SilentGap

1. One-liner

Catches the generative-AI exclusion buried in a client’s renewal packet before the agent lets them sign it.

2. Trend signal — why now?

In January 2026, ISO (Verisk) issued three new generative-AI exclusion endorsements for commercial general liability: CG 40 47, CG 40 48, and CG 35 08. These let carriers strip coverage for bodily injury, property damage, and personal/advertising injury arising out of generative AI — from an otherwise ordinary CGL policy.

The important part is not the forms. It’s how they arrive. They show up at renewal, quietly, with no premium credit and usually no conversation. One documented case: the exclusion sat on page 47 of a 63-page renewal package, the broker never raised it, and the risk manager hadn’t read that far. State regulators approved more than 80% of carrier filing requests for these exclusions. Berkshire Hathaway, Chubb, Travelers, Berkley, and Hamilton Select have all been named as filing or adopting AI exclusions.

Meanwhile HSB/Munich Re survey data puts small-business AI tool adoption at 74%. So you have a large population of insureds who use the exact technology their renewal is quietly carving out, and an agent channel that mostly isn’t flagging it. Aon, Gallagher, and Lockton have all publicly warned corporate clients. Nobody is warning the dry cleaner, the marketing shop, or the three-truck HVAC contractor — because their agent is a 6-person shop with 800 renewals a year and no tooling.

That’s the gap. The exclusion is new enough that it isn’t in anyone’s checklist, and boring enough that it hides in a forms schedule.

Provenance:

• Signal 1 (demand): AI exclusion found on page 47 of a 63-page renewal packet, broker never discussed it; brokers “generally not flagging” these — https://lawandkoffee.substack.com/p/your-insurance-policy-just-changed — 2026
• Signal 2 (feasibility): ISO CG 40 47 / CG 40 48 / CG 35 08 generative-AI exclusions effective January 1, 2026; >80% of state filing requests approved; Berkshire, Chubb, Travelers, Berkley, Hamilton Select named — https://www.ajg.com/news-and-insights/iso-introduces-generative-ai-exclusion-in-commercial-general-liability-policies/ and https://www.rmstudygroup.com/blog/does-your-gl-policy-still-cover-ai-the-exclusion-your-next-renewal-may-already-have — 2026-01-01
• Signal 3 (economic): ~39,000 US independent agencies, ~30,000 under $1.25M revenue; E&O defense costs average $35,000–$75,000 per claim; “failure to disclose material policy changes at renewal” is a named, recurring E&O claim pattern — https://www.producerflow.com/blog/us-insurance-agency-producer-statistics and https://www.agentsofamerica.org/articles/understanding-the-common-causes-of-insurance-agents-e-038-o-claims/ — 2026 Category: Regulatory arbitrage (insurance-form shift), with a tech-unlock delivery mechanism

3. The opportunity

Two things are true at once, and the space between them is the business.

One: carriers are adding a brand-new exclusion class that materially changes what a policy covers, and adding it silently at renewal.

Two: the agent who hands over that renewal carries E&O liability for not explaining material changes — and has no practical way to catch them. A mid-market commercial agency runs ~800 renewals a year, roughly 45 minutes of honest comparison each, on 30–80 page documents. Attention degrades around page 20. That math produces misses. It’s designed to.

The incumbents here are real but aimed elsewhere. Nomad Data does AI policy comparison built for “underwriting teams” and “risk selection” — enterprise carriers and large brokerages, controlled workflows, security review, no self-serve pricing published. Regure sells the same general full-policy comparison story to “mid-market commercial agencies with 800 active policies.” Both sell comprehensive diffing as an enterprise platform purchase.

Nobody sells the single most urgent question as a cheap, self-serve product: did an AI exclusion just get added to this client’s renewal, and can I prove I told them?

That’s the wedge. Full-policy diff is a platform sale with a long evaluation cycle. “Screen this renewal for the 2026 AI exclusions and generate the client notice” is a $99/mo credit-card purchase a principal makes on a Tuesday because they read something scary in IA Magazine. Same underlying capability, radically different sales motion — and it lands in the ~30,000 small agencies that the enterprise vendors structurally cannot afford to serve.

