Thinklytics

Regional P&C Insurer · Insurance · Columbus, OH · 13 weeks

Claims data unified across 4 legacy systems

A regional property and casualty insurer struggled with claims data scattered across four legacy systems, lacking a common policy ID. This prevented them from assembling training data, stalling their AI underwriting project for 18 months. We consolidated the claims data into a unified foundation in 13 weeks, enabling the AI initiative to launch six weeks after.

Challenge

The insurer bought two smaller carriers over eight years but never merged their claims systems. Each system had its own policy numbers, loss codes, and date formats. The data science team spent 18 months trying to build a usable training dataset and failed twice.

Approach

We created a claims data layer that matched policy identifiers from four different systems by normalizing policy numbers, matching insured names, and aligning effective dates. We standardized loss category codes to ISO standards and consolidated claims history into one table. This table now serves as the single source of truth for actuarial analysis and AI models.

Outcome

We consolidated 1.2 million policy records from four separate systems into a single claims data foundation. Six weeks after delivery, we launched the AI underwriting model, which went live within the next quarter. This work increased loss ratio reporting accuracy from 71 to 97 of every 100 by aligning data across systems.

Clearing up those annoying policy ID mix-ups after acquisitions

The client had three different policy numbering systems across four platforms, totally disconnected. We fixed this by building a deterministic matching process. We combined insured name, address, effective date, and coverage type into a single key. That key hooked all their data together and made sure everything matched up perfectly across the board.

How We Dug Into ISO Loss Categories to Tidy Up Our Messy Data

Here’s what happened: every system was speaking a different language when it came to loss codes. Pulling a combined loss ratio felt like herding cats. So, we rolled up our sleeves and mapped all those codes to the ISO standards. Then, we built a crosswalk table to connect the dots. That simple fix let us slice and dice historical data across all systems without any headaches.

Results

  • 1.2M Policy records unified across 4 systems
  • $8.4M AI underwriting initiative unblocked
  • 71 to 97 of 100 Loss ratio reporting accuracy
  • 6 weeks Time from delivery to AI initiative launch

Our data science team worked for 18 months on creating a training dataset for the AI underwriting model and ran into dead ends twice. Thinklytics pulled all our claims data together in 13 weeks, and we rolled out the model six weeks later. It’s up and running now.

Chief Actuary, Regional P&C Insurer

In the client's terms

The question

Where does this number come from, and can the path be trusted?

The situation

Two carriers acquired over eight years and four claims systems never merged, each with its own policy numbering, loss codes and date formats. The data science team had spent 18 months attempting a usable training dataset and failed twice.

What we found

Both failures had started from the model and worked backwards. The blocker was that a policy could not be followed across four systems, and that is a reconciliation job rather than a data science one.

What it produced

1.2 million policy records unified across four systems. The AI underwriting model launched six weeks after delivery and went live the following quarter, unblocking $8.4M of initiative. Loss ratio reporting accuracy rose from 71 in 100 to 97.

The roadmap

A claims data layer matched policy identity across the four systems by normalising policy numbers, matching insured names and aligning effective dates. Loss categories were standardised to ISO codes, and claims history was consolidated into one table designated as the source for both actuarial work and models.

The work behind this result

Related case studies

Frequently asked questions

How do you unify claims data across acquired carriers?

Build a composite key where no shared identifier exists. Four claims systems here used three policy numbering schemes, so insured name, address, effective date and coverage type were combined into one key, unifying 1.2 million policy records.

Why do insurance AI projects fail before the model is built?

Because a policy cannot be followed across systems. This insurer's data science team spent 18 months and failed twice, each time starting from the model and working backwards. The blocker was reconciliation rather than data science.

How do you make historical claims data comparable across systems?

Standardise the loss coding. Every system here used its own, which made a combined loss ratio impossible. Mapping all of them to ISO standards with a crosswalk took loss ratio reporting accuracy from 71 in 100 to 97.