ClaimComb — loss-run normalizer for independent insurance agencies

1. One-liner

Turns every carrier’s messy loss-run PDF into one clean, year-over-year claims comparison ready for a remarketing submission.

2. Trend signal — why now?

Every commercial-lines renewal needs loss runs — the carrier’s report of a client’s past claims. The problem isn’t getting them (that’s a carrier-motivation problem nobody can fix with software). The problem is what lands in your inbox: a different PDF layout from every single carrier, sometimes 200+ pages, with ~30 fields per claim buried in inconsistent sections. The account manager then hand-keys all of it into a spreadsheet to compare year-over-year and to build a clean claims summary for the remarketing submission.

Insurance Quantified is blunt about it: “almost every carrier has a different format for their loss runs, [so] underwriters will oftentimes spend countless hours manually reviewing and re-keying loss data into their own spreadsheets.” FurtherAI literally titled an engineering post “The Hardest Document Extraction Problem in Insurance.” The re-keying is “tedious and exhausting… the kind of work that can really only be done in short bursts before fatigue sets in.”

Two things changed in the last 18 months. (1) Multimodal LLMs got good enough and cheap enough to reliably parse arbitrary, never-seen-before PDF layouts — the exact thing that made this an unsolved “hardest problem” two years ago. (2) Money is visibly moving: InsurTech funding topped $1B/month in early 2026, and the AMS ecosystem now has a connector tool (Loss Run Pro, $29/user/mo) — but it only collects and transports loss runs; it does not read or normalize them. The hard, valuable middle step is still manual.

Provenance:

• Signal 1 (demand): “Underwriters spend countless hours manually re-keying loss data… every carrier has a different format” — https://insurancequantified.com/resources/the-hidden-costs-of-manual-loss-run-processing/ — 2026-06-23
• Signal 2 (feasibility): “The Hardest Document Extraction Problem in Insurance” — ~30 fields/claim, 200+ page docs, no two look alike; now tractable with multimodal LLMs — https://www.furtherai.com/engineering-blogs/hardest-document-extraction-problem-in-insurance — 2026-06-23
• Signal 3 (economic): InsurTech funding topped $1bn in Feb 2026; Loss Run Pro proves a paying market for loss-run tooling at $29/user/mo (collection only) — https://fintech.global/2026/03/06/insurtech-funding-tops-1bn-in-february-as-ai-investment-surges/ — 2026-06-23 Category: Workflow automation

3. The opportunity

There are roughly 39,000 independent insurance agencies in the US; ~30,000 of them do under $1.25M in annual revenue. They control ~62% of all P&C premium. Commercial-lines renewals are their bread and butter, and every renewal worth remarketing requires assembling a clean claims history from loss runs that arrive in a dozen incompatible formats.

The incumbents are pointed the wrong way. The well-funded extraction players — Indico, Convr, Cognisure, Insurance Quantified, FurtherAI — sell to carriers and MGAs for inbound underwriting intake, at enterprise prices and enterprise sales cycles. They have no reason to build a $99/mo self-serve tool for a 4-person agency. Loss Run Pro, the one tool aimed at agencies, deliberately stops at collection (“secure upload links… send anywhere your data needs to go”) — it moves the PDF, it doesn’t read it.

So the gap is specific: a self-serve, agency-priced tool that takes the messy loss-run PDFs an agency already has and turns them into a normalized, year-over-year, carrier-agnostic claims comparison plus a submission-ready summary. Collapse a 2-hour re-keying job into 2 minutes. That’s the whole pitch.

4. Target market

• Primary customer: The account manager / CSR / commercial-lines processor at a US independent agency with 2–25 staff doing meaningful commercial P&C (general liability, commercial auto, workers’ comp, property). The buyer is the principal/agency owner; the daily user is the AM.
• Why they buy: In their words — re-keying loss runs is “tedious and exhausting,” done “in short bursts before fatigue sets in,” and it’s the bottleneck that decides whether a remarketing submission goes out in time. Every account they don’t finish remarketing renews with the incumbent by default — lost commission.
• Rough TAM reasoning: ~30,000 sub-$1.25M agencies + a chunk of the ~9,000 larger ones. If even 10,000 are commercial-lines-active and a credible product lands 3–5% of them at ~$1,200–1,800 ACV, that’s a $4–9M ARR ceiling on the core wedge alone — comfortably inside the $1–5M target with room above.
• Why now for them: 2026 is a hard market — more remarketing activity, more loss runs flowing, more renewals where the claims summary has to be sharp to win a competitive quote. The manual bottleneck bites harder exactly when volume is up.

