Directions Brain

Technology Vision & Architecture (the 20-job machine)

6 min read

Technology Vision & Architecture (the 20-job machine)

Full doc: outputs/technology-vision-and-architecture.md. One-page architecture: outputs/architecture-map.html (desktop artifact ainsure-system-architecture). Built from Saar’s 20-job product outline plus the research in insurance-tech-integration-stack and insurance-checkup-funnel-design.

Definition

The engineering answer to “how do we build this”: a brokerage is a machine that moves a risk from a business exists to a policy is bound and stays correct, and every step of it is currently a human re-typing something a computer already knows. Five planes — Acquisition, Placement, Service, Ledger & Compliance, and the Brain — with all 20 jobs from the outline placed and phased P0–P3.

Key points

  • Three governing rules. (1) Buy every rail, own every representation — ingestion, quoting connectivity, market access, AMS and capture are all rentable and none is a moat. (2) Deterministic first, AI on the tail — the biggest available engineering mistake is pointing an LLM at what is actually a join; form number + edition date is a lookup, limits are arithmetic, renewals are a calendar. (3) The compounding assets are the product.
  • The four owned assets: ① observed-appetite graph (P(quote) × P(bind|quote) per carrier × NAICS-6 × state × size — every submission, especially every decline, is a labelled example), ② form corpus (each carrier form parsed once, keyed by number+edition, with precomputed edition deltas), ③ conversation corpus, ④ renewal graph. Competitors can buy every tool; they cannot buy these, because they are produced by running the machine.
  • ⭐ The benchmark evidence inverts the intuitive build order. The Vals Legal AI Report (blind, independent, vendor-participating) shows AI beating human lawyers on Document Q&A 94.8% vs 70.1% — but losing on redlining, 65.0% vs 79.7% (verified against the report). Redlining is two-document differential analysis with materiality judgment: the closest published analogue to policy comparison and policy checking, and the only task where humans win. Corroborated by CLAUSE (EACL 2026): omission detection “exceptionally challenging,” citation matching <14%; ContractEval: best F1 0.644 with per-category spread from 0.9 to near zero; the Quebec insurance RAG benchmark where Gemini-2.5-Pro collapsed 74.65% → 14.06% with naive retrieval enabled.
  • The shipped state of the art is lower still. Applied Epic’s policy checking is a paragraph-level text diff — automated highlighting, with “decision-making about significant versus insignificant differences” left entirely to humans. ReSource Pro tiers its AI product by complexity; Patra tiers its by who assumes the E&O risk. Nobody credible ships AI-only on complex commercial policies. And no vendor publishes an accuracy number for comparison — they all advertise speed (20x, 30x, 67% faster), because no benchmark exists.
  • Design answer: recall-optimised and over-flagging by design (a missed discrepancy is a ~$40k E&O claim; a false flag costs 30 seconds); every asserted term carries a verbatim quote plus form/page anchor that the pipeline string-matches back into the source before display — non-matching citations are dropped, not shown; measured per-coverage-element, never aggregate; a silent-miss metric as a release gate.
  • Portal automation: the quoted benchmark is the wrong one. Skyvern scores 85.8% on WebVoyager but 46.6% on WebBench write tasks — and write tasks are what a carrier submission is. A submission is 60–120 actions, so 60% end-to-end needs >99.5% per-action. Recommendation reversed from “build a portal agent” to record-and-replay with LLM repair, top 15–25 portals only, human/BPO tail for the rest. Failure mode that matters: agents don’t recognise they’ve failed — so “submitted” is an assertion, never an assumption, and every run ends with an independent verification read.
  • The legal constraint is commercial, not judicial. Carrier agency terms (e.g. Auto-Owners) prohibit credential sharing in standalone all-caps, and Ryanair v. Booking.com produced a CFAA finding for automated access behind a login. But the enforcement mechanism is appointment termination by email, not litigation — which should drive architecture. The strategic inversion: be the first brokerage to ask carriers to allowlist a signed agent (Web Bot Auth is shipping in Cloudflare and AWS WAF), turning the biggest legal risk into a moat.
  • Appetite is a data-acquisition problem, not an integration problem. Only ~1.2% of 509 P&C carriers annotate appetite at 6-digit NAICS; PDF guides over-assert inclusions 2.14× versus the same carrier’s own web page (κ=0.25). Coterie’s own docs state the API “does not provide an endpoint that provides all appetite information” and that the Integration Manager emails a mapping spreadsheet (verified) — while requiring an AKHash (NAICS + description MD5), making Ask Kodiak’s identifier a de facto standard. Meanwhile Federato sells carriers winnability scoring: carriers are scoring our submission quality right now, so cheap automation that sprays 10 markets destroys the relationship that is the actual moat. Hence the hard cap at 4–6 submissions per risk.
  • Two reclassifications from the outline. Renewal prep (job 13) is P1, not late — per H-INS-3 most check-up completers are mid-term and cannot convert, so the renewal graph is the acquisition engine. COIs (job 7) are an acquisition and retention channel, not back-office — the most-used surface in the product, and every request reveals a contract requirement, i.e. a coverage requirement and an upsell signal. Jobs 19–20 (“Operational”, “Broker↔Carrier”) are axes, not jobs; suggest retiring them as list items.
  • Tech-side GTM has four mechanisms: the funnel is the first demo; the renewal graph turns one-shot acquisition into an annuity (the CAC argument); submission quality buys carrier access faster than volume does; and publishing honest per-coverage recall numbers would be the first credible accuracy claim in the category, arriving just as InsureBench and Yardstick make their absence conspicuous.

Interpretation (not fact)

  • The strongest strategic read: the market’s inability to measure itself is the opening. Everyone competes on speed because nobody can compete on accuracy. A paired quote↔issued-policy corpus with planted defects is simultaneously the eval set, the moat, and the marketing claim — and it cannot be bought.
  • Hire the evals/golden-corpus owner earlier than feels comfortable (P1). Nobody accumulates measured accuracy by accident.

Open questions

  • Does Canopy’s commercial pull reliably return the forms schedule? Gates both Tier-B findings and asset ②.
  • Ask Kodiak commercial terms — and the conflict of buying an appetite prior from Ivans/Applied.
  • Three-portal pilot to validate the ~$1–4/carrier model cost and the replay-first architecture before committing engineering.
  • Will any carrier agree to sanctioned agent access? Cheapest experiment in the plan, highest strategic upside.

Axiom check

Consistent with AX-INS-8 (P0 is Plane 1 only), AX-INS-7 (the four assets are the continuity claim), H-INS-2, H-INS-3, H-INS-10. ⚠ AX-INS-4 tension, flagged explicitly: jobs 7–17 are in isolation exactly the broker back-office that FurtherAI/Qumis/Patra sell. The distinction holding us inside the axiom is that we build them for ourselves and never sell them to brokers — a distinction that is easy to state and easy to erode. A future “we could license the policy-checking engine” conversation is an axiom-amending decision, not a product decision.

Graph View

Backlinks