Hand-authored synthetic sample

Forwardable sample clearance memo

This is a representative packet shape on fictional facts: source excerpts, obligations, evidence gaps, affected contracts, conditions, and the counsel decision record. It is not connected to live tenant data or any live product endpoint.

Request access
Illustrative sample — not legal advice. This packet uses fictional facts to show what Newcomb produces. It is not customer-specific advice or production output. See our terms.

Clearance packet

Underwriting fraud model

Launch · Ready for review · v1 · Owner: M. Chen, Compliance · As of May 28, 2026

Appears clearable with conditions

Newcomb reviewed an in-house model that scores claims and applications for likely fraud during underwriting. Across 18 sources it found 23 requirements that apply, surfaced 4 obligations that are easy to miss — including an AI-notice duty buried in 14 customer contracts — and ruled out 11 provisions that do not apply, recording the reason for each. The position appears clearable once four conditions are met. Counsel reviews, edits, and owns the decision.

Scope

PurposeFlag potentially fraudulent claims and applications for a human underwriter to review.
DataApplicant PII, Claims history, Third-party fraud signals
ModelNorthwind Risk AI · fraud-scout v3.2
CustomersCommercial lines, Personal auto
GeographyColorado, Illinois, New York
AutomationAdvisory — informs a human underwriter; no fully automated denial.

It does not cover adjacent uses.

4 things you'd likely miss

Sorted by severity. Every catch is sourced.

  • Criticalburied · cross-document · customer-specific · confidence: high

    An AI-use notice obligation is buried in 14 customer master service agreements.

    These accounts contracted for advance notice before AI touches their data. Scoring their claims without notice is a contractual breach, not just a regulatory gap.

    How to clear it: Send the AI-use notice to the 14 affected accounts before the model scores their claims.

    Likely missed because It lives in the customer MSAs, not in any AI statute a questionnaire or a law-only tool would check.

    Master Services Agreement §11.2 · affects 14 accounts

  • Highvendor-term · fact-triggered · confidence: high

    The model vendor's terms bar training on submitted data — and that duty flows to you.

    If any pipeline reuses submitted claims to tune the model, you are in breach of the vendor terms you accepted. The obligation has to be enforced, not just acknowledged.

    How to clear it: Confirm in writing with Northwind that submitted data is excluded from training, and retain the confirmation.

    Likely missed because Vendor terms are rarely re-read after signing, and the restriction is a single clause in §4.

    Northwind Risk AI Terms §4.1

  • Highnon-obvious · confidence: high

    Colorado treats this as a high-risk insurance use, triggering a pre-deployment impact assessment.

    The assessment must exist before first production scoring. Deploying first and documenting later does not satisfy the statute.

    How to clear it: File the pre-deployment impact assessment with Compliance before the model scores in production.

    Likely missed because Fraud scoring reads like an internal control, not a regulated consumer decision — but Colorado scopes it as high-risk.

    Colorado SB26-189 §6-1-1703(3)

  • Mediumcross-document · customer-specific · confidence: medium

    Three reinsurance treaties require disclosure of automated underwriting.

    Undisclosed automation can give a reinsurer grounds to contest coverage on affected books later.

    How to clear it: Add the automated-underwriting disclosure to the three reinsurer notices.

    Likely missed because The duty sits in treaty boilerplate, not in the model documentation.

    Reinsurance Treaty §7.4 · affects 3 treaties

23 requirements apply to this clearance

Grouped by source. Each requirement quotes the exact provision it relied on.

AI law

  • Complete a pre-deployment impact assessment for a high-risk insurance use.

    Verified

    Colorado SB26-189 §6-1-1703(3) — Fraud scoring in insurance underwriting is enumerated as high-risk.

    A deployer of a high-risk artificial intelligence system shall complete an impact assessment for the system before the system is deployed and at least annually thereafter.
  • Provide consumer notice when AI meaningfully informs an adverse decision.

    Needs confirmation

    Colorado SB26-189 §6-1-1703(4) — Applies only if a fraud score can by itself drive an adverse action — an open fact below.

    If a high-risk system makes, or is a substantial factor in making, a consequential decision concerning a consumer, the deployer shall notify the consumer.
  • Maintain a documented governance program for the AI system.

    Verified

    Illinois HB 3773 — The model serves Illinois personal-auto applicants.

    An entity that deploys an automated decision system affecting Illinois residents shall maintain a written program describing its development, testing, and oversight.

Sector rules

  • Conduct annual bias testing and retain the records.

    Verified

    NAIC Model Bulletin §4.3 — Adopted by your domiciliary state DOI.

