Legacy Systems as AI Training Grounds in Insurance

Legacy systems as AI training grounds and hidden institutional memory

Legacy insurance systems are not just old technology; they are rich, empirical datasets that can train and guide modern AI-first modernization efforts when used under strong governance and with the right tooling. They capture decades of real decisions about pricing, underwriting, claims, and compliance that are rarely documented anywhere else.

Consider a 30-year-old policy administration platform. It may run on mainframe technologies that few engineers now master, but it also encodes every product filing, rule change, and operational workaround that has kept the business compliant and profitable. Boston Consulting Group notes that 74% of large core transformations underperform or fail, in part because this embedded knowledge is poorly understood before replacement (BCG). Legacy platforms are difficult not only because they are old, but because they are dense with undocumented rules.

Agentic AI changes how insurers can approach that density. Instead of relying solely on human subject matter experts to reconstruct rating algorithms or underwriting rules from memory, AI agents can observe how real work happens in production systems. They can watch which fields underwriters touch on specific screens, which documents are attached for certain risks, and how exceptions are escalated. They can align that behavioral evidence with historical transactions to infer patterns: which endorsements tend to trigger referrals, which combinations of deductibles and limits are offered in certain territories, or how billing plans vary by channel.

In practice, that means building a knowledge extraction layer around existing systems of record. AI models can parse legacy code, configuration tables, and batch scripts to highlight candidate business rules. Optical character recognition and document AI can read policy forms, state endorsements, and procedures, tying clauses to system behavior. Process mining tools reconstruct end-to-end workflows from event logs, revealing where manual workarounds compensate for system gaps. A consultancy case study on legacy modernization stresses that “visibility and reconstruction” of this kind is the prerequisite for safe transformation (Thoughtworks).

The result is a living map of institutional memory. Instead of scattered documents and fragile tribal knowledge, insurers gain a structured graph connecting products, rules, screens, state variations, forms, and test cases. For example, a carrier can ask an internal AI system: “Show me every place where coastal property accumulations affect rating or underwriting appetite in Florida,” and receive not only code snippets but also associated filings, forms, and historical decisions. This is very different from a generic knowledge base; it is grounded in the real behavior of the core.

There is practical near-term value as well. With this institutional memory exposed, insurers can build digital twins of key processes such as submission intake, quote-to-bind, or first notice of loss. These twins simulate how changes in rules or workflows would affect cycle time, loss ratio, or regulatory steps before any code is changed in production. A carrier could, for instance, experiment with new renewal routing rules in the digital twin, measuring expected impact on underwriter workload and service-level agreements before rollout.

For leaders, the main mindset shift is to stop treating legacy purely as a liability and start treating it as training data. Every policy, claim, endorsement, and billing change is an example of how the organization has interpreted risk, regulation, and customer expectations. AI gives insurers a practical way to surface, test, and refine that embedded memory rather than discarding it during replacement.

Why AI should precede core replacement in insurance modernization

Reversing the traditional sequence and applying AI-first modernization before core replacement allows insurers to de-risk transformation, preserve institutional knowledge, and generate early operational wins instead of waiting years for a big-bang cutover. The familiar recipe—replace the core, clean data, expose APIs, then deploy AI—was designed for a pre-agentic era.

That sequence fails for many carriers because it assumes they already understand their legacy estate well enough to map every rule, dependency, and exception into the new system. In reality, decades of changes, local adaptations, and market-specific products create complexity that even experienced teams struggle to articulate. Thoughtworks highlights the “it just works” paradox: everyone depends on the system, yet very few can explain why it behaves as it does (Praxent). When organizations move ahead without clarity, they risk rating discrepancies, compliance gaps, and operational surprises during migration.

Putting governed AI at the front of the sequence changes that risk profile. Instead of starting with large replacement projects, insurers start by observing reality. AI agents mine logs to reconstruct real workflows, compare those workflows with official procedures, and highlight where the two diverge. They read configuration and code to identify overlapping rules or contradictory thresholds. They generate candidate documentation that human experts validate, accelerating the creation of a reliable knowledge base.

From there, AI can support “outside-in” modernization. Rather than tearing out core systems, carriers wrap them with APIs, event streams, and modern user interfaces that rely on the knowledge extracted earlier. For example, a claims triage service might sit on top of an existing core, subscribing to events such as first notice of loss and coverage verification, then generating a prioritized work queue for adjusters. The underlying system remains unchanged, but the experience and control plane are modernized.

