The path to AI value in P&C runs through the unglamorous work of integration, modernization-in-place, governance, and operational discipline – not another giant rip-and-replace program wearing an AI costume.

Key Takeaways

• AI success starts with modernizing, not replacing. Most insurers can unlock AI value by improving integration, data, and governance rather than rebuilding core systems.
• Enterprise AI requires strong foundations. Secure architecture, quality data, and governance—not AI tools alone—determine long-term success.
• The biggest AI gains come from augmentation. Today’s highest-value use cases enhance underwriting, claims, and operations instead of replacing human expertise.
• Focus on business outcomes. Modernization should improve efficiency, customer experience, and operational resilience not simply deploy new technology.

When carriers talk about modernization in the AI era, the conversation too often collapses into two bad choices: stop everything and rebuild for AI, or ignore AI and keep running what already exists.

This is a false dilemma. It is also an expensive one.

The better question is not whether an insurance platform is new, but whether it is modern enough to operate sustainably, integrate cleanly, support the business, survive audit, and create room for AI-enabled improvement. A system that does those things is not necessarily legacy. It may simply be incumbent: stable, useful, and maybe not exciting. But insurance runs on reliability, capital discipline, and paying claims when a policyholder’s day goes sideways.

The industry is being told to move faster, replace more, spend more, and become fully agentic instantly or risk collapse. That is the new version of an old transformation sermon — replaced or be doomed. The technology has changed. The salesy call-to-action has not.

AI is real. Generative AI is powerful. Agentic AI may matter enormously. But none of that means carriers should abandon the hard-won architectural progress they have made. For most insurers, the fastest path to value is making existing estates AI-ready through integration, data improvement, workflow redesign, and targeted modernization.

AI is not a product

One reason the conversation gets confused is that people talk about AI as if it were a product. It is better understood as a set of capabilities and methods applied to business and technical problems. AI products exist – copilots, coding tools, agent platforms, underwriting assistants – but they package AI capabilities. They are not AI itself.

That distinction matters when part of the leadership is asking for an AI roadmap and expects a parade of apps, while the CIO is thinking about platforms, security, permissions, integration, telemetry, change management, and support. Some outside of tech still tend to think of AI as a standalone thing, like an app. The technologist is thinking of the plumbing.

Both are necessary. Only one gets applause at a conference.

That is why the proposal-to-production gap remains so stubborn. Pilots are easy. Production is hard. MIT’s 2025 State of AI in Business report found that despite $30 billion to $40 billion in enterprise GenAI investment, only about 5% of integrated AI pilots were extracting millions in value. McKinsey’s 2025 global AI survey showed broad adoption — 88% of respondents reported regular AI use in at least one business function — but most organizations still had not scaled AI broadly.

Insurance is not exempt. BCG reported in 2025 that only 7% of insurers surveyed had brought AI systems to scale, while roughly two-thirds remained in pilot. This is not because insurers are lazy or allergic to innovation. It is because regulated transactions require explainability, auditability, security, model governance, data lineage, human accountability, and operational support. In other words: enterprise-class technology.

The production gap is an enterprise technology problem

A proof of concept can be brilliant and still be nowhere near production. Many AI efforts start in digital teams, AI labs, data groups, or business-led experiments. Good. We need experimentation. But when a pilot aims to become a durable enterprise capability, the questions change.

Who monitors it? Who secures it? Who tests it? Who owns break-fix? What happens when the model changes? What data can it see? What answer did it give last month, and why? How does it fail safely? What does the regulator see?

Those are not innovation theater questions. They are serious operational and technology. Enterprise technology teams answer them every day.

The NAIC’s model bulletin on insurer use of AI systems emphasizes fairness, accountability, compliance, transparency, and safe, secure, fair, and robust systems. That is not a suggestion to move slowly for sport. It is a reminder that carriers cannot let machines make decisions they cannot explain and then hope the audit team brings donuts.

Modern enough beats new for the sake of new

Before replacing a core or complex incumbent system, carriers should ask a simple set of questions:

• Can the system integrate through APIs, events, messaging, or other reliable patterns?
• Does it deliver the base functional needs of the business without daily heroics?
• Does it run on a supported technology stack with a clear path forward?
• Does it operate at reasonable cost relative to value?
• Are resources and documentation available — and, where documentation is thin, can AI help reconstruct the estate?

If the answer is mostly yes, the system may be modern enough. It may need wrapping, refactoring, re-skinning, data remediation, or targeted engineering. But it may not need replacement.

