The Producer Your Scorecard Can't See

AI-Ready by Architecture, part 2

Premira Mutual is a composite. The carrier is fictional; the problems are real.

Marisol Vega knows exactly how much business her agency wrote for Premira Mutual this year. Premira's scorecard doesn't, and neither does the AI model its underwriters are piloting.

Vega is the principal of Vega & Lake Insurance Group, one of the several hundred independent agencies that produce Premira's book across 12 states.

PolicyCenter, Premira's policy system, has two agency records for Vega & Lake. The original, which its renewals are issued under, has the agency's National Producer Number (NPN) on file. The duplicate, set up under a misspelled name when the agency got a new producer code for submitted business, doesn't.

Premira's submission intake system has a third record, under a third spelling. Three records describe one agency, and no Premira system knows they belong together.

Resolve the agency in the data model, and Premira's scorecard and contingency payments count all of Vega & Lake's production while the AI model ranks the agency's submissions on its full binding history. Data Vault, the architecture approach behind this series, does it with two structures: a Hub keyed on the NPN, and a Same-As Link that ties the duplicate and the intake record to that key and keeps the evidence for each match. Because the fix lives in the data layer, any score the AI model logs with its as-of date, model version, and feature code version traces back to the source data as it stood that day.

Every report Dale Pemberton, VP of Underwriting, pulls on Vega & Lake comes up a third short.

The agency's renewals flow clean under the original record's producer code. Everything that runs through submission (new business, rewrites, remarketed accounts) is issued under the duplicate's code. The scorecard reads a crosswalk spreadsheet that maps producer codes to agencies.

One senior analyst, Walt Pruitt, maintains the crosswalk by hand, and it has never mapped the duplicate's code to Vega & Lake. Billing pays the agency correctly on its tax ID, so its commission statements are right while the scorecard never sees the premium.

On paper Vega & Lake is an average agency, and Premira treats it like one: the standard service queue and no call when appetite opens.

That AI model, a submission-triage model Dale's underwriters are piloting, makes the same judgment for the same reason. It ranks incoming submissions by appetite and by each agency's history of binding. Vega & Lake's submissions arrive under intake's spelling, and the policies they bind sit under the duplicate record, so the model never connects the two and puts the agency's submissions in the middle of its queue.

Meanwhile the agency's contingency comes in light, because it's computed on the same production totals the scorecard reads. Vega notices and starts placing business where it gets counted.

Dale learned the size of the gap on the call he made about pulling back the agency's appetite. Vega read him its production from her own agency system, to the dollar.

He reran the numbers three times that afternoon, from three different reports. All three agreed with each other and disagreed with Vega.

Resolving the agency is an architecture problem

The usual response is a cleanup project: merge the duplicate record, correct the spelling in intake, add the missing code to the crosswalk, and move on.

Each of those fixes holds only until the next file lands. Correct the spelling in intake on Monday, and Thursday's submissions arrive under the old one, because intake records the name each submission gives.

Merging the two PolicyCenter records, or adding the duplicate's code to Walt's crosswalk, fixes Vega & Lake alone. The next time someone sets up a code in a hurry, a new duplicate appears.

A cleanup script picks a string-match threshold and keeps no record of its reasoning, so when it guesses wrong, two agencies become one and the merge can't be defended six months later.

Where Data Vault earns its keep

Data Vault resolves Vega & Lake once, in the data model, and the missing third of its production comes back. When the scorecard and the contingency read the vault, a policy under the duplicate's code rolls up through the duplicate record, and the Same-As Link carries it to the agency's NPN, so its premium counts with the original record's.

Data Vault anchors each business key in a Hub. The agency Hub holds the NPN from the original record (licensed agencies receive one, as individual producers do) and the keys PolicyCenter and intake assigned to the other two records. Each source's details land in its own Satellite, which timestamps every change, so every file loads without waiting for someone to decide which version is right.

The Same-As Link ties the duplicate and the intake record to the NPN, and a Satellite on the link stores the evidence for each match, with the rule that fired and the date. The duplicate matches on the tax ID it shares with the original. The intake record matches on the agency name and address on the license record for that NPN, and on the submissions that bound under the duplicate's code.

