8. It is governed from the first line of code. NIST AI RMF aligned, SOC 2 Type II, HIPAA. Customer files, health information and prompts are never used to train models. Governance lives in the harness, so it applies to every answer regardless of which model is underneath.
These eight steps keep results consistent, and they are why a claims AI platform built this way stays dependable as the models underneath it change.
In the workshop map, we have invested in the harness, the tools, the sandbox and the knowledge layer. The brain is the part we rent.
You aren't locked into one model
This part matters to whoever signs the contract. Because the value sits in the harness and the domain layer around the model, ClaimLens doesn't depend on a single model. It runs on Amazon Bedrock and uses the right model for each task. When a better model comes out, it goes in underneath and the platform improves. The trust and governance work doesn't have to be redone.
Which brings us back to the question that claims leaders keep getting asked. For a system built this way, a new model release is good news, with no migration project attached.
What this looks like in results
The reason to build all this is claim economics. Across deployments with P&C carriers, we consistently see the following
About 20% operational efficiency gain, from less manual reading, re-keying and chronology building
2 to 4% loss-dollar savings, from spotting liability early, flagging unverified medical charges and negotiating with better evidence
95%+ extraction and chronology accuracy, validated by human expert review and ahead of general-purpose AI on complex medical records
Roughly 10x better data quality, with a person in the loop throughout
Minutes saved per adjuster is a fine number to track. What carriers care about more is better information leading to better decisions and better outcomes over the life of the claim.
The takeaway for claims leaders
The P&C carriers getting real value from AI in insurance claims aren't chasing the newest brain. They have invested in everything around it. The harness that orchestrates the work. The domain knowledge that makes it accurate. The sandbox that keeps it safe. The human sign-off that keeps it accountable.
That part is harder to build, and it keeps paying off. In complex liability claims, it decides whether AI stays a demo or becomes something you can stand behind.
If you are evaluating claims AI software for personal or commercial lines, start with the harness. Ask the vendor to show it to you, and then ask them to show you where an answer came from. In claims, an answer you can trace is the only kind worth acting on.
ClaimLens is an outcome platform, not an AI productivity tool. The difference is the system around the model.





