// Carriers
Ship it inside the core you already run.
Guidewire on the cloud, a mid-tier policy system on some lines, and fifteen years of premium history, policy records and claim files on a system nobody wants to touch. We build across all three — and write the answer back where the underwriter will actually see it.
// The situation
Between six and ten percent.
Five independent 2026 studies converge on the same number: between 6% and 10% of insurers have scaled AI as an enterprise capability, against the 44–48% who report something “in production.”
Capgemini · BCG · Sedgwick · Datos · EXL, 2026.
Everything we do lives in that gap. The first four weeks tell you which side of it your own program is on, and what it would take to move.
// The stack you actually run
The AI problem is rarely the new system.
Nearly every carrier above roughly $500M direct written premium runs a hybrid: a modern policy system for new and renewal business on some lines, and a retained legacy stack holding runoff books, older lines and, critically, the historical claims and premium data your models need for training.
// The build
Four builds, and the surface each one runs on.
A semantic layer over Cloud Data Access
What lands in your bucket is a replica of the operational schema — wide, deeply normalized, effective-dated, built for transactions rather than analysis. You have the data. You do not have a model of it.
Surface: Cloud Data Access + EventBridge lifecycle events.
Write-back into PolicyCenter
A decision that survives the next endorsement, appears in the audit trail, and does not break the effective-dated model.
Surface: Cloud API, versioned. Requires your entity-model depth and your change-control calendar.
Submission intake with cross-document reconciliation
Not extraction — reconciliation. When the SOV total disagrees with ACORD TIV, the system says so and shows where, instead of silently picking one.
A pipeline that reads each document independently produces a confidently wrong answer.
Claims document structuring into ClaimCenter
Estimates, police reports, medical records and demand packages turned into a structured, searchable file with every extracted fact citing its source page — so the adjuster checks it in one click rather than re-reading 300 pages to trust it.
Design rule: we build the check-path first.
What we measure
Baseline before build, every time: touch count per submission before and after; time from submission receipt to first underwriter decision, on the same cohort definition both sides; rework rate, meaning how often a file goes back for missing data; and correction rate on AI output, tracked as a first-class metric rather than hidden.
These are definitions, not forecasts. If we cannot define the baseline in week one, that is a finding, and it goes in the report.
// We ran one of these
The builds on this page are the ones that recur across carrier operations, and the hard parts are named here rather than discovered in week six.
The operating record →// What's hard about this
Two things we tell you in week two.
Release-train coupling.
Guidewire Cloud upgrades on Guidewire’s cadence. Anything built against a non-sanctioned surface — direct database access, custom Gosu plugins, the old event-messaging pattern — breaks on their schedule, not yours. So we build on Application Events, Integration Gateway and Cloud API, or we tell you the thing you want is not safe to build yet.
Effective dating will make a naive extract wrong.
Commercial policies are amended constantly — endorsements, audits, mid-term cancellations, reinstatements. Any model touching premium or exposure has to understand your transaction and version semantics, or it will produce a confident number that nobody in finance recognizes. So the semantic layer models it before anything reads from it.
// What ships with it
The governance file ships with the model.
Roughly half the states have adopted the NAIC AI Model Bulletin, and the variations are material. Virginia replaced “mitigate the risk” with eliminate the risk. Connecticut requires an annual AI compliance certification attested by a named officer. Iowa formally defines bias and outcomes testing.
NAIC and state bulletins · current at September 2026.
Every build leaves with a model inventory entry, data lineage, pre-deployment testing results, drift thresholds with remediation triggers, and a named human decision-maker specification. If your deliverable does not include the governance artifact, it cannot deploy.
Bring us the workflow you’ve already tried to fix.
A decade of running an AI-native carrier, pointed at one of your workflows for four weeks.