Skip to main content
Insights · Definitions · Insurance

The bordereaux data quality problem — and what MGA data infrastructure looks like from the source.

What is bordereaux data quality? Bordereaux data quality is the accuracy, completeness, timeliness, and defensibility of the periodic reports an MGA sends to its capacity carrier. When bordereaux fields don't tie back to a verifiable source, carriers reject or manually re-key them, delegated-authority audit cycles balloon, and reconciliation becomes the operating bottleneck of the MGA relationship. This piece is about what closes that gap at the source instead of managing it downstream.

Why bordereaux quality became the live conversation in 2026

Three pressures are converging on managing general agents right now, and every one of them ends at the same operating question: can this bordereaux be defensibly reconciled?

  • Carriers are tightening delegated-authority controls in response to NAIC AI Model Bulletin expectations, Lloyd's Coverholder audit posture, and the general regulatory tilt toward continuous-audit rather than periodic-audit oversight. The bordereaux is the daylight into what the MGA actually bound. If the fields don't tie back to a verifiable source, the carrier can't demonstrate the delegated authority was exercised inside the bound guidelines.
  • MGA tech stacks have historically been fragmented. Policy admin, rating engine, reporting layer, downstream claims — each hand-off is a reconciliation boundary. Bordereaux extract runs at the boundary, and every gap in the source data becomes a gap in the report. The stack was never designed to answer a per-policy audit question; it was designed to move policies through a workflow.
  • AI-assisted underwriting introduces new inputs. Verified-asset scores, external data feeds, model-derived features. These don't map cleanly to a legacy bordereaux schema whose field list was designed for hand-keyed operator entries. The MGA either invents ad-hoc columns (which the carrier's ingestion won't accept) or drops the AI-derived signal from the report (which the auditor then can't reconstruct).

What "closing the gap at the source" actually means

The gap isn't a reporting problem. It's a shared-record problem. The MGA prices from data. The carrier binds against that data. The auditor reconstructs from that data. When each of those parties is reading a different snapshot of the underlying record, the reconciliation gap is baked in — bordereaux is the periodic best-effort attempt to close it.

A verified-asset record shared by MGA and carrier changes the shape of the problem. The MGA reads the same signed per-asset record the carrier can also read. When the MGA quotes off that record, binds off that record, and reports off that record, the bordereaux fields tie back to a verifiable source by construction — the record itself is the tie-back. The carrier can query the underlying record from any bordereaux row and reconstruct exactly why the MGA bound at the rate it did.

The record shape is not novel; it is exactly the primitive Callisto Bridge built for Keystone's verified-asset rail in industrialized construction:

  • Signed at insert — ECDSA over canonical serialization; every field the MGA read to price is the field the carrier reads to audit.
  • Hash-chained lifecycle — every state change to the record appends a hash including the previous entry's hash. Any tampering, reordering, or field re-write is detectable by structure.
  • Deterministic scoring where it fits — the KeyScore is the same math for every party. There is no private MGA model vs public carrier model; both parties read the same number, computed the same way, off the same record.
  • Row-level tenant isolation — cross-MGA reads are impossible by database rule, not by policy. Each MGA's book is its own; carriers reading across their delegated MGAs read only their own capacity's records.

What this looks like operationally

The MGA runs its normal book. No new admin tool, no new integration to their existing PAS. What changes is that every bound policy carries a per-asset identifier plus the signed record that produced the rate. The MGA's own reporting, their carrier feed, and their reinsurance ceding statement all reference the same record ID.

The carrier's ingestion doesn't rewrite fields. Bordereaux stops being a periodic file transfer that requires the carrier's team to re-key half the values. It becomes a query against a shared source of truth. The carrier's audit team runs the same queries the MGA's compliance team runs; both see the same answers.

Delegated-authority audit cycles compress. The audit becomes a query, not a reconstruction. Weeks become hours. The auditor asks "on which of the 4,200 policies bound this quarter did the MGA exceed guidelines," runs a filter, reads the result. No inbox threads asking the MGA's data team to re-pull a report.

The claims cycle inherits the benefit. When a loss triggers, the parametric or condition-based reference reads the same record. If the coverage was written against an attribute the record captures, the trigger fires against that attribute at claim time. Trigger disputes close before they open.

Where this fits (and where it doesn't)

Where it fits: specialty MGA books whose underlying insured assets are already candidates for a per-unit signed record — industrialized construction, EV batteries, solar and wind components, refurbished industrial equipment, medical devices, and adjacent verified-asset classes. The MGA doesn't need to spin up its own registry; it reads the registry the manufacturer, lender, and program office already read.

Where it doesn't fit yet: personal-lines auto, small-commercial-property, and other high-volume, low-per-unit-value MGA books whose underwriting economics don't support per-asset record capture. For those books, the traditional bordereaux workflow with continuous data-quality tooling is the honest answer. The verified-asset-record pattern is not a universal MGA data-infrastructure replacement; it is the right shape for MGAs whose books are already asset-anchored.

The commercial pattern

Keystone's Insurance Rail is the pricing engine layer. Licensed MGA partners bind on carrier paper. Callisto Bridge doesn't itself hold binding authority, doesn't own carrier paper, and doesn't touch the regulated function. The MGA data infrastructure is the record + pricing math; the regulated binding stays with the licensed counterparty. Same licensed-partner posture Keystone applies everywhere else: infrastructure at Callisto Bridge, regulated activity at the licensed partner.

What to do with this if you're an MGA

  1. Ask your capacity carrier what percentage of your bordereaux rows they re-key or reject. If the number is above 5%, you have a real reconciliation gap the shared-record pattern would close.
  2. Identify which of your bound risks correspond to already-registered verified assets (Keystone-shape asset classes above). If a meaningful fraction do, the shortest path to closing the gap is reading the record instead of rebuilding it.
  3. Bring 25–100 anonymized policy records to Keystone's design-partner pilot. The pilot output is a working pricing-engine instance on your own book with the bordereaux fields tying back to the signed source, so you can measure the audit compression against your existing carrier relationship.
Sources + primary references

What this article draws on.

  1. NAIC AI Model Bulletin on Use of Artificial Intelligence by Insurersnaic.org
  2. Cognizant industry analyst content on MGA integration + delegated authority paincognizant.com
  3. Verisk / ISO advisory-loss-cost framework — comparable pricing infrastructureverisk.com
  4. Lloyd's Coverholder audit requirements — delegated authority frameworklloyds.com
  5. AAIS (American Association of Insurance Services) — advisory-loss-cost framework for specialty linesaaisonline.com
Next step

Explore where the research applies.

Every field note traces to something we build. Follow the thread into the specific engine or capability the research supports.