McKinsey Maps the Agentic AI Path to Core System Overhauls

From tracking carrier technology modernization programs across a dozen 10-K filings, the consistent pattern has been multi-year delays and cost overruns, which is exactly the problem McKinsey’s latest framework aims to break. When both McKinsey and BCG publish substantively the same thesis within weeks of each other, the signal is worth parsing for actuarial teams whose pricing, reserving, and reporting workflows sit directly downstream of whatever core system their carrier decides to migrate.

McKinsey’s Q1 2026 article, “Can Agentic AI (Finally) Modernize Core Technologies in Insurance?”, argues that the insurance industry has long understood the need to transform core systems but has consistently failed to execute. The firm introduces a “modernization factory” concept: instead of bespoke, one-off migration programs that spiral in cost and duration, carriers can build repeatable, scalable modernization capabilities where agents produce auditable outputs with human-in-the-loop controls at each step.

BCG’s January 2026 companion piece reinforces this thesis with two make-or-break best practices: deploying agentic AI through every phase of modernization and employing zero-based design that resets the conceptual approach rather than recreating legacy configurations on new platforms. The convergence of two competing strategy firms on the same structural argument within the same quarter suggests this is more than a single consultant’s pitch; it reflects genuine client demand and early implementation evidence.

The Core Problem: Why Insurance Modernization Programs Fail

The insurance industry’s legacy system burden is well documented. Research from Celent and PwC consistently finds that insurers allocate 70% to 80% of their annual IT budgets to maintaining existing systems, leaving 20% to 30% for innovation and digital projects. Over 70% of insurers continue to operate on legacy platforms built around COBOL and other aging architectures, where specialist contractors command premium rates because few engineers under 50 have worked with these languages.

The failure cases are equally well documented. BCG cites a Central European insurer whose core system program ran for eight years before producing a write-off exceeding $500 million. A Southern European insurer’s claims platform program completed at 500% over budget. Both failures traced to the same root cause: the programs began without sufficient understanding of what the legacy system actually contained. Business rules accumulated across decades of endorsements, riders, product filings, and regulatory patches were not adequately captured before migration design started.

McKinsey frames the bottleneck precisely. In policy administration migrations, the biggest constraints are rarely about writing new code. The binding constraints are the loops of discovery, mapping, testing, reconciliation, and cutover. Each loop involves subject-matter experts (SMEs) manually reviewing legacy logic, documenting business rules, verifying data transformations, and reconciling outputs between old and new systems. When defects surface during testing, the loop resets, consuming additional SME time and extending timelines.

This pattern explains why modernization programs routinely deliver results that are “mixed” at best, per Deloitte’s 2026 Global Insurance Outlook, with many carriers not fully realizing expected returns. The Deloitte outlook notes that moving into 2026, the emphasis has shifted from whether to modernize to how to execute at scale while strengthening data foundations and aligning architecture for AI workloads.

McKinsey’s “Modernization Factory” Framework

McKinsey’s central concept is that agentic AI transforms modernization from a project into a capability. Rather than assembling bespoke teams for each migration wave, carriers can establish a repeatable “modernization factory” that compounds learnings across projects. The factory reuses agents, patterns, and context layers, making each subsequent migration faster and more predictable than the last.

The framework distinguishes itself from previous AI-assisted coding tools by targeting the non-code bottlenecks that actually determine program success or failure. McKinsey identifies productivity improvements ranging from 10% to 90% depending on the specific step and degree of automation:

Migration Phase	Productivity Improvement	Key Agent Capabilities
Discovery & Requirements	15% to 90%	Capturing legacy knowledge at scale, extracting business rules from COBOL and mainframe code, generating documentation
Data Mapping & Quality	20% to 60%	Automated field mapping between source and target schemas, data quality profiling, transformation rule generation
Testing & Reconciliation	High (largest gains)	Automatically generating, executing, diagnosing, and reconciling tests; linking defects to root causes and financial outcomes
Cutover & Hypercare	15% to 90%	Orchestrating cutover runbooks, role-based training generation, policy servicing decision support

The wide range in productivity improvements is intentional, not imprecise. The lower bound (10-20%) reflects scenarios where human oversight remains primary and agents serve as accelerators. The upper bound (up to 90%) reflects scenarios where the agent can operate with minimal human intervention because the task is structured, the inputs are well-defined, and the outputs are independently verifiable.

What makes this a “factory” rather than a tool is the compounding effect. An agent that learns the business rules of one product line can apply that contextual understanding to adjacent lines. Testing patterns proven on one migration wave become templates for the next. The modernization factory model reframes the economics: instead of each migration starting from zero, subsequent waves inherit context, validated patterns, and trained agents from prior deployments.

BCG’s Zero-Based Design and the Discovery Breakthrough

BCG’s framework adds a critical design principle: zero-based thinking. Rather than replicating legacy system configurations on modern platforms (which locks in decades of accumulated complexity), zero-based design starts from the new platform’s native capabilities and asks which legacy behaviors genuinely require preservation versus which exist only as artifacts of historical constraint.

