The insurance industry spent 2024 and 2025 arguing about whether AI worked. That argument is settled, but it produced a strange result.
Datos Insights, reviewing AI implementations across property and casualty carriers this year, puts it plainly: most carriers have AI in production, but almost none have achieved widespread adoption. The top production use cases are all variations on AI reading and summarizing documents.
The pilot phase is over, but the industry stopped one step later.
In my experience working with VPs of claims at US P&C carriers, this is not a model problem, a budget problem or an ambition problem. It is what happens when AI is installed next to a claim decision instead of inside the process that produces it. The analysis gets better. But the decision takes exactly as long as it always did, because nothing around it moved.
"In Production" Became A Milestone Worth Gaming
Production and adoption are different states, and the industry has quietly started reporting the first while meaning the second.
Production means a model is deployed, serving requests, monitored and owned by someone. It is a real engineering achievement, and it is entirely compatible with nothing changing. A claims organization can have a summarization model in production for 18 months while every adjuster still opens the full file, because the process that governs their work was never rewritten.
Adoption means the work changed. Someone removed a manual assembly step, moved a review earlier in the file, or gave a supervisor a documented basis for an approval that used to rest on recall. The licensed claims professional still makes the determination. What changes is how much of the shift is spent getting to the point where determining anything is possible.
That requires touching process documentation, authority matrices, audit procedures and the core system. Almost none of it is AI work.
Datos Insights identifies data readiness and deployment maturity as the two variables that explain most of the variation in AI outcomes between carriers. I would put the same finding more bluntly. The wall is not in front of the model. It is between the model and everything the carrier already owns.
Summarization Is A Comfortable Place To Stall
The dominant production use case across carriers is reading and summarizing documents. I want to be fair to this: it works, it saves real time, and it was the correct place to start.
It is also structurally safe in a way that should make executives suspicious. Summarization touches no decision authority. It requires no model governance sign-off and no change to the policy administration system. A carrier can run it at scale for two years and never confront the question that actually gates value: is this organization willing to let a machine-produced analysis become part of the documented basis for a determination that a licensed claims professional signs?
That question has an answer, and it is not a technical one. It is a governance answer. Who reviews the analysis, against what standard, with what authority to accept, modify, reject or escalate it, and where that disposition is recorded.
I'd recommend treating a portfolio that is entirely summarization as a warning sign rather than a milestone. If every use case in production is one that no reviewer ever has to put their name to, the organization has not yet tested the thing it needs to test.
The Wall Is the Write-Back Path
Here is the pattern I see most consistently in stalled deployments. The model produces an assessment. The core system requires a transaction. Between those two facts sits a person, retyping.
A claims model produces a severity indicator at first notice of loss. The claims system has no field for it, no way to record which model version produced it, and no place to store what the adjuster did with it. So the indicator is displayed on a screen, the adjuster reads it, applies judgment, and then keys a value the system already recognizes. The analysis is in production. The cycle time is unchanged, because the constraint was never the assessment, and it was never the adjuster's judgment. It was the retyping between them, repeated across every file in the inventory.
This is why integration with the existing core, rather than model accuracy, is what I see stopping these programs. Carriers built AI capability on the read side, where data already flows outward and nothing needs permission. The write side was left for later, and the write side is where governance actually lives.
Six fields decide whether an AI-supported claim file is defensible: the analysis itself, the model version behind it, the evidence and policy language it drew on, its stated confidence and limitations, the reviewer who accepted, modified, rejected or escalated it, and that reviewer's authority level and timestamp. A carrier that cannot store all six has not built decision support. It has built a second screen. Later has arrived, and closing that gap is a core platform project, not an AI project.
McKinsey's research on AI in the insurance industry points at the same structure from the value side. Isolated proofs of concept, the firm argues, tend to lack workflow integration and sustainable benefit capture, while transforming a whole domain can move the bottom line by double digits.
Its budgeting rule is the sentence I'd pin above the desk of anyone sponsoring one of these programs: for every dollar spent developing digital and AI solutions, plan on at least another dollar to reach full adoption and scale. Most carriers I have looked at are running roughly a 10-to-1 ratio in the other direction.