External benchmarks support this sequence. BCG argues that agentic AI can handle much of the discovery and documentation work in core modernization, making later migration safer and faster (BCG). Other practitioners describe how generative AI can explain legacy modules, identify dependencies, and suggest refactoring options—especially when documentation is missing—turning an opaque codebase into something more navigable (Thoughtworks).

This AI-first approach also produces earlier business value. Instead of waiting for a multi-year core replacement to complete, carriers can deploy agents that assist with submission intake, policy servicing, or loss-run summarization within months. For instance, an agent can read loss runs and engineering reports, extract key fields, and pre-fill an underwriting workbench, cutting data entry time and improving consistency. At the same time, every interaction becomes new training data: corrections by underwriters help refine models, and workflow telemetry feeds the process digital twin.

Governance is the non-negotiable counterpart to going AI-first. Agent identities, least-privilege access, approval workflows, and complete audit trails are required, especially when agents interact with rating, billing, or claims payment. AI should never introduce unapproved changes to systems of record; instead, it should generate recommendations, tests, and documentation that human teams validate. A safe pattern is: observe, document, simulate, and then modernize bounded components, rather than letting agents autonomously refactor mission-critical logic.

For modernization leaders under pressure, the core argument is pragmatic. Delaying AI until after core replacement forfeits a powerful way to understand and de-risk the very transformation they are trying to execute. By putting AI first—within a governed framework—insurers can move faster while actually reducing uncertainty.

From technical debt to training data: reversing the modernization playbook

When insurers reframe legacy as a source of AI-first modernization training data, they can replace big-bang core projects with incremental, evidence-driven change guided by agentic AI, process intelligence, and digital twins. The modernization playbook shifts from a single high-stakes replacement to a sequenced program of observation, wrapping, simulation, and targeted refactoring.

The starting point is acknowledging that “technical debt” is only half the story. The same systems that slow integration also hold the clearest record of how the insurer actually operates. Claims notes, endorsement histories, abandoned quotes, and billing exceptions all capture cause-and-effect relationships that can train AI models. Carriers can mine this historical data to identify patterns such as which submission profiles lead to bound policies, which claim characteristics correlate with litigation, or which billing behaviors predict cancellation.

In a reversed playbook, this data is not just moved into a new core. It is used to build an insurance process digital twin: a live model of the journey from quote to bind, from first notice of loss to settlement, and from billing to collections. This twin reflects not only the intended workflows but also the real paths, including manual detours and exceptions. AI agents can then run “what if” experiments inside the twin: What happens to quote cycle time if certain documents become mandatory at intake? How does a new triage rule affect adjuster workload and settlement speed for specific claim segments?

At the same time, AI helps convert messy legacy behavior into cleaner, more modular services. For instance, a rating engine buried deep within a monolithic policy system can be treated as a black box while AI generates large test suites from historical policies. Those tests become the benchmark for a new rating service. The new service runs in parallel for a period, with AI highlighting discrepancies and providing explainable summaries to actuarial and compliance teams. Only after parity is proven and governance conditions are met does traffic gradually shift from the legacy engine to the new one.

External perspectives reinforce this modular strategy. Consultancies and engineering firms describe how generative AI now supports code explanation, interface mapping, and automated test creation for mainframe and midrange platforms, making it more practical to peel off specific capabilities rather than replace everything at once (Thoughtworks). BCG’s guidance on convergence—aligning local adaptations to a common enterprise design—echoes the same pattern: decide what must be consistent, then let AI help translate local variations into that framework (BCG).

For insurers, the benefit is not just technical. Treating legacy as training data creates a bridge between operations, technology, and the business. Underwriters, adjusters, and operations leaders can see their daily work reflected in the digital twin and in the behavior of AI assistants. They can correct and improve those systems in context, rather than reviewing abstract backlogs. Modernization becomes a continuous, collaborative effort grounded in observable outcomes: shorter cycle times, fewer errors, better regulatory traceability, and more resilient operations during peak events.

Ultimately, reversing the modernization playbook is about sequencing and mindset. Instead of assuming modernization is a prerequisite for AI, insurers treat AI as the mechanism that makes modernization safer, faster, and more transparent. By doing so, they preserve the best of their institutional memory, reduce the risks that have stalled past transformations, and position their organizations to evolve with confidence in a more automated, data-intensive insurance landscape.