This distinction matters for actuarial teams. Many legacy policy administration systems encode rating algorithms, factor tables, and territory definitions in ways that were shaped by mainframe-era computational constraints rather than actuarial intent. A zero-based approach allows actuaries to implement pricing logic in modern statistical frameworks rather than recreating COBOL rate tables in a new language. The result is not just a migrated system but a modernized pricing capability.

BCG identifies the discovery phase as potentially the most productive area for agentic AI, though usage is not yet widespread. AI agents can analyze legacy COBOL applications, automatically extracting business rules, generating documentation, and creating process maps. This directly addresses the root cause of the documented $500M+ failures: programs that started without understanding what the legacy system contained.

The combined effect of both practices, per BCG, is a program that is “financially more feasible, much shorter in duration, and considerably less risky.” Given that traditional modernization programs average 3-5 years in duration with budget overruns of 50-500%, even modest improvements in predictability and cost control represent significant value creation.

The Vendor Ecosystem Response

The consulting frameworks are not purely theoretical. Core system vendors are embedding agentic capabilities directly into their platforms, creating the infrastructure layer that McKinsey’s factory concept requires.

Guidewire’s Palisades Release (April 2026): Guidewire launched ProNavigator, an AI assistant built directly into InsuranceSuite and InsuranceNow, as part of the Palisades cloud release. The tool provides underwriters, claims adjusters, and billing specialists with context-aware insights within existing workflows. Pricing teams gained access to real-time pricing and quoting through PricingCenter, now integrated into PolicyCenter for personal and high-volume commercial lines. Finance teams received improved payment and refund reconciliation capabilities with detailed funds tracking and audit trails.

Duck Creek’s Agentic AI Platform (April 2026): Duck Creek launched what it calls an “insurance-native Agentic AI Platform” combining core system data, insurance domain models, and neuro-symbolic reasoning. The platform includes three layers: Agentic Intelligence (fine-tuned LLMs with insurance knowledge graphs), Agentic Orchestration (a unified layer for designing and coordinating AI agents), and AI Assurance (governance, audit trails, and compliance controls). Duck Creek also launched an Agentic Product Configurator claiming 50% acceleration in policy product implementation. BCG projects up to $80 billion in annual P&C impact from this category of platform capability.

The vendor launches validate McKinsey’s thesis from the supply side: if core system providers are building agent orchestration directly into their platforms, the infrastructure for the “modernization factory” is becoming available as commercial capability rather than requiring each carrier to build from scratch.

The Infrastructure Cost Challenge

McKinsey’s separate research on reimagining tech infrastructure for agentic AI introduces a complicating factor. With AI workloads expanding, IT infrastructure costs are projected to increase two to three times by 2030 while budgets remain flat. This creates a dual challenge for CTOs: they must upgrade infrastructure to support agentic AI while using agentic AI itself to contain rising costs.

Currently, only 10% of respondents in any given business function say their organizations are scaling AI agents, despite 62% experimenting with or piloting them. The gap between experimentation and scale mirrors the broader 82% adoption, 7% scale finding from Sedgwick’s insurance-specific data. McKinsey’s insight is that infrastructure itself must become “the platform that orchestrates and governs how work is executed across the enterprise” rather than merely a support function.

For insurance carriers planning modernization investments, this means the business case cannot assume static infrastructure costs. The modernization factory requires compute, storage, and orchestration infrastructure that will itself need investment. The economic argument depends on whether the productivity gains (10-90% per step) exceed the incremental infrastructure cost of running the agents that produce them.

Actuarial Implications: Why This Is Not Just an IT Problem

Policy administration system migrations directly affect actuarial workflows in ways that actuaries must anticipate rather than react to. The implications span pricing, reserving, reporting, and data governance:

Pricing system continuity. Rating algorithms encoded in legacy systems must produce equivalent results on modern platforms during parallel-run periods. Any deviation between legacy and target system premiums creates rate adequacy risk. Actuaries must define acceptable tolerance thresholds for premium calculations during migration and establish reconciliation protocols that verify rating factor application across platforms. The agentic AI approach to testing and reconciliation, where agents automatically compare outputs and link discrepancies to root causes, directly serves this need.

Reserve data integrity. Loss triangles depend on consistent data definitions over time. A policy administration migration that changes field mappings, date conventions, or transaction categorizations mid-triangle creates discontinuities that complicate development factor selection. Actuaries need to understand exactly which data transformations the migration introduces and how those affect historical development patterns. McKinsey’s emphasis on auditable outputs at each step provides the documentation trail that actuaries need for ASOP No. 43 compliance.

Regulatory reporting continuity. State statistical reporting, NAIC annual statement supplements, and statutory filing data all flow from policy administration systems. A migration that introduces mapping errors in statistical codes, coverage forms, or territory classifications can trigger regulatory inquiries months or years after cutover. The zero-based design approach must account for regulatory reporting requirements as constraints that cannot be simplified away.

Experience study disruption. Mortality, morbidity, and lapse studies depend on consistent policyholder data across observation periods. For life and health carriers, a policy administration migration that alters how policy status, coverage amounts, or exposure periods are recorded can compromise multi-year experience studies. Actuaries conducting assumption setting must know whether the underlying data experienced a platform change and how to adjust for any structural breaks.

