No context. Generic models don’t know your forms, your appetite rules, your endorsements or your adjusters’ shorthand — the 90% that changes carrier by carrier. You close the gap with ever-longer prompts and retrieval, and pay for it on every call.

Their control. The provider sets the price, the roadmap and the deprecation date. A model you validated last quarter can change under you, and your only options are to revalidate or fall behind.

Compounding cost. Every document read and every claim summarized is metered, forever. Volume that proves the business case is the same volume that makes the bill grow without end.

Data gravity and residency. Some books, some jurisdictions and some reinsurance arrangements simply don’t allow the data to leave.

Your advantage leaks. The judgment that makes your underwriting distinctive goes into someone else’s context window and comes back as a generic answer. None of it accrues to you.

A sovereign model isn’t a file you buy. It’s the output of a repeatable process run by people who know insurance, on tooling you keep. SovereignWorks brings all three.

People Insurance SMEs, instruction engineers and annotators working alongside your underwriters and adjusters. Your experts set the rubric; our team applies it at volume. This is the part that can’t be automated away — expert judgment is the raw material.

Process A five-stage factory with SME review before anything reaches the weights:

Continued pretraining — your corpus: policies, endorsements, manuals, claim files, bulletins.

Instruction tuning — your tasks, phrased the way your people actually ask.

Tool training — teaching the model to use your systems, not just talk about them.

Evaluation — measured against a rubric your experts own, before anything ships.

Preference tuning — aligning to how your best underwriters and adjusters decide.

Tools A four-layer stack: an open-weight base, our insurance model, your company model, your agents. Plus the pipeline, adapters and eval suite you keep and can rerun without us.

Most enterprise agents sit at L0–L1. EvalWorks moves you up the curve, and keep

Each layer is replaceable. Nothing in the design requires our continued involvement or a particular provider’s continued goodwill.

SovereignWorks is right when at least one of these is true:

Volume is high and permanent. Metered inference on a core process compounds into a line item you can never retire.

The domain is yours. Your appetite, your forms, your shorthand — and a generic model keeps getting the last mile wrong.

Residency or control is required. Regulation, reinsurance terms or group policy say the data stays put.

The capability is strategic. You don’t want the thing that differentiates your underwriting sitting in a shared service.

It’s the wrong answer for exploratory use cases, low volume, or anywhere a hosted model already performs well enough. We’ll say so.

Sovereignty assessment — 2–3 weeks. Corpus review, use-case economics, a rent-versus-own cost model over three years, and a recommendation that may well be “keep renting, for now.”

Factory build — the five stages run on your corpus with your SMEs, ending in a validated model in your tenancy with its eval suite.

Run and evolve — retraining cadence, drift monitoring, preference tuning as your book changes, and enablement so your team can operate the factory.