Why insurance system modernization still keeps getting stuck
Insurance core systems are often digital archaeology. Decades of operation layered actuarial rules, batch processes, custom interfaces and data semantics that today perhaps only two employees in the entire company understand. In its analysis, McKinsey states that insurance management has long understood the need for modernization, but structural costs and risks repeatedly demotivate them.
The biggest barriers represent underdocumented product logic and actuarial settings, semantic gaps that only manifest in the final phase of migration and force costly rework, and so-called cutover risk — the fear of outage of critical operations when switching from the old system to the new one. These factors create an expensive "double-bubble" period when the insurer simultaneously pays for legacy system operation and funds the modernization program.
However, McKinsey's interesting finding lies in where the most time is actually lost. Rewriting code or target platform configuration represents only a fraction of the work. A disproportionately large share of energy and investment falls on understanding and setting rules, data conversion, quality control, reconciliation, operational readiness and stabilization after migration. This is exactly where agentic AI, according to analysts, comes onto the scene.
What is agentic AI and how it differs from ordinary copilots
Agentic AI are not just smarter chatbots. While developer copilots like GitHub Copilot or Claude Code assist the user in real time, agentic systems are designed to independently pursue a goal, break it down into partial tasks, use available tools and context, and iteratively adapt to feedback and control mechanisms.
In the context of insurance policy management system migration, this means that the biggest bottlenecks are usually not writing new code, but discovery, mapping, testing, reconciliation and switching loops. Agents can read outdated programming languages that almost no one understands anymore, reverse engineer the logic and convert it into comprehensible English. McKinsey states that in many cases an agent can handle in a few days what would take a trained expert months or years.
Once agentic capabilities are introduced, the costs of modernizing other products and systems can quickly drop. The same agents, patterns and contextual layers can be reused across migration waves and different areas. McKinsey describes this as "a portfolio option that insurance technology leaders have not had until now".
Where the biggest gains will show: from 20% to 90%
The report divides the modernization process into several phases and assigns them specific productivity estimates:
- Discovery and reverse engineering of legacy systems: 20–50% productivity improvement
- Testing, reconciliation and compression of defect cycles: 15–90% productivity improvement
The highest numbers therefore come in phases that were previously the most manual and least attractive — repeated testing of data conversion, comparing outputs of the old and new system, finding bugs and fixing them again. Agents shorten these cycles through automation and parallel running of hundreds of scenarios.
Three fundamental shifts that McKinsey recommends
Analysts identified three critical changes without which insurers will not reap the full potential of agentic AI.
1. Modular agents instead of monolithic solutions
Insurers should build agents as a library of atomic capabilities with clear inputs, acceptance criteria and escalation paths to humans. This approach improves control, enables output auditability and allows reuse across discovery, data, testing and cutover phases.
2. From one big migration to a portfolio of opportunities
With a reusable agent stack, the incremental effort needed to modernize additional products or neighboring applications decreases. Leaders can thus evaluate platform migrations concurrently with selective rewrites of the long tail of legacy applications.
3. Reworking roles, governance and risk management
Because insurance system modernization is subject to regulation and is operationally sensitive, agentic workflows must contain built-in controls: human approval at key points (human-in-the-loop), traceability from requirements through configuration to testing evidence, and clear model validation procedures.
