AI is moving fastest where insurers price risk. Most treat it as a model the data team owns.
Underwriting, claims triage and pricing are where the models land first. The hard part is not building them. It is keeping the risk judgement that defines an insurer in human hands once a confident model is offering an answer.
Request a Strategic Briefing →Insurers are among the fastest adopters of AI, and they tend to adopt it where the work is most measurable: underwriting automation, claims triage and pricing models. So it gets framed as a data-science and actuarial-tech build, owned by the people who build models. The trouble is that the judgement that defines an insurer, which risks to select, where the appetite sits, how to reserve, where the regulatory and fairness line falls, the call on the cycle, is not a model output. It is reasoning, and an unchallenged model will quietly shape it. Gildoni installs the Havruta Methodology (formerly the Think Partner Methodology) into how leadership teams reason with AI, so the model informs the risk judgement instead of replacing it. This is not underwriting software. It is the reasoning discipline above the model.
Insurance is adopting AI faster than almost any sector
This is not a sector waiting to be convinced. The models are already in the pricing engine, the claims queue and the underwriting desk.
Europe's insurance regulator has measured it. In EIOPA's market-wide study, nearly two-thirds of the undertakings surveyed were already actively using generative AI, and the regulator is clear that this is happening at speed while supervisors work out how to keep oversight current. Petra Hielkema, EIOPA's Chair, put the posture plainly: "The findings show that European insurers have been carefully considering how Gen AI could benefit their businesses."
The adoption sits exactly where you would expect: in the functions that turn data into a number. And it is moving faster than the governance around it.
The pace, in the regulator's own numbers.
of the insurers surveyed across Europe are already actively using generative AI, with adoption concentrated in underwriting, claims and pricing.
of non-life insurers were already using AI in pricing, underwriting, fraud and claims, against 24% of life insurers.
of undertakings have a dedicated AI policy, up from 25% in 2023. Adoption is outrunning the governance built to hold it.
The model build is not in doubt. The reasoning around it is.
The findings show that European insurers have been carefully considering how Gen AI could benefit their businesses.
The model is owned. The judgement is not
Because AI lands first in the functions that produce a number, it gets treated as the data team's project. That framing is where the blind spot opens.
A model output is not a risk judgement, but it reads like one. Ask a pricing or underwriting model what a risk is worth and it returns a clean, confident figure. The trouble is that the figure carries assumptions about selection, appetite and fairness that were never put to the leadership team for a decision. The model did not decide the appetite. It inherited one, and then quietly hardened it.
The judgement that defines an insurer sits above the model. Which risks to write and which to walk away from. Where the appetite sits this cycle. How to reserve against what you cannot yet see. Where the regulatory and fairness line falls when more precise segmentation starts to price vulnerable customers out. None of that is a modelling task. All of it can be shaped by a model answer no one challenged.
A confident model answer on risk is the Mirror Principle at underwriting scale: if the reasoning that framed the model was generic, the risk position it produces is generic, however precise the decimal places look. EIOPA itself flags the tension, warning that the same segmentation that lowers cost can propose higher premiums and reduce access for high-risk or vulnerable clients. That is not a data-quality problem. It is a judgement the leadership team owns, whether or not it knows the model is making it.
Buying the model is the easy part
Here is what the model market leaves out. Every vendor will sell you a better underwriting model, a faster claims triage, a sharper pricing engine. Almost none of them will tell you how the leadership team reasons through the risk call once the model is offering one.
That is the real gap, and it is not a modelling gap. Put a risk-appetite or reserving question to a capable AI and it will hand back a fluent, confident position without ever asking what you have left out. At the altitude where the cycle call and the regulatory line are decided, that is the Mirror Principle at its most expensive. A model can tell you what a risk scores. It cannot tell you whether you should be writing it, where your appetite should sit, or what the answer does to the customers you are accountable for. That is a different discipline, and it is the one that decides whether the model is an instrument or the author.
The model produces one clean number, and it reads like a decision. The risk judgement that an insurer actually owns, selection, appetite, reserving and the fairness line, sits above it. The red is the distance the model quietly fills when no one reasons across it: a judgement made, but never decided.
What the Havruta Methodology installs at leadership level
The Havruta Methodology is that discipline. It changes the default behaviour of the machine from agreeing with you to reasoning with you, which is exactly what a risk-appetite call needs.
