The Architecture of Trust

I believe the key distinction is whether AI expands people's choices or quietly narrows them.

A good insurer should act like a navigation system: it can show risks, suggest routes, and warn about hazards—but it should not secretly choose the destination.

If insurers use AI to provide clearer information, personalized risk insights, or preventive recommendations while preserving customer autonomy, that creates value. But when algorithms steer people toward specific behaviors or products without sufficient transparency or meaningful alternatives, support becomes influence.

As insurers become embedded in healthcare, mobility, and broader digital ecosystems, this boundary will become increasingly important.

I do not think a single institution should define it alone. Regulators should establish principles around transparency, accountability, and consumer protection. Insurers should translate those principles into governance, and independent oversight should verify that they are being followed. At the same time, society should continue debating what level of algorithmic influence is acceptable.

Ultimately, trust depends on three things: whether people understand why a recommendation is made, whether they have genuine alternatives, and whether they remain free to choose differently.

The assumption I disagree with most is that Generative AI must produce 100 percent accurate answers before it can be trusted in insurance or financial services.

Of course, accuracy matters. These industries make decisions that affect people's money, protection, health, and future. But requiring perfection from the model itself sets the wrong benchmark. Human decisions, legacy systems, manuals, and existing operational processes are not 100 percent accurate either.

A navigation system is not useful because it is always right. It is useful because it combines data, recommendations, driver judgment, road signs, and the ability to change course. Generative AI should be evaluated in much the same way: as one component within a controlled decision-making system, rather than as an autonomous source of truth.

The relevant question is not simply, “Can the model make a mistake?” It is: “Does the redesigned process produce better outcomes than the process we have today?”

That requires different controls for different levels of risk. Low-impact tasks may be automated with monitoring. Higher-impact decisions require source verification, human review, escalation mechanisms, audit trails, and a clear allocation of accountability. In some cases, AI should generate an answer. In others, its role should be to identify missing information, challenge an assumption, or direct the case to the right specialist.

I believe the misconception persists because model accuracy is easy to discuss and measure, while the performance of an entire human-and-AI system is much harder to evaluate. It also allows organizations to postpone difficult questions about workflow redesign, accountability, and governance.

The goal should not be a perfect model. It should be a safer, more accurate, and more accountable system than the one it replaces.

When I developed this framework in 2024, we were comparing several foundation models, and most of them did not yet reach even the “Novice” level in insurance expertise.

We therefore tested how far they could improve through fine-tuning and retrieval-augmented generation, or RAG. The hypothesis was that if we could build a model with a reliable foundation of basic insurance knowledge, it could serve as a shared industry-level model rather than requiring every insurer to begin from zero. Such a model could help raise the baseline capability of the industry as a whole.

Since then, the situation has changed significantly. General-purpose models now possess much of the basic professional knowledge that we were trying to build into them. At least in terms of written examinations and knowledge-based tasks, many models have moved beyond the level we defined as “Novice.”

The central challenge today is therefore no longer how to teach AI basic insurance knowledge. It is how to teach it the proprietary knowledge, practices, judgment criteria, and tacit assumptions of an individual insurer.

Some insurers are developing proprietary or domain-specific models, but the economics remain a major barrier. Leaders must decide which business areas justify that investment, what level of expertise is actually required for each task, and whether fine-tuning, RAG, or a general-purpose model with appropriate controls is sufficient.

More importantly, the model cannot be designed in isolation. The operating model, human review, escalation paths, accountability, and governance must be developed at the same time.

In 2024, the question was how far we could develop the model. Today, the more important question is how precisely an insurer should develop it—and how the organization must change around it.

I do not think technological sovereignty means that Japan or Europe must build every AI model themselves. The real issue is avoiding excessive dependence on a small number of providers.

This concern has become much more concrete. In June 2026, the European Systemic Risk Board raised its assessment of systemic cyber risk to “severe.” It also warned that the concentration of leading AI providers outside the EU creates strategic dependency and geopolitical risk.

This shows that AI dependency and cybersecurity can no longer be treated as separate issues. Frontier models can increase the speed and scale of cyberattacks, while financial institutions themselves are becoming dependent on the same small group of model and cloud providers.

The risk is therefore not only that an insurer may be attacked. A provider may change a model, restrict access, increase prices, or withdraw a service. If many insurers rely on the same provider, one disruption could affect the market as a whole.

I would therefore consider AI dependency comparable to cybersecurity risk and potentially systemic. The biggest danger is concentration.

For Japan and Europe, the first priority should be practical interoperability: common evaluation standards, the ability to switch between models and cloud providers, and shared insurance-specific testing environments.

The goal is not to stop using American technology. It is to use the best technology without allowing one provider to become a single point of failure.

I think Japan offers a slightly different lesson from what many overseas insurers expect.

Many Japanese insurers, both life and non-life, now describe business diversification as a strategic priority. Elderly care has become one of the most important areas, with Nippon Life's acquisition of Nichii Holdings and SOMPO's long-term investment in care services being well-known examples.