5. Product sketch (MVP)

• Forward or upload any loss-run PDF (or batch) from any carrier — email-in address + drag-and-drop.
• AI extracts every claim line: date of loss, claim number, status, paid, reserved, incurred, cause/peril, claimant, line of business — the ~30 fields that matter.
• Auto-normalizes across carriers into one consistent schema, regardless of original layout.
• Year-over-year and carrier-over-carrier comparison view: loss ratio trend, large-loss flags, open vs. closed, frequency by peril.
• One-click export to a clean claims-summary sheet (Excel/PDF) formatted for a remarketing submission.
• Confidence flags on any field the AI wasn’t sure about, so the AM verifies the 3 shaky cells instead of re-keying 300.
• Light request-and-track helper: pre-filled ACORD 611 generation + a follow-up reminder cadence (table stakes, not the core).

6. AI angle — what’s load-bearing

Entirely load-bearing. The product is the extraction-and-normalization engine. Remove the AI and you’re back to a human re-keying 30 fields across 200 pages from a layout they’ve never seen. The defining difficulty — explicitly called “the hardest document extraction problem in insurance” — is that no two carrier loss runs look alike and there’s no schema to template against. That’s precisely the class of problem multimodal LLMs newly solve and rule-based OCR templating never could. If the extraction isn’t trustworthy, there’s no product; if it is, the 2-hour job becomes a 2-minute review.

7. Localization angle (if any)

N/A — this is a US-first play. Loss runs are a US/Canadian commercial-insurance artifact tied to ACORD standards and state-level claims-history-disclosure rules. No language or payment-rail wedge. The “localization” here is vertical depth (carrier-format coverage), not geographic.

8. Business model — path to $1M–$5M ARR

• Pricing: $99/mo (solo/small, ~50 loss-run docs/mo) → $249/mo (busy commercial agency) → $499/mo (multi-location). Anchored below the $85–150/user AMS spend and framed as “one tool, agency-wide,” not per-seat — agencies hate per-seat creep.
• ACV: ~$1,500 blended.
• Rough math to $1M ARR: ~670 agencies at ~$1,500 ACV. Out of ~10,000 commercial-active agencies, that’s <7% — achievable.
• Rough math to $5M ARR: ~3,300 agencies, or fewer agencies at higher tiers plus an AMS-integration upsell. Requires the format library to be deep enough that the product “just works” on every carrier an agency throws at it — that’s the trust threshold that drives word-of-mouth.
• Expansion path: usage tiers as document volume grows; AMS write-back (push the normalized summary straight into Applied Epic / AMS360 / HawkSoft); adjacent extraction (dec pages, policy schedules) once you own the loss-run workflow.

9. Go-to-market wedge — first 100 customers

• The Big “I” + state PIA/IIABA chapters: state independent-agent associations run active forums, listservs, and CE events. Sponsor/demo at 3–4 state association meetings and post a “stop re-keying loss runs” walkthrough — this is the watering hole where commercial AMs actually hang out.
• InsuranceJournal forums + r/InsuranceProfessionals + the AMS Facebook groups (AMS360 Users, Applied Epic Users): these are full of “loss runs are killing me” posts. Reply with a 90-second before/after video on a real (redacted) multi-carrier loss run. Cold-DM the people who complained.
• Loss-run-format bounty / freemium hook: free tier that normalizes up to N docs/month. The agency feels the magic on their own worst PDF, then hits the paywall at renewal-season volume. Self-serve, no demo call needed for tier 1.
• AMS consultant + Loss Run Pro complement angle: the independent consultants who configure Applied Epic/AMS360 for agencies are trusted referrers; ClaimComb is the “reads it” layer that pairs with Loss Run Pro’s “moves it” — partner rather than compete.
• Wholesalers/MGAs as a referral wedge: they receive submissions with garbage claims summaries daily and would happily tell their producing agents “use this.”

10. Build complexity — justification

Medium. The web app, upload pipeline, export, and ACORD-611 generation are off-the-shelf. The hard, time-consuming part is reaching trustworthy extraction across the long tail of carrier formats — the first 10 carriers cover most volume, but the 80/20 tail is where trust (and the moat) is won, and it needs an eval harness + human-in-the-loop correction flow that compounds. Realistic v1 for a strong technical founder + a domain advisor: 3–4 months to a tool that’s reliable on the top ~20 carrier formats.

11. Gating checklist

Gate	Pass?	Note
Legal in target market	✅	Processing documents the agency already lawfully holds; standard data-handling/PII obligations apply.
Ethical — no harm / dark patterns	✅	Removes drudgery; confidence-flagging keeps a human in the loop on accuracy.
Market exists (evidence above)	✅	39K agencies, documented “countless hours” pain, paying adjacent tools.
1–5 person team can build this	✅	Technical founder + domain advisor; Medium complexity.
Launchable with <$50K / ₹40L	✅	Off-the-shelf stack + LLM inference; main cost is the founder’s time.

