Directions Brain

Directions — operational-AI reversal, YC scan, and the AX-INS-4 challenge

6 min read

Directions — operational-AI reversal, YC scan, and the AX-INS-4 challenge

Source: raw/meetings/notetaker/2026-08-09-directions-inssurance-14-30.json

Summary

Evening internal session on 2026-08-09 (after the timor-arbel-sadras advisory call earlier that day). The team read their own Aug-7 one-pager aloud and then spent most of the session pressure-testing it. The conviction that emerged: brokerage is a labor business and today’s real AI disruption sits in internal operations (broker/agency/MGA back-office), not in end-customer personalization — the opposite of the team’s founding intuition. The session’s most energizing concrete concept, “Gong for insurance,” is explicitly AI-for-brokerage tooling, the category AX-INS-4 forbids. Nothing was decided; the team closed on “this is an interesting opportunity” but “we’re too high in the air — sharpen the segment.”

Key takeaways

  • The stated reversal. “However much we thought the AI wins would go to end users and personalization — in the end most of the AI optimization is in the internal operations of the MGA and the carrier, not at the broker level.” Underwriting was explicitly ruled out as an opportunity (“insurance was one of the first data-science professions” — LLMs didn’t unlock it).
  • Brokerage as a labor business, quantified in-room: 60–70% of a brokerage’s activity is operations, not sales (“he has 50 agents, net three of them do sales”); >50% of broker cost is labor. Jobs named: services, quoting, chasing carriers, renewals, certificates, back-office.
  • AMS, not CRM — the systems insight. Brokers run on Agency Management Systems (policies, endorsements, carrier downloads, commission accounting, certificates), whose embedded CRM is thin. AMSs are non-proactive and capture no conversation context — the identified wedge.
  • “Gong for insurance.” Record broker↔client calls → unstructured to structured → auto-feed the AMS → coverage-gap detection, cross-sell, agent coaching, preference memory (“he hates Phoenix — never quote Phoenix again”). Worked example: client mentions a new Dallas warehouse, +20 hires, and direct online sales → derive property/GL schedule review, workers’ comp and payroll exposure, GL product-liability check → update AMS, request missing info, prepare endorsements. GTM idea: two-way AMS API integration, then displace upward. Counter-argument raised in-room: “Gong is CRM — are you building CRM for insurance?” and revenue intelligence doesn’t touch the actual job-to-be-done.
  • A real design partner surfaced: goldfus, an Israeli agency. Most of his time goes to operations; he can’t grow his book because of the operational load; says he would pay tens of thousands of shekels for AI removing “90% of the operation”; has been pitched by AI vendors and none nailed it. Caveat flagged by the team: stated willingness to pay ≠ payment.
  • B2C doubted on user-rationality grounds: “because we come from SaaS we think users are rational… the guy’s policy renews every year and he has no idea what he’s signing.” Landed on a four-phase autonomy spectrum, leaning to phases 2–3 (D2C at sale, human agent when claims get complicated), with the leverage argument that AI operations let one agent carry a book ~20x larger.
  • New axis raised for the first time: heavy lines (pension/life) over P&C. Argument: “anything that doesn’t require a human is much harder to break into — it’s already digitized (Lemonade); the heavy lines have far more room for an operations revolution” for B2B.
  • cover’s moat is regulatory, not technical — Israel’s mandated rails (Har HaBituach, pension clearinghouse) have no US equivalent, which is why its value prop fragments in the US. Convergent with policy-ingestion-requires-a-data-rail, reached independently in-session; canopy-connect named as the closest US analog (“the Plaid of insurance”). A “Cover B2B in Israel” play was floated and rejected as unambitious.
  • “Sexy MGAs don’t invent categories” — “they took an existing category and made it far more accessible” — versus corgi, which insured a previously uninsured risk and thereby bought a simple GTM. Consensus lean for a niche: be an MGA cooperating with existing carriers.
  • YC scan (~15 insurance startups, last ~2 years): “everyone wants to be a brokerage.” Notable: RiskyTix (?) — commercial insurance for robotics, autonomy, data centers, energy — “took your Corgi and merged several verticals,” i.e. the robotics wedge is partly pre-empted; Copycat (?) — “saves brokers 5–10 hours a week,” identified as exactly the Goldfus use case; Kinro, Casey, Fanta (?), Fernstone, Accolite (?) (construction wedge, “wraps your business with a specific wedge”); plus underwriting/claims plays (Florin (?), Verdex (?), Solva (?), Avalon AI (?), Amra (?)) and AI-agent-liability carriers (Mount (?), Climbing (?)).

Axiom tension (unresolved — no decision taken)

  • AX-INS-4 (no “AI for brokerages” tooling) is openly contested. After reading the axiom aloud: “and that’s one of the axioms I’m not sure I agreed to.” Grounds: >50% of broker cost is labor (a large savings pool), and B2C brokerage isn’t a high-margin software business once CAC and integrations are paid. The session’s best product idea sits inside the forbidden category.
  • AX-INS-2 (AI brokerage serving END customers) is strained by the same reversal and by the pension/heavy-lines axis.
  • AX-INS-5 (category creation) reaffirmed as an option but not advanced — Corgi’s move called “exotic, not really repeatable,” and RiskyTix (?) appears to have already taken the robotics cross.
  • Per the axioms protocol these remain active axioms until a logged team decision amends them. Recommended next step: an explicit decision session, since the team is currently operating against an axiom it no longer fully endorses.

Decisions

  • None formal. Explicit non-decision: “better to wait and close on something good than close on garbage.”
  • Soft conclusions: operations (not underwriting, not consumer personalization) is where AI value sits today; pure rational-user B2C is unrealistic; for a niche, MGA-with-existing-carriers over inventing a category; “Cover B2B in Israel” rejected as unambitious.

Action items

  • Team — sharpen the target segment before the next alon-huri conversation (“we’re too high in the air”); the Tuesday meeting is expected to be answers, not brainstorm.
  • saar-arbel — re-interview goldfus properly as a design partner and validate the “would pay tens of thousands” claim.
  • guy-barkat — research which AMS/CRM systems brokers actually run on and their capability gaps; identify Goldfus’s specific legacy system.
  • Team — clarify the regulatory position on recording/consent for insurance calls containing pension, salary, and premium data.
  • Team — talk to more and better-targeted people (self-criticism: ~6 people so far, “just not the right people”; Huri may have expected ~20).
  • Team — dig into Accolite (?) (construction wedge) and Kinro specifically.

Open questions

  • Broker, MGA, or both? Which processes exactly do we automate, internal and client-facing?
  • Simple lines (P&C, already digitized) vs heavy lines (pension/life, more operational upside)?
  • Is “AI for brokerage” genuinely a derivative-of-a-derivative TAM, or was that axiom adopted too fast?
  • Is “Gong for insurance” a CRM play in disguise?
  • Is the monday enrichment/data advantage differentiated or commodity?
  • Is goldfus’s pain representative — “how much is his problem everyone’s problem?”
  • Does a broker have any incentive to cut a client’s duplicate coverage? (retention vs commission conflict)
  • What is the US equivalent of Har HaBituach — is canopy-connect it?

Graph View

Backlinks