Policy Administration Migration: Converting 9,200 Exceptions
Cedar Valley Insurance is 11 months into a 14-month migration of 1.15 million personal and small commercial policies from a 1998 mainframe policy administration system to a cloud platform. Two dry runs converted 99.2% of records. The remaining 0.8%, about 9,200 policies, fail validation, mostly because coverage limits, deductibles, and endorsement forms were stored in free-text notes maintained by agents. Separately, the conversion team reports that 14% of converted homeowner records show a different protection class than the legacy system, because the new platform geocodes addresses instead of using the legacy rating territory table. The CFO has committed to a hard cutover on the first of the quarter to stop $220,000 per month in duplicate licensing costs. The CIO proposes converting the 9,200 exceptions as-is with a flag and letting agents clean them up at renewal. The personal lines underwriting manager is worried about renewal rating accuracy and about a claim arriving on a policy with the wrong limit.
1. Evaluate the CIO's proposal for the 9,200 exceptions. What would you do instead, and why?
Read the worked answer
Converting as-is transfers a carrier data-integrity problem to agents who have neither the records nor the incentive to fix it, and pushes discovery out to renewal dates that may be two or three years away. If a loss occurs on a policy whose converted limit understates what was sold, the carrier faces an errors-and-omissions and bad-faith exposure that dwarfs the licensing savings, and a records-accuracy finding in a market conduct exam.
A better approach is triage by materiality rather than a single as-is decision:
Reconstructable. Many exceptions can be rebuilt from declarations pages, agent correspondence, billing history, or by reverse-engineering premium. Run automated inference, then human QC on a statistically valid sample. Convert with the reconstructed value and a source citation in the record.
Not reconstructable. For this residual, do not convert a guessed limit. Keep a read-only legacy lookup available to claims and underwriting for at least 24 months, and run a targeted outreach campaign with a written confirmation deadline.
Controls. The exception flag must surface at point of quote, renewal, endorsement, and claim, not just in a data-quality report. Require claims adjusters to verify limit against the legacy image for any flagged policy. Set a hard exit criterion: zero policies in force without a system-of-record limit after 12 months, tracked by an accountable owner.
2. The protection class discrepancy affects about 14% of homeowner records. How do you resolve it, and what makes this more than a data-quality issue?
Read the worked answer
Geocoding is not automatically more accurate than the legacy table, and this is a rating question, not merely a data question. Rating territories and protection classes are filed and approved by state; the legacy territory table may be the filed artifact. If the new platform silently applies a different territory, the carrier may be charging a rate that does not match its filed program, which is a rating error and potentially an illegal rate.
Method: (1) Obtain the currently filed territory and protection class definitions for each state and confirm which methodology is on file. (2) Sample 200 records and classify each variance as legacy data error, geocoder error (P.O. box, rural route, centroid fallback, new construction), or methodology difference. (3) Quantify premium direction and magnitude by state. (4) Correct only where the new value is defensible against the filed definitions; hold the legacy value where it is not. (5) Where a methodology change is intended, file it, disclose the impact, and handle any customer premium change under state notice rules, ideally at renewal with advance notice.
Why it matters beyond data quality: territory is a rating factor with anti-redlining sensitivity. A geocoder that systematically misplaces rural or low-income addresses shifts premium by geography and can produce a disparate-impact allegation even where no intent exists. Track premium change by geography before go-live.
3. Recommend a cutover decision to the CFO, including the conditions you would attach.
Read the worked answer
Recommend a conditional deferral or a phased cutover, not the hard date as proposed. The CFO's cost is real but it is bounded and known: roughly $220,000 per month. The downside of a botched conversion is unbounded, because it includes claims-handling failures on policies with wrong limits and rating errors across 14% of the homeowner book.
Conditions I would attach:
Scope the cutover. Move a line or state at a time, beginning with the cleanest population. Small commercial with simple forms before homeowners with complex endorsements.
Resolve territory first. No homeowner conversions until the filed-methodology question is settled and premium impact by geography is quantified.
Parallel run. Thirty days of dual processing with daily delta reporting on premium, limits, forms, and producer commission.
Rollback plan tested. Demonstrated ability to revert within 24 hours, with the legacy system retained in read-only for the full run-off period.
Exit criteria in writing: exception rate below an agreed tolerance with reconstructed values; zero unexplained territory variances; claims lookup tested end-to-end on flagged policies; reconciliation signed off by underwriting and finance, not only IT.
Governance. Present the residual risk to the board or risk committee as a documented acceptance with a named owner. Frame the trade to the CFO honestly: a four-to-six week deferral costs about $300,000 in licensing and buys the validation that protects the book.