The second-order insight: the deliverable isn’t the detection. It’s the signed client acknowledgment. Detection protects the client. The acknowledgment protects the agent — and the agent is the one holding the credit card.

4. Target market

• Primary customer: Principal or commercial-lines manager at a US independent P&C agency, 3–25 employees, under $1.25M annual revenue, 200–1,500 active commercial policies. The person who signs the E&O check and personally reviews the ugly renewals.
• Why they buy: Not “compliance.” Fear of a specific, nameable E&O claim: client’s AI-assisted work causes a loss, carrier denies on CG 40 47, client’s lawyer asks the agent “when did you tell my client about this exclusion?” Defense costs alone run $35,000–$75,000 per claim before anyone decides who was right. A $1,200/year tool that produces a dated, signed acknowledgment for every renewal is cheap insurance against that question.
• Rough TAM reasoning: ~39,000 US independent agencies; ~30,000 under $1.25M revenue. Realistic serviceable slice — agencies with enough commercial lines to care and enough tech-comfort to buy self-serve — call it 8,000–12,000. At $1,200–$3,600 ACV that’s a $10M–$40M addressable line. More than enough for a $1M–$5M business, far too small to attract a funded competitor.
• Why now for them: The exclusions turned on January 1, 2026. Renewal cycles are rolling through them right now, and the first denied AI claims — plus the first E&O suits naming an agent who didn’t flag it — are the trigger event this market is waiting for. Getting in before that news breaks is the whole timing play.

5. Product sketch (MVP)

• Drop in a renewal packet PDF (or the expiring policy plus the renewal quote) and get a verdict in under two minutes.
• Flags the specific 2026 AI exclusion family — CG 40 47, CG 40 48, CG 35 08 — plus carrier-proprietary AI exclusion wording that isn’t an ISO form.
• Side-by-side forms-schedule diff: what’s on the renewal that wasn’t on the expiring policy, ranked by coverage impact, not document order.
• Auto-generated client notice letter in plain English — what changed, what it means, what their options are — ready to send.
• Signed acknowledgment capture: client e-signs, timestamped record stored in the agency’s file. This is the E&O artifact.
• Renewal watchlist: upload the book once, get a 60-day-ahead queue of which renewals to screen.
• Exportable audit log the agency can hand to its own E&O carrier at its renewal.

6. AI angle — what’s load-bearing

Remove the AI and this product doesn’t exist. The core task is reading a 63-page unstructured PDF — declarations, forms schedule, endorsements, manuscript wording, referenced forms — and answering a semantic question: has coverage for generative AI been removed, by any wording, ISO or proprietary?

Regex finds “CG 40 47.” Regex does not find a carrier’s own endorsement titled “Emerging Technology Limitation” that accomplishes the same thing in different words. That’s the actual job, and it’s exactly what modern long-context document models are now good enough and cheap enough to do. Two years ago this was an OCR-plus-brittle-rules project with an accuracy problem. Today it’s an API call over a parsed PDF with citation back to the source page — which matters, because the agent must be able to verify the finding before sending a letter under their own name.

The second load-bearing use: drafting the client notice. Turning “CG 40 47 01 26 attached” into a paragraph a bakery owner understands is a translation task, and it’s the part agents most want to avoid writing 800 times.

7. Localization angle (if any)

N/A — this is a US play. The wedge is ISO form numbers, US state filing approvals, and US agent E&O exposure. The UK, Canadian, and Australian markets are developing similar AI exclusions but on different form architectures with different intermediary liability rules; each would be a separate product, not a translation. Deliberately staying US-only keeps the form library small and the accuracy high.

