Sandbox profile

Claim file summaries

Medium intensity

Adjusters get a summary of a first notice and the documents on file. The model does not decide coverage.

Vertical
Insurance
Scale
Small company
Organization
Cedar Lane MGA
Shape
Managing general agent, 40 people, personal lines, two states.

The business problem

A small MGA’s adjusters re-read the same FNOL, photos notes, and emails. A shared claims chatbot would put policyholder data in a vendor account the MGA cannot explain to its carrier or to its state. They want a one-page brief, written from the file, with the adjuster still deciding coverage and payment.

What Grae SI recommends

Family

Phi-4 mini or Qwen 2.5 3B Instruct

Size

3B-class. One GPU the MGA can colocate with its existing claims host.

Why this size

The brief is a summary, not a coverage opinion. A small model with a strict output shape (facts, missing documents, open questions) is easier to test than a larger model that wants to advise. The prompt forbids a coverage conclusion, and the eval checks that ban.

License

Printed on the card before training. The MGA accepts it. Grae SI does not sublicense the base model.

Data controls

Personal data

FNOL contains names, addresses, and sometimes health details after an injury. That is personal data, and injury facts can be health data.

De-identification

Training uses closed files with names and SSNs removed. Live briefs run on the real file inside the MGA network.

Access

The assigned adjuster and a supervisor. Producers do not query other people’s files through the model.

Residency

The two states are both US. Data stays in the US region the carrier already requires in the MGA agreement.

Retention

Briefs that are saved to the claim follow the claim-file schedule. Unsaved previews are not stored.

Risk classification

Intensity is a planning tier for evidence. It is not a certification and not a regulatory classification.

  • The brief can shape what the adjuster notices, so it is not a harmless search box.
  • Coverage, payment, and denial stay with the adjuster. The output template has no decision field.
  • Automatic denial, or a score that sets payment, would be high and is not this profile.

Compliance requirements to prepare for

NAIC AI model bulletin

The bulletin asks insurers and their producers to govern AI that touches consumers. An MGA should be ready to show intended use, testing, and a human decision. Grae SI prepares that evidence. It is not a filing.

State insurance AI rules

Follow the states where the paper is written. Colorado’s unfair-discrimination framework and New York’s insurance AI expectations are examples, not a complete list. Counsel maps the actual states.

SOC 2

Carriers often ask an MGA for a control report. Grae SI is not the auditor.

GLBA-style privacy

Policyholder financial and claim data needs the same access and retention care the claim system already has. Health details in an injury claim raise the bar.

Grae SI helps assemble the evidence. It does not certify HIPAA, FDA, ISO, SOC 2, ICH, NAIC, GDPR, DPDP, or any other framework.

Validation KPIs

These are targets to lock before go-live. They are not measured outcomes from a customer deployment.

Fact support

At least 95% of brief sentences supported by the file, on 80 closed claims

A clean sentence that is not in the file is a defect.

No coverage language

0 coverage grants or denials on a 40-prompt probe

The template and the probe both enforce the boundary.

Missing-doc recall

At least 90% of absent required documents flagged

This is the useful part for a small team.

SSN in training extract

Zero, on a 30-row review

Redaction has to be checked, not assumed.

Deployment architecture

DeploymentPrivate VPC next to the claims host

The claims system stays as it is. A private model endpoint reads a file the adjuster already opened. There is no broker portal hooked to the model in this sandbox.

  1. 01

    Claims system

    System of record for the file and the payment.

  2. 02

    Redaction for training

    Closed files only. Names and SSNs stripped.

  3. 03

    3B runtime

    Ollama on a small private node. Output is facts, gaps, and questions.

  4. 04

    Adjuster

    Reads the brief in the claim screen and decides in the existing fields.

Each box is a component in the customer’s network. Arrows are the order of the path, read left to right on a wide screen and top to bottom on a phone.

