Cluster parent

Reinsurance AI: operations software that extracts, traces, and packs every field

Reinsurance AI is software for the B2B chain between brokers, cedents, and reinsurers. It turns inboxes of PDFs and spreadsheets into a structured pack: named insured, limits, attachments, TIV, gaps, and a chase list. Every field should show the span it came from. Data stays in the tenant.

What reinsurance AI is not

Reinsurance AI is not a strategy memo about transforming insurance. It is not a public chatbot you paste a treaty into on a laptop at a conference. It does not replace a catastrophe model, a capital model, or the underwriter who decides whether to write. If a vendor answers a missing deductible with a confident number, that is a product failure, not a feature.

It is also not a retail claims bot, not a call-centre script, and not a dashboard that invents TIV so the chart looks complete. The work is B2B and document-heavy. Files arrive as PDFs, spreadsheets, forwarded threads, and scans. People re-key. Quotes wait. Bordereaux miss the treaty year. Wordings are argued after the binder because nobody can point at the hours clause.

If you need a definition of the market itself, start with What is reinsurance? and the answer graph. This page is about the operations loop: inbox to pack, pack to bordereaux, wording, then claims. Software that cannot show a source span for a field is a summariser. Treat it that way when you buy.

The buying test is whether the product can name the job it is doing on Tuesday afternoon. Inbox of slips to a pack with a chase list is one job. Outbound premium bordereaux against a signed wording is another. Inbound claims rows against a layer is a third. A demo that hops between all three without saying which mailbox it connected is a tour. Operators do not buy tours. They buy a first file they can inspect, revoke, and argue with.

The operations map

The chain is five jobs. Each job has its own documents, its own failure modes, and its own SLA. Mixing them in one undifferentiated inbox is how a facultative limit gets posted to a treaty year, or a claims bordereaux row gets treated as premium.

Inbox to pack

A placing broker or an inbound reinsurer mailbox is a pile of attachments. The job is not to summarise the email. The job is to turn those attachments into a structured pack: named insured, period, limits, deductibles, attachments, TIV, schedule highlights, and a list of what is still missing. The Submission Agent is the named worker for that ingest. Humans still bind. The pack is the working file the market can price against, not a pitch deck.

Walk a fictional file before you send a live one. The public pack inspector uses a sample only. Production ingest is tenant-scoped and read-only.

A real submissions mailbox is not one PDF. It is a zip, a forwarded thread with three versions of the same SOV, a scan of a slip with a handwritten share, and a loss run saved as an image. The pack has to survive that mess. If the product only works when you pre-name the files Slip.pdf and SOV.xlsx, you will still be sorting mail. The connector should be read-only, scoped to that mailbox or folder, and revocable. Writing as the broker is a different grant. Do not let it hide inside the ingest pitch.

Pack to bordereaux

Once something is written, reporting starts. Cedents send premium, claims, and commission schedules. Reinsurers have to post them against the wording they actually signed. Bordereaux are not a vibes report. They are rows that must map to treaty year, currency, class, and attachment. The long form of that job is the bordereaux automation guide. For a short definition, see What are bordereaux in reinsurance?.

Wording

Quota share, surplus, and excess of loss are standing capacity. The ops work is keeping the contract, the bordereaux, and the technical inputs on the same definitions. Treaty reinsurance is automatic for a class. Facultative reinsurance is optional on both sides and lives or dies on the pack. Glossary entries sit at treaty reinsurance and facultative reinsurance. If the hours clause is only in a scanned PDF, extraction that cannot point at the page is not extraction.

Claims

Claims in this product are loss runs, claims bordereaux, and recoveries against layers — not a consumer FNOL chatbot. IBNR language, as-at dates, and follow-the-settlements arguments all depend on whether the row you posted matches the wording. A claims total with no source span is an argument waiting for month-end.

Capital and accumulation sit on top of clean extracted exposures. Clash and peak-zone views that invent TIV are worse than no view. Do not buy a heat map that cannot open the cell it came from.

If two facultative packs both contain ACME Construction Ltd and nobody recorded that the TIV on one file conflicted with the schedule, your peak-zone total is a story. Accumulation is a downstream consumer of the pack, not a reason to skip provenance. Get the span first. Map the zone second. The same rule applies when a treaty bordereaux and a facultative certificate both mention the same insured: two contracts, two files, two postings. Merging them because the name matched is how recoveries get argued twice.

Why generic LLMs fail on slips and wordings

A London slip and an SOV often disagree on TIV. An hours clause lives in a scanned PDF. A reinstatement is a formula, not a slogan. Language models complete text. They do not, by default, refuse to emit a field without a citation. That is the design split, not a prompt-engineering footnote.

Paste a facultative pack for ACME Construction Ltd (acme.example) into a public model and ask for the occurrence limit. You will often get a fluent sentence. You may even get USD 10,000,000 if that number appeared somewhere. You will not reliably get document, page, and character span. You will not reliably get a chase item when the hours clause is absent. You may get a blended TIV when the slip says 42 million and the SOV totals 47.1 million because a warehouse was added after the slip was typed. Averaging those figures is not underwriting. It is hiding a conflict.