12. Feasibility score

Axis	Weight	Score	Notes
Problem intensity	20	15/20	Felt every renewal, “tedious and exhausting,” directly costs lost commissions. Not quite hair-on-fire because workarounds (offshore back-office, brute-force re-keying) exist.
Demand evidence	15	12/15	Multiple independent sources naming the exact pain; adjacent paid tools; 39K-agency base. Short of 13–15 only because I lack a clean count of agency-side WTP vs. carrier-side.
Build feasibility	15	11/15	Standard stack, but trustworthy long-tail extraction is real engineering with an eval/correction loop. 3–4 months, not 6 weeks.
Distribution clarity	15	11/15	Named watering holes (Big “I”, AMS user groups, IJ forums) and a freemium self-serve hook; conversion math still unproven.
Revenue mechanics	15	12/15	Pricing benchmarked below AMS spend; $1M needs <7% of the active base. Per-agency (not per-seat) pricing is the assumption to watch.
Time to first revenue	10	7/10	Self-serve freemium can convert in weeks once the top-20 formats are solid, but the product has to actually work before anyone pays.
Defensibility	10	5/10	Execution + a compounding carrier-format library + correction data. Real but copyable; a funded carrier-side player could turn downmarket. A 12-month head start + agency trust is the moat, not patents.
Total	100	73/100

13. Qualitative modifiers

Founder-fit tags

technical-heavy · domain-expertise-required — needs someone who can ship a reliable extraction pipeline AND someone who has actually built a remarketing submission and knows which fields matter.

Key assumptions to validate (3–5)

1. Assumption: Account managers will trust AI-extracted loss data enough to skip re-keying, given confidence flags. How to test: Run 30 real (redacted) multi-carrier loss runs past 10 AMs; measure how many fields they feel compelled to re-verify. Trust threshold ≈ they verify <10% of cells.
2. Assumption: Agencies will pay $99–499/mo for this as a standalone, not demand it be free inside their AMS. How to test: 25 pricing interviews with agency principals; a fake-door pricing page measuring checkout-intent by tier.
3. Assumption: Top ~20 carrier formats cover enough of a typical agency’s volume to feel “it just works.” How to test: Collect loss-run samples from 15 agencies, measure carrier-format coverage of the first 20.
4. Assumption: Freemium self-serve converts without a sales call. How to test: Ship the free tier to one state association’s members; measure free→paid at renewal-season volume.

Risk flags

1. Competitive risk: A well-funded carrier-side extraction player (FurtherAI, Insurance Quantified) decides to launch an agency-priced self-serve tier. Mitigation: own the agency relationship and AMS write-back before they look downmarket.
2. Trust/accuracy risk: One bad extraction on a large-loss claim that costs an agency a quote erodes trust fast. Mitigation: confidence-flagging and conservative defaults — never silently guess on dollar fields.
3. Platform dependency: Reliance on third-party LLM APIs for the core function; inference cost and model behavior can shift. Mitigation: model-agnostic pipeline, cache aggressively, keep an eval harness to swap models.
4. Defensibility ceiling: The wedge alone is copyable; the business only gets durable via the format library + AMS integrations + brand in the agency niche.

14. Structured verdict

Score:                  73/100
Verdict:                GO
Confidence:             Medium
Best-fit builder:       Technical founder (document-AI) + commercial-lines domain advisor
Time to revenue:        8–14 weeks (freemium self-serve once top-20 formats are reliable)
Capital to launch:      $15–30K ($ mostly founder time + LLM inference + sample data)
Top 3 assumptions to validate first:
  1. AMs will trust flagged AI extraction enough to stop re-keying — 30-doc trial with 10 AMs.
  2. Agencies pay $99–499/mo standalone — 25 pricing interviews + fake-door page.
  3. Top-20 carrier formats cover enough real volume — sample 15 agencies' loss runs.
Kill criteria:
  - Abandon if <50% of trial AMs trust the output enough to skip re-keying after seeing confidence flags.
  - Abandon if top-20 carrier formats cover <70% of a typical agency's real loss-run volume.
  - Abandon if a funded incumbent ships an agency-priced self-serve normalizer before your v1.

15. Next step — 1-week validation sprint

• Day 1–2: Collect 25–40 real (redacted) loss-run PDFs from 8–10 friendly agencies, spanning as many carriers as possible. Build a quick extraction prototype on the top ~8 formats.
• Day 3–4: Put the normalized output in front of 10 account managers. Watch them work: which fields do they re-verify, would they trust the export in a real submission, what would they pay?
• Day 5: Go/no-go. Falsifiable bar: ≥6 of 10 AMs say they’d trust the output enough to skip manual re-keying and name a price ≥$99/mo. If fewer, the trust/WTP wall is too high — kill or rework toward an AMS-embedded model.