This is not an argument for keeping everything forever. If a platform cannot integrate, cannot be secured, cannot support the business, cannot be staffed, or is running on a dying stack, replacement may be the only responsible option. Some systems are not old; they are dead. But many are simply unfashionable.

And unfashionable is not a business case.

The distraction tax

Our industry, and the consultancies and PE firms that fuel its change, underestimate the distraction tax of giant transformation and massive system replacement programs. They consume capital, attention, goodwill, executive oxygen, business patience, and the scarce people who understand how the company actually works.

A large core replacement often takes years and too often becomes a lateral move with a slightly better user interface and a much larger invoice. During that time, underwriting workflow stalls, data work gets deferred, digital initiatives wait, and AI enablement becomes a talking point. The business is told to be patient because transformation is coming. Eventually. Maybe after release 17.

There are cases where big transformation is necessary. But skepticism should be the default. If 80% of the value can be achieved with 10% to 20% of the cost through modernization-in-place, integration, workflow improvement, and selective bolt-ons, that deserves a serious look. This is not timid. It is disciplined. It is also how you avoid spending hundreds of millions to end up roughly where you already were, only more tired.

AI changes the build-versus-buy math

Generative AI and AI-assisted development are changing build-versus-buy. Coding tools, test generation, documentation tools, and agentic development practices can reduce the cost and risk of building around existing systems. They do not eliminate engineering discipline. If anything, they make it more important because more people can create more code more quickly. That can be wonderful. It can also create a spectacular mess at machine speed.

Letting someone vibe-code directly into production in insurance is a basic control (and thought-process) violation, not to mention an audit disaster waiting to happen. The better pattern is to use AI to accelerate ideation, prototyping, code analysis, documentation, QA, and targeted engineering – then bring the output into proper enterprise SDLC practices leading to productionization.

We do not have to replace the whole Christmas tree when the issue is that several ornaments are ugly and one strand of lights is suspiciously warm. Build or buy the missing capabilities. Integrate them properly. Keep the stable parts stable. Replace what is broken.

Where AI value is real now

The most credible AI value in insurance today is not full replacement and autonomy. It is augmentation: document summarization, submission ingestion and triage, document classification, underwriting workbench support, correspondence drafting, coding, QA, documentation, workflow orchestration, exception routing, better search, better synthesis, and less swivel-chair work.

Accenture’s 2025 underwriting survey of 430 senior insurance underwriting executives found that respondents expected AI adoption in underwriting to rise from 14% today to 70% over the next three years. Directionally important. But the practical word is augmentation: better information, faster analysis, fewer manual steps, and improved decision support. Not the disappearance of underwriters into a puff of agentic smoke.

Gallup’s late-2025 workforce research is a useful reality check. AI use at work was rising, but only 10% of U.S. employees reported daily use in Q3 2025, and 23% used it a few times per week or more. By Q4 2025, daily use had risen to 12% and frequent use to 26%. Meaningful growth, but not yet enterprise transformation.

Agentic AI will likely accelerate the curve. Celent reported that 22% of insurers in its 2025 GenAI survey planned to have an agentic AI solution in place by year-end 2026. Notable, but most insurers will not. For now, the sensible agentic pattern in insurance is guided automation with humans accountable for decisions, not unsupervised autonomy in regulated transactions.

Data, integration, and governance are the actual AI strategy

The uncomfortable truth is that AI makes old problems more visible. Bad data becomes more dangerous. Weak access controls matter more. Unclear process ownership gets harder to hide. Brittle integration becomes the bottleneck. AI does not magically fix these issues. It amplifies them.

That is why the path to AI value in P&C is really an enterprise architecture conversation. Carriers need clean data, usable metadata, sound identity controls, sane integration, observability, model governance, test harnesses, and human-in-the-loop controls matched to transaction risk.

This is not glamorous work. It will not trend on social media. No one gets a standing ovation for data lineage or API rationalization. But this is where AI value becomes bankable: the difference between a clever demo and a capability the business can rely on.

A pragmatic modernization playbook

For carriers deciding where to go next, be brutally practical:

• Inventory the estate honestly. Separate systems that are old-but-working from systems that are truly unfit. Use Gartner’s TIME method (Tolerate, Invest, Migrate, Eliminate) or your favorite equivalent.
• Apply the modern-enough test. Integration, functionality, supported stack, cost-to-value, and maintainability should drive decisions — not age alone.
• Decompose transformation into smaller value releases. Avoid giant one-shot programs unless there is genuinely no alternative.
• Modernize around business outcomes. Underwriting speed, submission quality, claims cycle time, service quality, product agility, expense reduction, and risk control beat architectural vanity metrics.
• Use AI first where risk-adjusted value is obvious. Summarization, ingestion, classification, drafting, coding, testing, and workflow routing are better early bets than autonomous underwriting or claims adjudication.
• Put enterprise technology in the center of production AI. Innovation can begin anywhere. Production cannot live just anywhere.