Generic models are trained to be helpful. Helpful, in a treaty, is dangerous. A missing deductible should be empty. A conflict should stay a conflict. Source-grounded extraction is the rule: no span, no value. If the vendor cannot show that loop on a sample pack you control, you are buying a summariser with a reinsurance theme.

Public endpoints also move the file. A treaty wording is not a blog post. If the path is copy, paste, third-party, you have already left the tenant. That is a procurement fact, not a vibe about innovation.

Reinstatement language is a useful torture test. The wording may say one reinstatement at additional premium, calculated as pro rata of the layer, with a 72-hour hours clause that you still have not found in the pack. A generic model will often emit a tidy sentence that sounds like a reinstatement clause. An operations product should extract the formula only where the span exists, and leave the hours clause empty until a page appears. If the vendor cannot fail that test on a file you bring, the rest of the pitch is decoration.

MRC layouts make this worse. Limits live in a grid. Conditions live in a paragraph. The SOV is a different workbook. A model that reads left to right across the email will mix a covering-letter TIV with a schedule total. The pack must keep those as two fields or a labelled conflict. Helpfulness that merges them is the bug.

What shows its work means

Shows its work is not a slogan on a hero image. It is a data shape.

Field, then document, then page or sheet, then character span or cell. A gap is an empty span: the hours clause is not in the pack, so the field stays empty and becomes a chase item. A conflict is two spans that cannot both be true: slip TIV versus SOV TIV for ACME Construction Ltd. A chase list is the ordered queue of gaps and conflicts. Nothing in that loop requires you to believe a percentage on a landing page.

Operators should be able to click a limit and land on the sentence it was copied from. If the product shows a confidence bar and no span, you still cannot defend the number in a query. If it shows a span on the wrong page, that is a defect you can file, not a mystery. Provenance is how you mystery-shop. It is also how an auditor, a claims adjuster, or a following market reconstructs what you thought you extracted.

Do not accept a narrative summary as a substitute. Summaries drop exclusions. Binders argue about what was said. Citation is the product. Pair any model, including a domain model, with that loop or you are back to unsourced paragraphs.

The chase list is a first-class output of that shape, not an appendix. If the hours clause has no span, the item is on the list with a reason: CAT underwriters will assume a window. If the SOV has no as-at date, the item is on the list because values may be last year. If two spans disagree, the item is a conflict, not a gap. Calling everything a missing document is how technicians stop reading the list. The product should regenerate the list when a new attachment lands, keep spans that still hold, and reopen items the new file did not close.

Three buyers, three inboxes

The same extraction loop serves three inboxes. The documents overlap. The SLAs do not.

Brokers live in submissions. The mailbox is slips, SOVs, loss runs, and broker correspondence. The output is a pack the market can quote, plus a chase list for missing pages, stale schedules, and unsigned slips. Speed without a complete file is a faster wrong quote.

Cedents live in outbound reporting. The job is premium, claims, and commission bordereaux that map to the wording they actually wrote — attachment, hours clause, territory, excluded classes, currency, treaty year. Reinsurers reject or query files that do not map. Late files are not just an ops nuisance. They change who owns IBNR and reinstatement.

Reinsurers live in inbound packs, inbound bordereaux, and accumulation. They triage what brokers send, post what cedents send, and try not to double-count TIV across facilities. MGAs sit on both sides of reporting: they look like cedents to the reinsurer and like a carrier to the binder.

Do not buy one chatbot and hope it learns all three inboxes. Name the mailbox, name the output, and name who still binds. If a demo cannot say whether it is placing, reporting, or inbound triage, it is a tour, not an implementation.

SLAs differ even when the extraction rule does not. A facultative pack that sits overnight can miss a market. A bordereaux file that sits for six weeks can move who owns IBNR. Inbound reinsurer triage that treats every zip as urgent will bury the complete file under the incomplete one. Implementation starts by picking the inbox whose missed SLA hurts this quarter, then proving spans on that inbox, then expanding. Three logos on a slide is not a rollout order.

Reinsure-8B versus a public LLM

Reinsure-8B is a domain model you can run in a tenant. A public LLM is a third-party endpoint. Neither is a system of record. Pair the model with provenance or you are back to unsourced summaries.

A public model will talk fluently about quota share and CAT XOL. Fluency is not posting. It will not, by default, keep your slip inside your perimeter. It will not emit a chase list unless you build that product around it. It will complete a missing hours clause because completing text is what it does.

A domain model can be better at market language: reinstatement, clean-cut, follow the settlements, hours clause, slip versus schedule. Better language is still not a pack. If Reinsure-8B (or any model) fills a limit without a span, refuse the fill. The operations product is the loop: ingest, extract with spans, gap, conflict, chase, human decision. The model is a component. If a vendor sells you the component and calls it the loop, you will still re-key.

Use the public inspector and a tenant pilot to see the difference. Do not use a consumer chat window as the control. The control is: given this pack, which fields have spans, which are gaps, and which conflicts are still visible.

How implementation is supposed to look

Implementation is not a rip-and-replace of the treaty admin system. It is a read-only mailbox or a tenant drop zone, scoped to the inbox that actually receives submissions or bordereaux, revocable when you are done looking.