    Insurers should test predictive models for unfair discrimination at least annually and retain documentation of the testing methodology and results.
  • Keep a model inventory entry with owner, purpose, and validation status.

    Verified

    NYDFS Circular Letter 7 — The model serves New York commercial lines.

    A regulated entity should maintain an inventory of external consumer data and information sources used in underwriting, including the basis for their use.

Customer contracts

  • Give written AI-use notice before the model processes the account’s data.

    Verified

    Master Services Agreement §11.2 — 14 accounts — The same AI-notice clause appears in 14 active MSAs.

    Provider shall give Customer at least thirty (30) days’ written notice before applying any automated or artificial-intelligence system to Customer Data.
  • Disclose automated underwriting to the reinsurer.

    Needs confirmation

    Reinsurance Treaty §7.4 — Three treaties carry the disclosure duty.

    The Cedant shall disclose to the Reinsurer any material change in underwriting methodology, including the introduction of automated or model-based decisioning.

Vendor terms

  • Do not allow submitted data to be used to train the vendor model.

    Verified

    Northwind Risk AI Terms §4.1 — The restriction binds your downstream pipelines, not just the vendor.

    Customer Submissions will not be used to train, retrain, or fine-tune the Services or any other model, and Customer shall not configure the Services to do so.

Company AI policy

  • Keep a human in the loop for any adverse action informed by the model.

    Verified

    Internal AI Policy §3.2 — Your policy is stricter than the statute here.

    No adverse action affecting a customer may be taken solely on the output of an automated system; a qualified employee must review and approve the decision.

Newcomb's packet for this system lists all 23 requirements; this sample shows representative ones.

Affected customer contracts

Accounts touched by at least one catch on this packet.

  • Cedar MutualMaster Services Agreement1 catch
  • Granite State InsuranceMaster Services Agreement1 catch
  • Meridian ReReinsurance Treaty 2025-A1 catch

+12 more accounts carry the same AI-notice clause. Every affected contract links to the exact clause and the catch it triggered.

Coverage manifest

What the analysis examined, and what it ruled out — with reasons.

18

Sources examined

23

Provisions cited

11

Considered & excluded

Sources examined

Representative — 18 examined in full.

  • Colorado SB26-189 · version 2026 session4 cited
  • Illinois HB 3773 · version as enacted2 cited
  • NAIC Model Bulletin · version Dec 20253 cited
  • NYDFS Circular Letter 7 · version 20242 cited
  • Customer MSAs · version as executed8 cited
  • Northwind Risk AI Terms · version v2026.12 cited

Considered and excluded

Representative — 11 considered and excluded in full.

SourceProvisionReason
Texas TRAIGA§552.001 et seq.The model is not deployed to Texas residents.
EU AI ActArt. 6 (high-risk)No EU data subjects are in scope.
NYC Local Law 144§20-870Applies to automated employment decisions; this is an underwriting tool.

Required controls & conditions

1 of 4 met. The clearance moves to fully cleared when all conditions are satisfied.

  • File the pre-deployment impact assessment with Compliance before first production scoring.

    Open

    Owner: Risk · Colorado SB26-189 §6-1-1703(3)

  • Send the AI-use notice to the 14 affected accounts.

    Open

    Owner: Account Management · Master Services Agreement §11.2

  • Obtain written confirmation that submitted data is excluded from Northwind training.

    Satisfied

    Owner: Vendor Management · Northwind Risk AI Terms §4.1

  • Add the automated-underwriting disclosure to the three reinsurer notices.

    Open

    Owner: Reinsurance · Reinsurance Treaty §7.4

What's still missing

Open facts the review needs before the position is final.

  • Can a fraud score by itself drive an adverse action, or does it only route to a human?

    Determines whether the Colorado §6-1-1703(4) consumer-notice duty applies.

  • Which additional states will the model serve at launch?

    New states can pull in additional sector rules and notice duties.

Proposed disposition — draft, your decision required

Appears clearable with conditions

Based on the reviewed sources and the facts provided, this is what the engine currently observes. Newcomb proposes; you decide. The recommendation is an observation, never a verdict.

Generated
May 28, 2026
Sources examined
18, at pinned versions
Proposed disposition
Appears clearable with conditions (draft)
Decision
Awaiting counsel review
Illustrative sample — not legal advice. This packet uses fictional facts to show what Newcomb produces. It is not customer-specific advice or production output. See our terms.

Want this for the AI systems landing on your desk?

Tell us about one, and we'll show you the real thing.

Request access