Conversational quote intake

A quote conversation is an interview that fills a risk form. Chat or voice asks the applicant questions, then writes answers into named fields. Quality means a populated field cites the applicant answer that supports it. If they never said it, the field stays empty.

The conversation is not a quote. It does not price, bind, or confirm cover. Rating and referral sit after the record is complete enough for a person to stand behind it. The common failure is treating a fluent chat as if it were already a quote. A transcript can read as finished while occupancy, construction, claims, or a required coverage question is still missing. That file is incomplete. It must not enter auto-quote.

This pattern fits personal lines and SME when the applicant is the household or the owner, answering in their own words. Broker-led commercial packs usually start from documents, not from a live interview. Those belong on a submission extraction and normalization path.

Downstream systems care about fields, not about the chat UI. Policy admin, agency management, and CRM stacks (Guidewire, Duck Creek, Salesforce, Applied Epic among them) should receive the same structured objects your quote desk already uses. Map occupancy to occupancy. Do not dump the transcript into a notes blob and call intake done.

Run the loop as ask, store, cite, stop.

Ask one risk question at a time, in words the applicant already uses. "What year was the building constructed?" is enough. Voice and chat can share the same script. Branch only on answers you have already stored with a cite. If they said the premises are a shop, ask shop questions. If they said a home, ask dwelling questions. Do not open a fifty-field questionnaire in one bubble.

Store the value in the field it belongs to, together with a cite. A cite is a pointer to the turn or transcript span (a message id in chat, a time range in voice). Anyone who opens the file should see year built = 1987 and be able to open the sentence where the applicant said 1987.

Do not store a tidy paraphrase as if it were the answer. If they said "late eighties, I think eighty-seven," you may parse 1987 only when your rules allow that parse, and the cite still points at that phrase. If they said "old building," leave the structured year empty. Keep the raw line in a notes field that rating must not read.

The failure mode is a field with no answer cite. If year built is filled and the cite is missing, treat the value as untrusted. Clear it or send it to the intake owner. Models fill schemas easily. Without a cite you cannot tell applicant speech from model invention. Make cite presence a hard validation rule before the record can leave intake.

A skipped question is not a no, and it is not a yes. If they closed the chat before the alarm question, alarm type is empty. If they said they would check and come back, alarm type is empty. Empty is a stored state, not a hole to paper over.

Do not invent a risk fact to make the form look complete. Writing "standard alarm" when nobody said it is invention. So is assuming masonry from a city postcode, assuming no prior losses because they did not volunteer any, and assuming a van is social-only because they described a weekday commute. Guesses do not belong on applicant-stated fields.

Third-party attributes can still be useful. Keep them on a separate overlay, labeled as enriched, never as applicant-stated. external data enrichment at quote can add building characteristics or claims history. Enrichment does not fill a skipped interview question. If the product requires an applicant answer, a person still has to ask.

A café owner starts a property conversation on their phone. They confirm the trading name, the unit address, that they lease the ground floor, and that cooking happens on site. The assistant asks about the fire suppression system. They say they need to ask the landlord and hang up. The record should show cooking on site as cited, leasehold as cited, and fire suppression as blank. The intake lead receives an incomplete file. The next action is a follow-up to the owner or the landlord. Writing "wet chemical hood, assumed" is inventing a risk fact.

Do not auto-quote from a partial conversation. Quote engines will price whatever schema you send. If blanks were coerced to defaults, you get a number that looks official and is wrong. Incomplete files never enter rating automatically.

An intake person owns those files. Ownership is a queue: conversations that ended with required fields empty, cites that failed validation, and answers that conflict (1987 in one turn, 1979 in another). A person resolves them. The model does not close the file by guessing.

Define required fields per product before you go live. Home, personal auto, and a simple SME package will not share the same required set. The conversation may end when every required field has a cite, or it may end early when the applicant leaves. Early end always creates a human task.

You can still run a cheap filter on what you do have. appetite and complexity triage can look at class of business, geography, and a crude size signal, then say keep collecting or stop and hand off as a specialist risk. Triage must not write missing construction or missing claims into the risk record. It is not a quote.

When two cited answers disagree, keep both cites and flag the field. Do not pick the friendlier number. If they correct themselves later, store the correction with a new cite and mark the old value superseded. Silent overwrite destroys the reason you captured speech in the first place.

Once required fields are cited and remaining empties are honest empties, the record can move to rating, referral, or a request for documents. ml-augmented indicative rating should read the structured, cited fields, not the chat log. A rating model that consumes free text will re-introduce invention.

Map into the objects the quote desk already trusts. Whether the destination is Guidewire, Duck Creek, Salesforce, Applied Epic, or a rating worksheet, the contract is field-level: applicant-stated with a cite, blank, or enriched from another source. If a value looks surprising, the desk should open the cite. If they cannot open a cite, the field should not have been populated.

Voice needs a playable cite: audio or a time-aligned transcript, not only a model summary of the call. Chat needs a stable message id. Summaries are for the human who is catching up. They are not the system of record.

Before you scale, enforce four checks: every populated applicant field has a cite; empty required fields create a human task, not a quote; no job writes a risk fact without an applicant span; quote and bind refuse records flagged incomplete. Fail any of those and you have a chatbot sitting in front of a questionnaire, not conversational intake.