Policy RAG coverage assistant

A usable answer names the issued form or endorsement, a locator a reviewer can open, and the language that maps to the question. If that cite cannot be produced, return nothing on the coverage point. A fluent sentence without a form page is not an answer. It is a record you will have to unwind.

Broker desks and policy-service leads get one class of question all day: does this policy cover this, what is the limit, what is the deductible, which endorsement applies. Speed of the first paragraph is not the quality bar. The bar is whether someone else can open the same PDF or form edition and find the same sentence.

Policy administration and agency systems (Guidewire, Duck Creek, Salesforce, Applied Epic) hold the policy of record, the customer thread, and the servicing task. They do not, by themselves, prove which edition was issued or which manuscript endorsement sits on the file. Retrieval has to run against the documents that were bound, not against a summary field typed at last renewal.

Do not treat assistant output as a coverage confirmation. Confirmation is a human act. The machine puts the right page next to the question, or it stays silent.

Start with the inbound question as it arrived: email, chat, call notes, or a broker portal ticket. Keep the customer's wording. Capture only the identifiers you need to fetch the right file: named insured, policy number, term, and location or vehicle when the question is specific to one. Do not rewrite the question into a legal issue you have already decided.

Load that packet against the issued file for that term: declarations, base form, schedule of endorsements, and endorsement text. Mid-term pages are in scope only when they are on the file. Changing the contract is a different job. Hand that off to mid-term endorsement processing agent. This page is retrieval and citation, not issuance.

Pull identifiers from the system of record so you do not retrieve a sibling account or a prior term. Guidewire or Duck Creek can supply policy number, term, and form schedule when those fields are trustworthy. Salesforce or Applied Epic can supply the broker's question text and the servicing queue. Neither class of system is the citation. The citation is the form page.

If the question is oral, write it down before retrieval. An ai voice agent for policy servicing can take the call and capture the question. It must not answer coverage on the phone from memory of the conversation.

Search the bound packet for the coverage part, exclusion, condition, limit, deductible, or endorsement the question names. Prefer the specific endorsement over brochure language. Prefer the page that was issued over a specimen that merely looks similar.

When a match exists, return three things together: the excerpt, the form or endorsement identifier, and a locator (page, section, or clause). Then stop. Do not add that the insured is covered, or that this should be fine. Those are opinions.

One path through the desk, not a measured result: a commercial broker writes about a businessowners policy. A contractor flooded a tenant suite after a sewer backup. The broker asks whether water backup is covered, what limit applies, and whether a deductible other than the property deductible sits on that coverage. The assistant loads policy number and term, retrieves the issued form and the endorsement schedule, and looks for water-backup language. If an issued endorsement states a sublimit and a separate deductible, it quotes those lines and points at that page. If the schedule has no such endorsement and the base form excludes backup of sewers and drains, it quotes the exclusion and the form page. If the endorsement is listed on the declarations and the PDF is missing from the archive, it does not guess last year's specimen and does not invent a sublimit because that is what the desk usually binds. It returns empty on the limit and flags the missing document for a human.

That is the how-to in motion: load the question, retrieve the form, cite or stay silent, leave the opinion to the person with authority.

Limits and deductibles are easy to fill from habit. Habit is not a source. If the declarations show a limit and the form explains how that limit applies, cite both. If only one of those pages is in the packet, cite what you have and leave the other field blank. Inventing a limit is a hard fail, even when the invented number is the one the desk expected.

An answer with no form cite is a failure mode, not a draft. Suppress it. Do not soften the empty state with typical wording, or with what most policies like this say. Typical wording is how people invent coverage.

Stay silent on the coverage point when the policy file cannot be identified, the form edition on the declarations does not match the PDF in the archive, an endorsement is scheduled but the text is not attached, the question depends on facts the file does not contain (date of loss relative to a mid-term change, occupancy, who was operating the vehicle), or retrieval returns a near-match from a different state or year.

Empty does not close the ticket. Empty means the assistant does not speak as if it read the page. The desk then fetches the missing form, calls the carrier, or tells the broker the file is incomplete. That is slower than a confident paragraph. It is the quality outcome.

A second failure mode sits next to the missing cite: treating the chat as a coverage confirmation. Once a sentence appears in the broker portal or a Salesforce thread, people screenshot it. If that sentence confirmed coverage, you created a record you did not intend to create. Allowed outputs are excerpt-plus-cite, or a structured "no cite available" state. Do not offer a yes/no coverage switch.

Complaints often start when someone was told they were covered and the claim file later disagrees. Routing those complaints is a different workflow: complaint classification and deadline routing. Prevention on this desk is refusing to emit an uncitable answer.

After a cite is on screen, a licensed person (or the authority your firm assigns) decides what to tell the broker or policyholder. That person reads the excerpt in context: other exclusions, conditions, who is an insured, the facts of the loss or the risk. The assistant does not reconcile conflicts between forms. It does not pick which endorsement probably wins. If two pages appear to conflict, surface both cites and stop.

Gray-area calls stay human: whether a fact pattern meets a definition, whether an exclusion's exception applies, whether a manuscript endorsement changes the base form in the way the broker hopes. Retrieval exists so that person is looking at issued language instead of a summary.

Do not auto-confirm coverage in email templates, chat macros, or voice wrap-up. A cite in the workspace is input to the person who will write the reply. The reply can attach the same page. The reply must not say the assistant confirmed anything.

Renewal conversations will try to reuse this retrieval for shopping. Finding that an endorsement is absent is not a bind order and not a cross-sell. Talking about a gap at renewal is a separate motion: coverage gap cross-sell at renewal. Keep this assistant scoped to questions about the in-force file.

Keep vendors in their lane. Policy admin (Guidewire, Duck Creek) can supply identifiers and, where you have built it, a pointer to the form archive. Agency and CRM platforms (Applied Epic, Salesforce) can hold the question, the task, and the human's sent reply. Index the issued documents, versioned by policy and term. If a field in admin says water backup is present and the PDF does not contain the endorsement, trust the PDF or stay silent. Do not let a checkbox outrank a missing page.

Run a simple sample on the queue. For each question, did the assistant return a form or endorsement cite a reviewer can open? If yes, did it refrain from confirming coverage and from filling a blank limit? If no cite, did it stay empty on the coverage point? Anything else is noise.