Model validation demands. Under ASOP No. 56, actuaries must evaluate whether models remain appropriate when inputs change. A core system migration changes the input layer for every downstream model. Pricing models, reserving models, and capital models that received data from the legacy system must be revalidated against the new platform’s output format and content. This creates a temporary surge in model validation workload that actuarial departments must plan for.

The Trust and Governance Dimension

McKinsey’s State of AI Trust in 2026 report provides context for the governance challenges that modernization programs will face. The average Responsible AI (RAI) maturity score increased to 2.3 out of 4 in 2026, up from 2.0 in 2025. Only about one-third of organizations report maturity levels of three or higher in strategy, governance, and agentic AI governance specifically.

Nearly 60% of respondents cite knowledge and training gaps as the primary barrier to implementing responsible AI practices, up from approximately 50% the prior year. For insurance carriers, this translates directly to the governance challenge of deploying AI agents in core system migration work where errors propagate to pricing, billing, and claims outcomes.

The “human-in-the-loop” controls that both McKinsey and BCG emphasize are not decorative. They represent the mechanism by which carriers maintain accountability when AI agents perform tasks that historically required actuarial or underwriting judgment. The modernization factory concept explicitly preserves human checkpoints at each phase transition, creating verifiable decision points where a qualified professional confirms that outputs meet requirements before the next phase begins.

This governance architecture aligns with the NAIC’s emerging agentic AI governance framework from Spring 2026, which identified the need for defined oversight boundaries, output verification, and decision traceability in autonomous AI systems. Carriers that build modernization factories with proper governance from the start will face less friction when regulatory requirements mature.

Early Adopter Advantage and Market Timing

McKinsey’s argument for moving early rests on the compounding nature of the factory model. Carriers that build modernization capabilities now accumulate agent training data, validated patterns, and organizational experience that make each subsequent migration faster. Late movers do not merely delay benefits; they compete for scarcer consulting and vendor capacity once the approach reaches mainstream adoption.

The timing signal from Deloitte’s 2026 outlook reinforces this urgency. With AI-driven workloads expanding and cyber insurers treating legacy systems on unsupported platforms as grounds for policy non-renewal or premium increases of 40-60%, the cost of maintaining legacy systems is rising faster than the cost of migration. The crossover point where maintaining legacy becomes more expensive than migrating is approaching for many carriers.

Morgan Stanley’s projection of $9.3 billion in AI-generated operating income for P&C insurers by 2030 provides the revenue context. If core system modernization is the prerequisite for capturing those AI gains (because modern platforms enable the data access, real-time processing, and API connectivity that AI applications require), then the modernization decision is not an IT infrastructure question alone. It is a competitive positioning question with direct implications for expense ratios, combined ratios, and market share.

From tracking these consulting firm publications across cycles, the convergence of McKinsey and BCG on the same thesis within the same quarter is unusual. These firms typically differentiate their frameworks rather than validate each other’s conclusions. When they align, it generally reflects strong client demand and early evidence from live engagements rather than speculative thought leadership.

What Actuaries Should Do Now

Engage early in modernization planning. If your carrier is evaluating or initiating a core system migration, actuarial teams must be at the table during the discovery phase, not introduced during testing. The business rules that agents will extract from legacy systems include pricing logic, rating algorithms, and factor tables that only actuaries can validate for correctness. Waiting until user acceptance testing to discover that a rating factor was mapped incorrectly is the exact rework loop that McKinsey’s framework exists to prevent.

Define actuarial acceptance criteria before migration starts. Establish quantitative thresholds for pricing reconciliation (e.g., premium calculations must match within 0.01% for in-force policies), reserve data continuity (e.g., no field mapping changes without documented triangulation impact), and regulatory reporting accuracy (e.g., statistical plan codes must map one-to-one with zero transformation errors). These criteria become the requirements that AI agents test against during reconciliation.

Plan for temporary model validation surge. Every downstream model that consumes data from the migrating system will require revalidation under ASOP No. 56. Build this workload into departmental planning 6-12 months before projected cutover dates. Identify which models are most sensitive to input format changes and prioritize their validation.

Evaluate zero-based design opportunities for pricing. BCG’s zero-based principle means actuaries can advocate for implementing modern rating algorithms rather than recreating legacy rate table structures. If the migration is happening regardless, use it as the opportunity to modernize pricing logic, territory definitions, and factor structures that have accumulated complexity without actuarial justification.

Monitor vendor platform capabilities. Guidewire’s PricingCenter integration and Duck Creek’s Agentic Product Configurator represent a shift toward platforms that natively support actuarial workflows. Evaluate whether your carrier’s chosen platform offers the real-time pricing, rating algorithm flexibility, and data auditability that modern actuarial work requires.

Document the baseline before migration. Before any system change occurs, ensure that current-state rating logic, data definitions, and reporting mappings are documented thoroughly enough that post-migration reconciliation has a clear reference point. If the legacy system’s documentation is incomplete (as BCG’s failure cases demonstrate), the discovery phase is the time to fill those gaps, not cutover week.