Connect Microsoft 365 or Gmail to one mailbox or folder. The connector copies. It does not send as you. Writing mail as the broker or the cedent is a different trust boundary and should be sold separately, if at all. You should be able to see what was read and revoke the grant. Files stay in the tenant. This website does not take live client packs. The pack inspector is a sample so you can see fields, conflicts, and gaps without uploading a real slip.

Then you pick a job, not a platform tour. Brokers start with inbound submissions and the pack. Cedents start with outbound bordereaux against one treaty wording. Reinsurers start with inbound triage or inbound bordereaux. Humans review evidence and bind. Agents do not bind.

If a vendor needs a six-month data lake before you can see a span on a limit, you are not implementing operations software. You are funding a science project. An honest first week looks like: one mailbox, one sample of your own in the tenant, one pack or one bordereaux file, a gap list you can walk, and a person who still signs the quote or the settlement.

Week two should still be recognisable to the technician who owns the inbox. Brokers: the next live submission, same chase-list shape as the sample. Cedents: one treaty, one period, premium file first, claims file second, wording used as the validate-against target. Reinsurers: inbound packs or inbound bordereaux, not both on day one. If the vendor is already talking about replacing the admin system, you have left the implementation that this product is supposed to be. Layer on top. Keep the grant narrow. Keep bind with the human.

What a complete pack contains

A pack is complete when an underwriter can price without guessing. If a field has no source page, it is not in the pack yet. It is on the chase list.

Named insured, as written on the slip, with a span. Period. Interest and territory. Occurrence and aggregate limits. Deductibles and attachments. TIV, with the slip figure and the SOV figure both visible when they disagree. Schedule highlights: locations, occupancies, values that actually drive the quote. Loss history with an as-at date. Stale runs belong on the chase list, not in a chart that pretends they are current. Wording excerpts that govern: hours clause, exclusions, warranties — not a marketing brochure. Source spans on every populated field. Gaps. Conflicts. A chase list a placing team can work.

Fictional walkthrough: ACME Construction Ltd, acme.example, property facultative. Slip.pdf page 1 names the insured and the period 1 Jan 2026 to 31 Dec 2026. Slip.pdf page 2 states USD 10,000,000 any one occurrence and TIV 42 million. SOV.xlsx totals 47.1 million because a warehouse was added after the slip was typed. Hours clause: not found. SOV as-at date: not found. The pack records the limit with a span, records the TIV conflict with both spans, and puts hours clause and as-at date on the chase list. It does not average the TIVs. It does not invent 72 hours.

That is the file you send to market, or the file you refuse to send until the chase items close. Completeness is not a page count. Completeness is whether the next person can act without asking you what you meant.

A treaty placing file is a different completeness test. You still want spans. You do not want a single-location SOV pretending to be the board. The following market needs structure, attachment, exclusions, bordereaux samples, and loss experience for the class. Mixing a facultative pack and a treaty submission in one folder is how an occurrence limit gets discussed as a layer. Name the pack type on the first screen. The public inspector is a facultative sample on purpose. Do not treat it as a treaty demo.

Loss history belongs in the pack only with an as-at date. A five-year run with no as-at is a chase item. Plotting it as current is how development surprises the quote. The same is true of a schedule with no valuation date. Stale is a status. It is not a reason to invent a date.

Failure modes you can mystery-shop

Do not buy on a slide. Buy on a pack you can break.

Hand the product a slip with no deductible and ask for the deductible. If it returns a number, fail the demo. Hand it ACME Construction Ltd with slip TIV 42 million and SOV TIV 47.1 million. If it emits one TIV, fail the demo. Remove the hours clause. If it writes 72 hours because that is common, fail the demo. Give it an unsigned slip. If the pack looks bindable, fail the demo. Give it a loss run with no as-at date. If the product charts the losses as current, fail the demo.

On bordereaux, feed a file whose treaty year does not match the wording. If it posts anyway, fail the demo. Change a column header from Policy No. to Pol Ref. If extraction dies or silently maps the wrong column, you are looking at brittle IDP, not ops software. Put two currencies on one schedule without an FX as-at. If a single total appears with no gap, fail the demo.

On wording, ask where the attachment lives. If the answer is a paragraph with no page, fail the demo. Ask the model to split quota share from excess of loss in a hybrid contract. If it cannot point at the clauses, you will fight that split at settlement.

Mystery-shopping is allowed to be rude. The failure mode is always the same: a value with no span, a conflict flattened into one number, or a file that left the tenant. If the vendor will not run those tests on a sample you bring, you already know the answer.

Score the session in writing. For each field you care about, record: populated with a span, gap, conflict, or invented. Invented is an automatic fail. Then ask where the file lived during the demo. If the answer is a vendor laptop or a public paste box, that is a second fail even if the spans looked pretty. Then ask who binds. If the answer is the agent, stop. This website will only ever show a sample. Your live packs stay in the tenant. That constraint is part of the product, not a footnote under a signup form.

Longer guide (existing) · Answer graph · Submission Agent