The people problem: The Other Bubble

There is also a people issue. We are asking technologists to remain fluent in 20 years of existing systems while mastering AI, data, and security. All the old systems and projects, and all the new ones. At the same time. Forever, apparently.

This is a Skillset Bubble. It is not sustainable unless leaders make choices. If AI and modernization matter, something else must be deprioritized. The same people keeping the policy admin system alive at 2 a.m. cannot also scale every AI pilot, remediate every data issue, migrate every platform, and attend every vendor demo involving the word agentic.

The companies that win will not be the ones with the longest AI backlog. They will focus people, money, and leadership attention on fewer high-value use cases and the architecture required to scale them.

The answer is not stagnation

None of this is an argument for complacency. Modern enough does not mean leave everything alone. It means understand what you have, improve what matters, replace what must be replaced, and avoid confusing novelty with value.

The P&C industry has made enormous progress. It is easier to underwrite, bind, service, and claim than it was 15 years ago. Insurtechs are more mature. Core platforms are more capable. Data expectations are clearer. Cloud and API patterns are more established. AI now gives us capabilities that can make the next decade materially better.

But AI does not excuse bad architecture, and it does not justify waste. If anything, it raises the premium on disciplined enterprise technology. Carriers that make existing systems AI-ready, modernize surgically, integrate intelligently, and govern carefully will move faster than those waiting for a perfect AI-native core platform to save them.

Insurance is built to manage risk. We should take safe chances, experiment, modernize, and use AI aggressively where appropriate. But we should not let machines outrun our ability to govern and keep them safe.

The path forward is not reckless replacement or stagnant preservation. It is pragmatic modernization: stable foundations, flexible edges, enterprise-grade AI, and a ruthless focus on business value and customer experience. Not quite as sexy as a clean slate. Much more likely to work.

Frequently Asked Questions

Why should insurance carriers modernize legacy systems instead of completely replacing them for AI?

Replacing core systems is expensive, time-consuming, and carries a high “distraction tax.” Often, 80% of the value can be achieved at a fraction of the cost through targeted modernization, integration, and workflow improvements. Complete replacement should be reserved for systems that are truly dead or unsupported, not just “unfashionable.”

How can a carrier determine if their current system is “modern enough”?

Carriers should ask a few practical questions: Can the system integrate via APIs or messaging? Does it deliver base functional needs reliably? Does it run on a supported tech stack? Does it operate at a reasonable cost? If the answers are mostly yes, the system likely just needs wrapping, refactoring, or data remediation rather than a full replacement.

Where is the most credible value for AI in the insurance industry today?

The real value right now lies in augmentation, not full autonomy. High-value use cases include document summarization, submission triage, underwriting workbench support, correspondence drafting, and workflow orchestration. It’s about improving decision support and reducing manual work, not replacing human underwriters.

Why do so many AI pilots in insurance fail to make it into full production?

The gap between pilot and production is an enterprise technology problem. Regulated insurance transactions require explainability, auditability, security, model governance, and human accountability. Pilots are easy to build in a lab, but scaling them requires robust enterprise-class plumbing, data lineage, and governance.

What is the “distraction tax” of large transformation projects?

The distraction tax refers to the massive drain on capital, executive attention, and business patience caused by giant system replacement programs. During these multi-year projects, other important work—like underwriting workflow improvements, data remediation, and AI enablement—often stalls while the business waits for the new system to eventually launch.

Does AI change the “build versus buy” calculation for insurers?

Yes. Generative AI and AI-assisted coding tools reduce the cost and risk of building around existing systems. However, this makes engineering discipline even more important to avoid creating a mess at machine speed. AI should be used to accelerate development, but the output must still go through proper enterprise software development life cycles (SDLC) before hitting production.

What is the actual foundation of a successful AI strategy?

The uncomfortable truth is that AI makes old problems more visible. A successful AI strategy is actually an enterprise architecture strategy. Carriers need clean data, usable metadata, sound identity controls, sane integration, and human-in-the-loop controls. Without this unglamorous foundational work, AI cannot scale safely.