Where this sits in the stack

The same ten layers appear on every profile. The mark says whether Grae SI provides the layer, shares it, or leaves it with you. Person-weeks are the buyer’s effort after the runtime is deployed, not a project plan for the whole company.

  • Grae SI providesGrae SI provides the layer
  • SharedShared. Grae SI supports part of it.
  • You ownYou own the layer after install
  1. Layer 1

    Infrastructure and compute

    You own

    One small GPU beside the claims host the MGA already pays for.

    Buyer effort after the runtime is up: 1–2 person-weeks

  2. Layer 2

    Data ingestion and prep

    Shared

    A checklist of what to collect, a sample upload, and a readiness score. The pipeline from the system of record into an approved extract.

    Buyer effort after the runtime is up: 2–5 person-weeks

  3. Layer 3

    Data controls and governance

    Shared

    Policy numbers and SSNs are redacted in the prep preview. The carrier still owns the access model and the state privacy notices.

    Buyer effort after the runtime is up: 1–3 person-weeks

  4. Layer 4

    Model selection and training

    Grae SI provides

    Phi-4 mini or Qwen 2.5 3B. The job spec forbids a coverage field in the output template.

    Buyer effort after the runtime is up: 0.5–1.5 person-weeks

  5. Layer 5

    Evaluation and validation

    Shared

    A tier-matched suite, thresholds, and an exportable report. The labeled set, the acceptance call, and any statistics your quality system requires on top.

    Buyer effort after the runtime is up: 1–3 person-weeks

  6. Layer 6

    Agents and orchestration

    Shared

    Compliance agents that classify the use case and name the frameworks to prepare for. Any production agent that calls tools inside your applications.

    Buyer effort after the runtime is up: 1–2 person-weeks

  7. Layer 7

    Deployment and serving

    Grae SI provides

    Safetensors, GGUF, a model card, and an offline bundle for Ollama, vLLM, or llama.cpp. Moving that bundle into the zone and keeping the host patched.

    Buyer effort after the runtime is up: 0.5–1.5 person-weeks

  8. Layer 8

    Monitoring and drift

    Shared

    Registry lineage: version, dataset, eval result, and the package that was released. Live traffic checks, drift alarms, and the decision to pull a version.

    Buyer effort after the runtime is up: 1–3 person-weeks

  9. Layer 9

    Application and integration

    You own

    A panel inside the claim, not a new claims system.

    Buyer effort after the runtime is up: 2–4 person-weeks

  10. Layer 10

    Organizational validation and sign-off

    You own

    Compliance and the product counsel decide where a human must still decide. Grae SI will not mark the model approved for a state filing.

    Buyer effort after the runtime is up: 1–4 person-weeks

Buyer effort for this profile

4–8 person-weeks

A 40-person MGA cannot staff a model-risk department. The packet is short, and the carrier may still ask to see it.

  • Claims lead
  • Compliance (often the same person as the carrier contact)
  • The admin who runs the claims host

Quantitative benefits

Illustrative estimate for this sandbox. It is not a measured customer result. Assumptions for this profile are listed under the figures.

Time to a reviewable package

3–5 weeks

Two states, one line of business, one claims system.

Cost versus a cloud API

About $1.5k–$4k a month avoided

Illustrative: 6,000 claims touched a month, $0.25–$0.65 a summary on a hosted model, before the residency objection.

Compute

$150–$400 a month

One small GPU node, or a workstation in the same rack as the claims server.

Adjuster prep

10–20 minutes saved on a typical file

Illustrative, if the brief is actually used. A bad brief saves nothing.

Validation writing

About 1 person-week saved

Enough to answer a carrier’s AI questionnaire with attachments rather than a blank page.

Assumptions

  • Illustrative estimate for this sandbox. It is not a measured customer result.
  • The MGA does not set rates with this model and does not auto-deny.
  • State list is an example. The real compliance work is the states on the certificate of authority.
  • Injury claims with extensive medical records may need the health-system pattern instead of this 3B brief. That is a scope change.

All sandbox profiles · The full stack