The Bordereaux Problem: Data Quality Someone Else Controls

There's a data problem in insurance that's uniquely frustrating because the data isn't yours to fix. It arrives from someone else, in their format, at their quality, on their schedule — and your reserving, reinsurance recoveries, and portfolio view all depend on it.

It's the bordereaux problem, and anyone doing delegated authority, MGA, coverholder, or reinsurance business lives with it.

What a bordereaux actually is

When you delegate underwriting authority — to an MGA, a coverholder, a broker — they write business on your paper and report it back to you periodically as a bordereau: a spreadsheet listing the risks written (premium bordereaux) and the claims incurred (claims bordereaux).

That spreadsheet is your only window into a book of business you're carrying the risk on. Everything downstream — reserving, aggregate exposure, reinsurance recovery, regulatory reporting — is built on it.

And it is, almost universally, a mess.

Why it's such a mess

Every counterparty has their own format. Different columns, different orders, different names for the same field. You receive twenty bordereaux from twenty coverholders in twenty layouts, monthly, forever.

The formats change without warning. Someone adds a column, renames a header, or changes a date format. Your ingestion breaks silently — or worse, doesn't break and quietly misreads a column.

Coding is inconsistent. Peril codes, occupancy classifications, cause-of-loss — each counterparty uses their own vocabulary, and mapping it to yours is a judgment call someone makes in a spreadsheet.

Data arrives late and restated. Last month's figures get revised this month. Which version is the truth, and does your reserving reflect the restatement?

You can't enforce quality. This is the crux. You can ask for better data. You can put standards in the contract. But you can't fix the coverholder's system, and if they're producing profitable business, "improve your bordereaux or we terminate" is a threat nobody wants to make.

What it costs

Aggregate exposure is a guess. You cannot see your true concentration if location data arrives inconsistently across counterparties. You'll discover the accumulation after the event, not before.

Reinsurance recoveries leak. If claims data is incomplete or miscoded, claims that should be recoverable under a treaty aren't identified. That's money you're entitled to and don't collect.

Reserving is delayed and imprecise. Late, restated data means your view of the book lags reality — which is a solvency issue, not an inconvenience.

People do the ingestion by hand. Skilled staff spend their month reformatting spreadsheets. It's expensive, error-prone, and impossible to scale as you add coverholders.

The approach that actually works

You can't control the input, so build a pipeline that assumes chaos at the boundary.

1. Ingest per-counterparty, normalise to one canonical schema. Accept each bordereau in whatever shape it arrives; maintain a mapping per counterparty; transform into a single internal model. The mapping is the asset — treat it as versioned code, not tribal knowledge in someone's head.

2. Validate at the door, and reject loudly. Row counts, control totals, date ranges, required fields, referential checks. A file that fails validation should raise an exception, not flow through. Silent acceptance of a broken file is how a misread column becomes a misstated reserve.

3. Detect schema drift automatically. Compare each file's structure to the last accepted one from that counterparty. A new or renamed column should alert someone before the load, not after a quarter of wrong numbers.

4. Score data quality per counterparty. This is the highest-leverage move. Track completeness, validity, timeliness, and restatement frequency by coverholder, and report it to them. You suddenly have an objective, unarguable conversation — and, in renewal negotiations, real leverage. Some will improve simply because it's now visible.

5. Version everything. Restatements are normal. Store each submission immutably, so you can reproduce what you believed at any past date. Reserving and audit both need this, and it can't be reconstructed later.

6. Enrich what they didn't send. Missing geocodes, construction characteristics, and classifications can often be derived or looked up from what you did receive. Do it on your side rather than waiting for data that will never come.

The strategic point

Delegated business is growing across the market, which means the share of your book you see only through someone else's spreadsheet is growing too. Insurers treating bordereaux ingestion as clerical work are accumulating a blind spot proportional to their growth.

The ones treating it as a real data-engineering problem — canonical schema, validation, drift detection, quality scoring per counterparty — get something valuable: an accurate, timely view of a book they don't directly control, and the evidence to improve it at renewal.

You can't make your counterparties send clean data. You can build a pipeline that tells you precisely how dirty it is, fixes what's fixable, and refuses what isn't. In delegated business, that's the difference between managing a portfolio and hoping about one.

We build ingestion, validation, and quality-scoring pipelines for delegated and third-party data. More at IntelliBooks.