Interestingly, when I speak with insurers in Europe and Asia, this is one of the topics they ask about most often.

What makes Japan different is that these care businesses are generally not viewed simply as a way to sell more insurance.

Of course, insurers gain valuable customer relationships and real-world data through these services. But the primary objective is to build a sustainable care business, not merely to generate future insurance sales. I believe this focus has actually strengthened public trust.

Looking ahead, I think this becomes even more important.

As AI becomes more capable, many traditional insurance functions—such as product comparison, underwriting support, and even parts of claims handling—will become increasingly standardized. Insurance itself may become more of a commodity.

Competitive advantage will therefore come less from the insurance contract itself and more from the broader services that surround it: healthcare, elderly care, child and family support, wellness, and risk prevention.

In other words, AI may make insurance products more similar, but it will make the surrounding ecosystem more important.

The lesson from Japan is not simply to diversify. It is to build services that customers genuinely value, even if they never lead directly to an insurance sale.

I would not frame the main obstacle as regulation.

Regulation matters, but for financial institutions it is only the floor. The deeper question is whether customers can still trust institutions when AI is involved in decisions that are increasingly difficult to explain.

Trust cannot rely on perfect transparency. Customers have never understood every actuarial model, underwriting rule, credit model, or investment algorithm. They trust financial institutions because they expect them to remain competent, controlled, and accountable.

Generative AI makes this harder because it can produce convincing but inaccurate answers. The risk is not only error. The greater risk is blurred responsibility.

So I think the future foundation of trust should be accountable control.

This means three things.

First, institutions should clearly disclose when AI is used and provide a process through which customers can ask for the basis of an AI-assisted decision.

Second, AI systems should be continuously tested and monitored, especially for errors, bias, and unexpected behavior after deployment.

Third, AI should be used in a hybrid model. Where decisions must be deterministic, auditable, or strictly compliant, rule-based systems are likely to remain important. By contrast, Generative AI is better suited for tasks such as summarization, triage, pattern detection, recommendation, and decision support.

Customers do not need to understand every parameter of a model. They need to know when AI is used, how they can challenge a decision, and whether the institution remains responsible for the outcome.

So the key point is this: trust in AI will not come from perfect transparency. It will come from accountable control, practical explainability, and clear boundaries around where AI should and should not be used.

I think the biggest misunderstanding is that Japan often sees Europe as too slow and regulatory, while Europe often sees Japan as too cautious and consensus-driven.

Both views contain some truth, but they miss the deeper logic.

Europe is not simply regulating innovation. It is trying to make innovation socially legitimate. In areas such as AI, data, sustainability, and operational resilience, Europe asks: under what conditions can society trust this innovation at scale?

Japan, on the other hand, is not simply resistant to change. It often approaches innovation through long-term relationships, service operations, and gradual refinement. The process may look slow, but the focus is often on how a service can actually be delivered, maintained, and trusted over time.

So I would put it this way: Europe is strong at designing trust. Japan is strong at embedding trust into operations.

Japan should learn from Europe how to make governance more explicit—through clearer rules, accountability, data governance, and model oversight.

Europe should learn from Japan not necessarily the speed of AI implementation, but the discipline of embedding new services into long-term customer relationships and daily operations.

A good framework is important, but long-term value is created only when innovation becomes part of the customer experience and the operating model.

The future needs both: European-style governance and Japanese-style implementation discipline.

The most likely structural transformation is that insurers will move from being risk carriers to becoming long-term risk and life-service partners.

Insurance companies already hold some of the most important customer data: health, family, assets, income, retirement, mobility, accidents, and business continuity. As insurers expand into healthcare, elderly care, wellness, family support, and prevention services, this data will become broader, more continuous, and more valuable.

This creates a role that few other industries can play. Insurers can support customers not only at a single point of loss, but across the entire timeline of risk: before, during, and after an event.

So the future question will not simply be, “Who provides the best insurance product?” It will be, “Who can help customers live more safely, healthily, and confidently over time?”

The assumption I think is most underestimated is the value of human involvement.

Many people assume that as AI becomes faster and more capable, human interaction will naturally become less important. I think the opposite may happen. As products, advice, and claims processes become more automated, the moments where humans are involved will become more valuable.

People still want to be understood by other people, especially in moments of illness, ageing, loss, uncertainty, or family transition. This is why high-touch services do not disappear even when cheaper digital alternatives exist.

The real opportunity for insurers is not to replace humans with AI. It is to use AI to remove routine work and allow human expertise, empathy, and judgment to be used where they matter most.

In that sense, insurers have a unique opportunity. They can combine data, capital, prevention, services, and human trust to design systems that help people live better lives, not only manage financial losses.

So I would say the next decade will not only redefine insurance products. It will redefine the human role inside insurance.

The winners will be the companies that use AI to make insurance more human, not less.