Life and health underwriting class assignment

The useful output is a life or health class recommendation with three citations, not a policy that has already been issued or declined. Machine learning may propose preferred, standard, rated, or decline. That proposal stays a suggestion until an underwriter assigns the class.

Each recommendation must cite the medical file that was read, the guideline vintage that was applied, and the application answers that were treated as true. If any of those cites cannot be produced, the class field stays empty. Empty is a valid state. A filled class without a vintage, without a named medical document, or without the application is not a recommendation worth putting on the file.

Do not auto-issue from a preferred score. Do not auto-decline from a decline score. Do not invent a rating class because the file looks close to a table in last year's manual.

Policy administration and underwriting workbenches from vendors such as Guidewire and Duck Creek, and guideline or reinsurance material associated with firms such as Milliman and Munich Re, already store applications, medical packets, and rule books. Use those systems as the source of the three artifacts. Do not treat the model as a second underwriter that can bind or refuse coverage on its own.

Class assignment sits next to underwriter decision support briefing. The briefing explains the file. The class rec proposes a bucket. Neither one issues the policy.

Load three artifacts into the same decision context before any class is recommended.

The application is the first artifact: product, amount, tobacco, build, disclosed conditions, medications, family history, and the exact wording of the questions the applicant answered. If a required question is blank or the application is not signed, stop. Do not complete the form from similar cases.

The medical file is the second: attending physician statements, labs, prescription history, industry database hits your shop already uses, and any paramedical exam. A cite names the document and the date. Writing "medical file" without an APS date, a lab panel, or a page reference is not a cite.

The guideline vintage is the third: the underwriting manual or reinsurance guide identifier and its effective date, pinned to this product and age or amount band. Vintage is not the phrase "current guidelines." It is a version you can retrieve when someone asks why this case was preferred and a similar case last month was standard.

Run the work in this order. Confirm the case is open in the workbench with a stable application ID. Attach or retrieve the medical packet and refuse to score if the packet pointer is empty. Pin the guideline vintage for this product band. Only then run the class model against that locked triple.

If application answers and external medical data disagree, do not pick a class to cover the conflict. Send the file through application vs external data reconciliation first. A class assigned on an unreconciled contradiction is a class assigned on a guess.

A class with no guideline vintage is a failure mode, not a minor data-quality issue. Two reviewers will not be able to tell which build chart or impairment table the model used. Leave the class empty until the vintage is pinned.

The recommendation payload should look like work an auditor can follow.

Illustrative example, not a measured result: a 47-year-old applicant for term life discloses treated hypertension, lists a beta blocker, and denies tobacco. The medical file includes an internist APS from March of this year, office blood pressures, and a lipid panel. The shop's life underwriting guide, vintage 2026.04, maps controlled hypertension on a single agent, with readings in the guide's stated range, to standard non-tobacco rather than preferred. The model may recommend standard, citing the application tobacco and medication answers, the March APS with those pressures and the panel, and guide 2026.04, hypertension table, controlled-on-one-agent row.

That rec is complete. The underwriter can agree, apply a rate class, or order more requirements. The system does not mail a policy and does not mail a decline.

If the APS is missing and only a pharmacy fill list exists, you do not have a medical-file cite for the condition the guidelines actually test. Do not recommend preferred because the fills look orderly. Do not invent Table B because the case feels like a mild extra. Leave the class blank.

Rated recommendations need the same three cites. Point to the guideline cell that produces that table or flat extra, the medical finding that lands in that cell, and the application answers that do not contradict it. If the guide has no cell for the finding, the model does not mint a new rating class. It leaves the class empty and surfaces the unmatched finding.

Decline uses the same shape. A decline rec cites the guideline's decline criteria and the evidence that meets them. It remains a recommendation. Auto-decline from the model is out of scope. An underwriter, or the shop's documented decline authority, assigns decline.

Blank is more honest than a guessed preferred.

Leave the class field empty when the guideline vintage is unknown, expired for this product, or not loaded for this age or amount band. Leave it empty when the medical file pointer is empty, or when the cited page is not in the packet. Leave it empty when the application is unsigned, a relevant question is blank, or a required disclosure is missing. Leave it empty when reconciliation has flagged an unresolved conflict. Leave it empty when the finding has no matching guideline cell, because any class in that situation would be invented.

Desks will want a default. Resist it. Defaulting to standard to keep the queue moving teaches the desk that missing evidence is acceptable. Defaulting to decline to be conservative is auto-decline with extra steps.

What you can still show on a blank class is which cite failed, which documents arrived, and which vintage was requested. That is work-queue information. It is not a class.

An underwriter assigns preferred, standard, a specific rated treatment, or decline, using the shop's authority limits. That assignment is the operational act. The model does not write the class onto the policy record as issued.

Treating the recommendation as issued is a failure mode that appears later as a contract that does not match the file: preferred language in the policy, standard evidence in the APS, no human sign-off. Catch that before anyone calls the case bound.

If the underwriter changes the class, they record why, especially when they override a cited rec. Keep the original recommendation and its cites on the file. Do not silently rewrite the rec to match the assignment. That destroys the trail.

Do not auto-issue when the rec is preferred. Do not auto-decline when the rec is decline. Preferred still needs the underwriter, or a documented straight-through rule that is not this model. Decline still needs authority. Rated still needs a human to confirm the table or extra. Rounding or inventing a rating class is how you get a policy you cannot defend.

Once the underwriter has assigned a class, the case can move to remaining requirements and, if approved for issue, to document production. Class is an input to automated policy document assembly, not a substitute for it. The assembled contract must print the assigned class, not the model's earlier guess.

Then run issuance accuracy verification: issued class, ratings, exclusions, and amounts match the underwriter assignment and the cited file. A mismatch here is an issuance defect.

Keep the class model inside issue-stage underwriting. It does not replace briefing, it does not skip reconciliation, and it does not bind coverage.