The Flip
The Flip puts the machine on the other side of the risk. Instead of confirming the price or the position, it argues against it: where is this appetite too generous, what is the reserving missing, which customers does this segmentation push out, what would have to be true for the cycle call to be wrong. The leadership team gets challenged before the loss ratio does the challenging.
Ground Truth
Ground Truth keeps the reasoning anchored in the insurer's real book, its actual exposures, claims history and regulatory obligations, rather than the generic risk language a model produces by default. A reserving or appetite decision built on a plausible average is worse than no model at all.
Decision Velocity
And Decision Velocity lets the team decide at the speed the market is moving, compressing the path from model output to a defensible risk position without surrendering the judgement to the model.
The fuller account of how all of this works is on the methodology page.
What this is not
This is not underwriting software and it is not a claims platform. It is not actuarial or pricing tooling, not a model-risk-management suite, not AI training or general AI literacy. The models and the platforms are a separate market. This is the thinking that sits above them.
It changes how the leadership team reasons about the risk it already owns: the appetite call, the reserving position, the regulatory and fairness line, the decision on the cycle.
Where a leadership team starts
The methodology is installed along a ladder, and a leadership team enters at the rung that fits.
Most begin with the Eye-Opener Workshop, a half-day in which the team sees the shift on its own real risk work.
A leadership group embeds the practice through the Havruta programme, taking the discipline across the team.
A single high-stakes question, a risk-appetite call, a reserving position, a pricing-fairness decision, can be worked through Advisory Havruta.
How a CEO and the leadership team reason with AI
For the chief executive specifically, the role page takes the same discipline to the seat that answers for the risk the insurer carries. A Strategic Briefing is how to decide where to begin.
Go to the CEO pageLeadership questions about AI in insurance
Is AI in insurance a data-science build or a leadership question?
It is both, but the part that gets neglected is the leadership question. Building the underwriting, claims and pricing models is a data-science and actuarial task. Deciding which risks to write, where the appetite sits, how to reserve, and where the fairness line falls is risk judgement, and it belongs to the leadership team. The mistake is treating the whole thing as a model rollout. The model produces a number; the judgement around that number is what defines the insurer, and an unchallenged model quietly shapes it.
Why is treating AI as the data team's project a problem?
Because AI lands first in the functions that turn data into a number, so it looks like a build. But a pricing or underwriting model carries assumptions about selection, appetite and fairness that were never put to the leadership team for a decision. The model did not choose the appetite; it inherited one and hardened it. When the model is treated as the data team's tool, the judgement that the board answers for is being made by a system no one reasoned across. That is how accountability and authorship drift apart.
Can an AI model decide risk appetite or reserving?
A model can score a risk and inform the call, but it cannot own the appetite or the reserving position. Those are judgements about what the insurer should write, what it should refuse, and how it answers to the regulator and the policyholder. A model returns one confident number; it does not ask whether you should be writing the risk at all. When a leadership team accepts the figure without reasoning across it, the model has effectively made the call, and the team is accountable for a decision it never consciously took.
What about AI fairness and pricing of vulnerable customers?
Europe's regulator has flagged exactly this: the same segmentation that lowers cost can propose higher premiums and reduce access for high-risk or vulnerable clients. That is not a data-quality issue to be tuned away. It is a judgement the leadership team owns about where the fairness and regulatory line sits. A model will optimise to its objective without asking the question. The discipline is to make the AI argue the fairness case, anchored in the real book, before a pricing decision becomes a position the insurer has to defend.
Is this underwriting software or a model-risk platform?
No. It is not underwriting or claims software, not actuarial or pricing tooling, not a model-risk-management suite, and it does not touch your stack. Those build and govern the models. This addresses the thinking above them: how a leadership team reasons through a risk-appetite, reserving or pricing decision so the answer is genuinely theirs, anchored in the real book, and stress-tested before the loss ratio or the regulator tests it for them.
Where should we start?
With a Strategic Briefing, or with the Eye-Opener Workshop, where a leadership team sees the difference between instructing AI and reasoning with it on its own real risk work. From there the path depends on whether you are setting the appetite, embedding the practice across a leadership group, or working a single high-stakes question such as a reserving position or a pricing-fairness call.