Because the link persists, Thursday's submissions and next month's policies under the duplicate's code arrive on keys already tied to the agency. When someone sets up another code in a hurry, the new duplicate enters the Hub under its own key, and the same match rules tie it to the NPN.

The link holds computed results, so it sits in the Business Vault, and the records it ties stay in the Raw Vault as delivered. Six months later, anyone can read the evidence behind a match and defend it. Reversing a wrong one adds a closing row to the link's Satellite, so a triage score from January still traces to the keys as they were tied together in January.

A rollup on the tax ID would catch the two PolicyCenter records and restore the scorecard's missing third. Intake carries no tax ID, so the triage model still couldn't connect the agency's submissions to its binds. The Same-As Link resolves all three records once and keeps the NPN as the key, so the scorecard and the model read the same agency.

When Walt leaves, Premira loses the only person who knows how to keep his crosswalk current. With the agency Hub and the Same-As Link in place, he reviews only the matches the rules can't make, and each answer he gives is stored as evidence.

What your AI model inherits

The scorecard judged Vega & Lake once; the triage model judges it again on every submission. With the vault in place, the model joins submissions to bound policies through the agency Hub and the Same-As Link, so it reads the agency's full binding history.

A language model can sometimes tie the three spellings together on its own. Researchers at the University of Mannheim found that GPT-4, with no training examples, beat fine-tuned matching models on three of six benchmark datasets, and that the best prompt depended on the model and the dataset.

A match made at query time can change with the prompt, and its explanation disappears with the session. Run the language model as one of the match rules instead, and the Same-As Link's Satellite stores its decision, the prompt version, and the date, so every model reads the same result.

The first article in this series sets five criteria a data platform must meet before an AI agent can safely use it. One is lineage: tracing any prediction back to the data behind it, as of the day it was made. This design answers it once each score is logged with its as-of date, model version, and feature code version, because the Same-As Link carries the trace through the match to the source records.

The vault also lays the foundation for two more: stable access paths and classification tags that belong to the dataset itself. The Hub's keys never move when a new duplicate appears, so the route an agent takes into agency data can be documented once. The agency's tax ID and commission bank details sit in a Satellite of their own, tagged as sensitive, so an agent asked what Vega & Lake wrote gets the premium while the tag's masking policies hide those details from its role.

Where to start Monday

Revenue gets attributed through producer codes, so before you build anything, measure how often one agency sits under several names. Pull your 20 largest commission payees from the system that pays commissions, then look each one up in every other system that holds agency records and count how many match without a person stepping in. The premium under the rest is your dollar figure for this problem.

The build comes next: the agency Hub on the NPN, a Satellite for each source, and one Same-As Link with one match rule and a confidence score. Send every match that scores below a threshold you set to whoever maintains your producer-code crosswalk today.

Then check it the way an AI model will use it. Ask a test agent how much your largest agency wrote with you last year, and compare the answer with the agency's own number. A mismatch means the agent is reading one of the records, and it won't say which.

After that, when Dale asks about Vega & Lake, he gets the number Vega read him on the phone. The next time appetite opens, Vega gets the call.

More in this series

Two more articles follow, each built on one part of the vault architecture. Next are the data contract and the Satellite that catch a vendor's schema change at the boundary and keep every change, in "One renamed column," and then the Business Vault and semantic layer, in "Teaching your data to speak insurance." The series opens with "Will your AI pilot stall at the data layer?"

About Premira Mutual

Premira Mutual is a regional P&C mutual founded in 1903 in Savannah, Georgia. The group writes $1.2 billion in premium across roughly 250,000 policies, about 90 percent of it inside 12 Southern states, about 60/40 commercial to personal.

Three brands carry the book. Premira Mutual is the flagship, sold through several hundred independent agencies. Premira Go is the direct and digital sub-brand, admitted across the 48 contiguous states plus DC. Live Oak Specialty writes coastal and excess-and-surplus property through MGAs and brokers. Every story in this series happens at the flagship.

Its data platform is where the stories happen: ten source systems, Guidewire on premise, billing on an AS400, reporting in Power BI, and a Snowflake and dbt migration that is real, funded, and unfinished.