Appetite and complexity triage

The classifier scores an inbound submission against appetite and complexity, then proposes a route: toward an auto-quote path, or toward a human underwriter. That is the job.

A route is a recommendation about who works the file next. It is not a bindable quote, not a decline, and not an appetite score you can paste into a rating worksheet. If you treat the route as a decision, you have already left the operating rule.

Intake has to be trustworthy before this step. If the ACORD, loss runs, and schedule have not been extracted and normalized, you are scoring a guess. Start from submission extraction and normalization. Conversational intake does not change the rule: chat can collect missing fields, but it cannot invent appetite. See conversational quote intake.

What the classifier may output:

• A route toward an auto-quote path, only when both cites are present and the complexity file places the risk in the low-complexity band that path is built for.
• A route toward a human underwriter, when appetite is in-band but complexity is not, or when the file needs judgment the path is not allowed to apply.
• An empty route, when the appetite rule cannot be cited, the complexity file cannot be cited, or both.

What it may not output: a quote, a decline, a numeric appetite score that did not come from the loaded rule set, or a filled route with a missing cite.

Score nothing until both sources are loaded for the same product, effective date, and branch.

Appetite is the written rule set: classes, geographies, limits, occupancy, construction, and the exclusions that take a risk out of the book. Complexity is a separate file: which combinations are simple enough for a quote path, which need a specialist, and which attributes (layered programs, unusual occupancy mix, incomplete loss history) force a person.

Load them as files with identifiers you can cite. A policy admin screen in Guidewire or Duck Creek, a submission record in Salesforce, or an agency download in Applied Epic is not a substitute for the rule identifier. Treat them as a class of systems of record for workflow and account state. They do not authorize you to skip the cite.

If appetite and complexity disagree, do not average them. Do not invent a blended score. Record the conflict and leave the route empty so an underwriter can see both sources.

An appetite memo from last renewal and a complexity file from this quarter are not a pair. If either file is stale, missing a product, or scoped to a different branch, treat the cite as missing.

A complete route names two things, in language a reviewer can follow without opening a model card:

1. The appetite rule that applied, including the rule identifier and the clause that matched, or the clause that put the risk outside appetite.
2. The complexity file entry that applied, including the file version and the band or flag that matched.

If either cite is missing, the route stays empty. Empty is a valid state, not a failed job. Downstream should send an empty route to a human queue, not retry the model until it invents a number.

Do not backfill a missing appetite cite from a similar account. Do not infer complexity from premium size, broker, or the claim that this looks like last week's habitational. Similarity is not a cite.

The auto-quote path is a destination, not a result. Routing toward straight-through processing for low-complexity risks still requires both cites. Straight-through processing may produce a quote later, under its own gates. This classifier does not.

ML-augmented indicative rating can inform a conversation. It does not replace appetite, and it does not fill a blank cite.

A mid-market package lands with a warehouse, a small retail storefront, and a tenant that does light assembly. Extraction has filled occupancy, construction, limits, and a five-year loss run. Appetite for the product includes wholesale and retail in the stated states, and excludes manufacturing above a defined process list. Complexity treats single-occupancy wholesale as in-band for the quote path, and mixed occupancy with any assembly as out-of-band.

The classifier should load those two files, match the occupancy mix to the appetite clause on manufacturing and assembly, and match the mixed-occupancy flag in the complexity file. If both matches are citable, the route should point to a human underwriter and show both cites. The assembly question is a judgment the quote path is not allowed to make.

If the appetite file has no clause for mixed occupancy, do not analogize from mostly warehouse. The appetite cite is missing. The route stays empty even if complexity is clear.

If complexity has no mixed-occupancy flag, do not treat two locations as a substitute. The complexity cite is missing. The route stays empty even if appetite looks open.

The underwriter who picks up the file should see the cites that exist, the fields that stayed blank, and nothing that looks like a quote number or a decline stamp. That is the shape of a correct run, including when the route is empty.

The person with authority still decides whether the account is quoted or declined. The classifier does not.

A route toward the auto-quote path means both cites were present, appetite did not exclude the risk, and complexity placed it in the band that path is built for. The path may still fail its own checks. Bindability, forms, and referral rules live there, not here.

A route toward human underwriting means the file needs a person. It does not mean lean yes or lean no. Do not encode a shadow decline by routing every awkward class to a graveyard queue that never quotes.

An empty route means do not pretend the model finished. Queue it with the missing cite named.

When a human takes the file, they should get the cites, the blanks, and the extracted facts, not a recommended premium. Pair this step with underwriter decision support briefing so the workup leads with evidence, not with a model mood.

Do not auto-quote from this classifier. Do not auto-decline from this classifier. Do not invent an appetite score to make a dashboard go green. If a reviewer cannot point at the rule and the complexity file, the output is not ready for any path that issues terms.

A route with no appetite cite. The complexity file matched, the queue shows a complete chip, and intake sends the file to the quote path. That is an incomplete route wearing a complete badge. Empty should have stayed empty.

Treating the route as a quote. Operations copies the destination (auto-quote) into a broker status as if terms exist. They do not. The classifier never produced a premium, form, or binder.

Inventing an appetite score. Someone wants a 0-100 so they can sort the queue. The score is not in the rule file, so a model supplies one that will be treated as appetite. If the rule set does not publish a score, you do not have one.

Other patterns that produce the same harm: scoring before extraction is complete; citing a screen in Guidewire, Duck Creek, Salesforce, or Applied Epic instead of the rule identifier; filling blanks from a lookalike account; using indicative rate as a proxy for appetite; declining by routing to a queue that never works files.

Can a reviewer open the appetite rule and the complexity file from the cites on the route? If yes, the route may stand. If either cite is missing, the route stays empty and an underwriter decides what happens next.