Governance Is The Gate, Not The Brake
There is a common view that regulation is what slows AI down in insurance. I disagree. I think the causality runs the other way.
The NAIC adopted its Model Bulletin on the use of AI systems by insurers in December 2023, and it asks carriers to maintain a documented program covering governance, risk management, internal controls, and third-party AI. State departments have been adopting it since, and more states regulate insurer AI under their own frameworks. None of that prevents a carrier from scaling. What it does is make scope expansion conditional on being able to explain the system.
A carrier that cannot show which model version produced an analysis, on what data, reviewed by whom, and under what authority, is not blocked by the regulator. It is blocked by its own inability to answer the question, and it self-limits to use cases where the question never comes up. That is the summarization trap described from a different angle.
The defense of an AI-supported claim file in litigation is not that the model was correct. It is that the organization ran a controlled process, reviewed the relevant information, documented the reasoning, applied the required authority, and kept ownership of the outcome with a licensed professional. Carriers that can produce that record can widen scope. Carriers that cannot, stay where they are and call it caution.
McKinsey's 2026 survey on AI trust, fielded across roughly 500 organizations, found that nearly two-thirds of respondents cite security and risk concerns as the top barrier to fully scaling agentic AI. It also found that organizations assigning clear ownership for responsible AI averaged a maturity score of 2.6, against 1.8 for those with no clearly accountable function. Governance capability and scaling capability turn out to be the same capability.
What The Carriers That Got Through Do Differently
I don't have inside visibility into every carrier that crossed from production to enterprise deployment, and some of them got there with brute force. But in my experience working with claims and technology leadership on these programs, four patterns recur.
- They pick a domain, not a use case. Underwriting submission intake end to end, or first notice of loss through assignment. Isolated use cases produce demonstrations. Domains produce a P&L line.
- They build the write-back path before the model. Analysis, model version, evidence, confidence, reviewer, disposition. If the core system cannot record both what the analysis said and what the human did with it, everything downstream is theater.
- They fund adoption at parity with build. Process rewrites, authority changes, training, revised audit procedures. This is the unglamorous half, and it is usually the half that is cut when budgets tighten.
- They instrument the decision, not the model. Model accuracy is an engineering metric. The business metrics are the ones a chief claims officer already reports: cycle time from FNOL to a decision-ready file, the share of files that reach an adjuster fully assembled rather than half assembled, rework and reopen rates, and LAE per claim. Carriers that report the first and none of the rest generally cannot tell whether anything changed.
The Question Worth Asking
The gap between production and adoption will not close by buying better models. It closes when a carrier can point to a determination its people now reach faster, on better-assembled evidence, with the reasoning on the record, and can show a regulator exactly who made it and under what authority.
So here is the question worth putting to your leadership team this quarter. For every AI system you currently run in production: what work was removed from the process it touches, and what does the claim file now show that it did not show before?
If the answer is that a person has better information but performs the same administrative sequence, and the file records nothing about how the analysis was used or by whom, the system is deployed. It has not been adopted, and the difference is where all the value was.
Sources
- Datos Insights. "AI Implementations in 2026: The Pilot Phase Is Over for Property and Casualty Carriers." 2026. https://datos-insights.com/reports/ai-implementations-in-2026-the-pilot-phase-is-over-for-property-and-casualty-carriers-ins-2026-102369/
- McKinsey & Company. "The future of AI in the insurance industry." https://www.mckinsey.com/industries/financial-services/our-insights/the-future-of-ai-in-the-insurance-industry
- McKinsey & Company. "State of AI trust in 2026: Shifting to the agentic era." 2026. https://www.mckinsey.com/capabilities/tech-and-ai/our-insights/tech-forward/state-of-ai-trust-in-2026-shifting-to-the-agentic-era
- National Association of Insurance Commissioners. "Artificial Intelligence." Insurance Topics. https://content.naic.org/insurance-topics/artificial-intelligence
