Skip to main content
DoneThat

AI Adoption GuideInsuranceQuote

ML-augmented indicative rating

Machine learning overlay generates risk-adjusted price recommendations with expected loss and margin estimates per submission.

Insurance processQuoteUnderwriteBindIssueBillServiceRenewClaim

By Don, DoneThat’s AI coach · updated

Cite the table vintage, the overlay vintage, and the fields you used

An indicative price is usable only when it names the rating table vintage, the overlay model vintage, and the submission fields the overlay used. If any of those cites is missing, leave the dependent recommendation empty. Do not backfill a vintage from memory, from the current table, or from last week's batch. The underwriter still issues the quote.

Pricing and underwriting leads see this fail as a dollar amount that appears in the quote workspace with no provenance. Filed or scheduled tables change. Overlay models retrain. A recommendation you cannot reproduce is not a rating output. It is an uncited number, and you treat it as missing.

The table vintage identifies which table produced the starting premium. The overlay vintage identifies which trained adjustment produced the risk-adjusted recommendation, including any expected-loss or margin estimate the overlay is allowed to emit. The used-field list identifies what the overlay actually consumed, so a reviewer can see whether occupancy, construction, protection, TIV, a claims flag, or an enrichment attribute entered the score.

A price with no table vintage fails even if the overlay vintage and field list look complete. You do not stamp the vintage after the fact because the policy-admin system usually points at the current table. Withhold the indicative premium until the table cite exists.

Run this only after appetite and complexity triage. An overlay on a submission that should have been declined or referred before rating wastes a model run and pollutes the file with a recommendation that should never have been priced.

Load the table and the overlay, then draft with cites

Load both engines before you draft. Rating and policy-admin platforms such as Guidewire and Duck Creek commonly hold the table the underwriter is accountable to. Pricing and overlay systems such as Earnix commonly hold the ML adjustment. Data and analytics vendors such as Verisk commonly sit upstream as enrichment. Treat them as a class of sources. Do not rank them, and do not assume a named vendor supplies a vintage you have not loaded.

If location, hazard, or similar attributes are still being attached, finish external data enrichment at quote first. An overlay that scores a half-complete submission will cite fields that were empty or stale. That cite cannot be defended.

Draft in this order:

  1. Confirm the rating table identifier and vintage are in the payload.
  2. Confirm the overlay model identifier and vintage are in the payload.
  3. Record the submission fields the overlay used. If a listed field was null, keep it on the used-field list as null rather than dropping it so the file looks clean.
  4. Write the table premium, the overlay recommendation, and any expected-loss or margin figures the overlay actually produced, each attached to those cites.

If the table does not load, stop. There is no indicative price. If the overlay does not load, you may surface the table premium with its table vintage only. You do not attach a risk-adjusted premium, expected loss, or margin that pretends the overlay ran.

The draft is a recommendation object inside underwriting. It is not a quote letter, not a binder, and not a number the broker can bind against.

Leave expected loss and margin blank when a cite is missing

Expected loss and margin are overlay outputs. They inherit the overlay vintage and the used-field list. If either cite is missing, those cells stay empty.

Do not invent a loss ratio. Do not invent an expected-loss percent so the worksheet looks finished. A fabricated ratio that sits in the same grid as the table premium becomes the model number in later review. Underwriters can work from a table premium and their own judgment. They cannot reliably unsee a made-up loss ratio.

If the overlay returns a premium recommendation but omits expected loss, leave expected loss blank. Do not back into a ratio from premium and a guessed pure premium. If the overlay returns expected loss without a model vintage, discard the expected loss. An uncited number is not an estimate. Apply the same rule to margin: a margin figure without the overlay vintage is not a margin estimate.

One walkthrough, with no invented results. A mid-market warehouse submission reaches quote after appetite screening. Construction, occupancy, protection class, TIV, and a prior-claims flag are on the submission. Enrichment has already attached a location attribute. The rating table loads and reports its vintage. The overlay loads, reports its vintage, and lists those fields as used. The draft may show a table premium, an overlay-adjusted recommendation, and whatever expected-loss and margin figures the overlay actually emitted. Each row carries the three cites.

Change one condition: the overlay service times out. The table vintage is present. The overlay vintage is not. The draft shows the table premium only. Expected loss is blank. Margin is blank. The overlay-adjusted premium is blank. Nobody types a familiar occupancy loss ratio into the empty cells so the file looks complete.

Change a different condition: the table service returns a premium amount but no vintage identifier. Withhold the entire indicative price. That is a price with no table vintage. The underwriter can still quote after a reload or from manual rating. They must not send the uncited amount as an indicative rate.

The underwriter quotes; nothing auto-binds

The overlay does not bind. The draft does not bind. Auto-bind on an ML-augmented indicative rate is not part of this control.

Treating the indicative as bound is the failure that follows a clean-looking draft. A broker who receives an overlay recommendation as if it were a firm quote will work that number. If the underwriter later changes it, you have both a market problem and an authority problem. Keep the recommendation inside the underwriting workspace until an underwriter issues the quote.

The underwriter may accept the overlay recommendation, change premium, or ignore the overlay and quote the table. Each of those is a quote. None of them is the draft.

Quote letters, specimen wording, and portal labels need the same line. If a document can be mistaken for a bindable quote, it is the underwriter's document. It does not carry the overlay as the quoting party.

If the draft sits while the table or the overlay model moves, do not let the old recommendation travel with a new vintage implied. Apply stale-quote repricing lock so an expired indicative cannot be issued without reloading current vintages.

Renewal pricing is a different stage. A similar overlay at renewal belongs with dynamic renewal pricing, not with this quote-time draft.

Hand the file to underwriting with blanks intact

The file the underwriter opens should show cites that exist and blanks that do not. A referral note can say the overlay did not run, or that the table vintage was missing and the indicative was withheld. It should not contain a substitute loss ratio, a guessed margin, or a vintage the system did not return.

Check the handoff:

  • Table vintage present, or the indicative premium is absent.
  • Overlay vintage present, or overlay premium, expected loss, and margin are absent.
  • Used-field list present whenever overlay outputs are present.
  • No auto-bind flag, no portal bind, and no broker-facing number until the underwriter quotes.
  • No filled expected-loss or margin cells that lack the overlay vintage.

If those checks pass, the underwriter has a recommendation they can use, challenge, or set aside. If they fail, the underwriter still quotes, but they quote without pretending the overlay produced a complete rate.

Is this worth automating for you?

Whether this pays back depends on how much time it takes your team today. Most teams estimate that from memory, and the estimate is usually wrong in one direction or the other. This one is rated high effort to implement, so the baseline matters more than usual.

DoneThat reconstructs where the time actually went, with no timers to forget, so you can measure the baseline before committing to a project and check the gain afterward.

Measure the baseline first