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

• Pricing: $99/mo Solo (up to 25 screens/mo), $249/mo Agency (100 screens, multi-user, acknowledgment capture), $499/mo Book (unlimited screens, full-book watchlist, E&O audit export).
• ACV: ~$2,400 blended. Skews toward the $249 tier — that’s the natural fit for a 6-person agency running 800 renewals.
• Rough math to $1M ARR: 350 agencies × $249/mo × 12 ≈ $1.05M. That’s roughly 3% of the ~12,000 serviceable agencies. Achievable.
• Rough math to $5M ARR: ~1,700 agencies on the $249 tier, or ~1,100 with meaningful mix-shift to $499 plus a per-seat expansion. Requires winning the association channel (state Big “I” chapters) and becoming the default checkbox in E&O risk-management guidance. Plausible, not guaranteed.
• Expansion path: Start with the AI exclusion, expand to the full “material change at renewal” screen — cyber sublimits, communicable disease, PFAS, abuse/molestation, wildfire deductibles. Every new exclusion class the market panics about is a feature release and a price increase. The AI exclusion is the wedge; “material change detection” is the durable product.

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

• The E&O fear letter. Scrape state Big “I” chapter member directories (most publish member agency lists publicly — NY, TX, CA, FL, NC alone cover thousands). Send a plain-text email to the principal: “Did your January renewals include CG 40 47? Here’s a free screen of one policy — send us the PDF.” Free single-policy screen as the hook. Expect 2–4% reply on a well-written fear-based email to this audience; 3,000 emails → 60–120 conversations → 20–40 paid.
• Own the search term before anyone else. “CG 40 47” and “generative AI exclusion” have near-zero commercial competition right now and a highly qualified audience. Publish the definitive free form-by-form breakdown plus a live carrier-adoption tracker. This is the [consultancy-maintains-the-list] play — somebody is going to hand-maintain that list; it should be the product’s funnel.
• State association webinars. State Big “I” chapters run continuing-education sessions and are actively hunting 2026 E&O content. Offer a free 45-minute CE-eligible session on AI exclusions. One webinar to 200 agency principals with a soft product mention converts far better than any ad. Target 6 chapters in the first two quarters.
• The E&O carriers themselves. Agency E&O writers (Swiss Re Corporate Solutions, Utica, Westport) publish risk-management bulletins and sometimes offer premium credits for documented procedures. Getting listed as an accepted risk-management tool is slow but is the single highest-leverage distribution asset in this market. Start the conversation at month 3, expect nothing until month 12.
• IA Magazine / Insurance Journal comment sections and LinkedIn. The agency-principal crowd is unusually active on LinkedIn. Post the carrier-adoption tracker updates weekly; it’s genuinely useful and it’s how this audience discovers tools.

10. Build complexity — justification

Low. PDF parsing plus a long-context model plus a small hand-curated library of known AI-exclusion forms and carrier-proprietary wordings. No custom model training, no integrations required for v1 — the agent uploads a PDF. E-signature is an off-the-shelf API. A competent solo builder ships a credible v1 in 6–8 weeks; the real work is the form library and getting false-negative rate low enough to be trustworthy, which is curation effort, not engineering difficulty. AMS integrations (Applied Epic, EZLynx, HawkSoft) are a month-6 concern, not an MVP concern.

11. Gating checklist

Gate	Pass?	Note
Legal in target market	✅	Document analysis tool. Sells to licensed agents; product flags and drafts, the agent advises. Must avoid rendering coverage opinions itself.
Ethical — no harm / dark patterns	✅	Strictly protective: surfaces coverage changes clients would otherwise sign blind.
Market exists (evidence above)	✅	~30,000 small agencies, documented E&O claim pattern, enterprise vendors already selling the adjacent capability upmarket.
1–5 person team can build this	✅	Solo-buildable v1.
Launchable with <$50K / ₹40L	✅	Under $10K: inference, PDF parsing, e-sign, landing page.

All five pass.

12. Feasibility score

Axis	Weight	Score	Notes
Problem intensity	20	16/20	Real and dated, with a $35K–$75K E&O tail. Not quite hair-on-fire yet — the pain is anticipatory until the first denial makes the news. That’s the honest discount.
Demand evidence	15	12/15	Strong indirect evidence: enterprise vendors selling the adjacent product, documented broker misses, named carriers, association warnings. Missing: verbatim small-agency principals asking for this specific thing.
Build feasibility	15	13/15	6–8 weeks solo on off-the-shelf components. Accuracy curation is the only real work.
Distribution clarity	15	12/15	Named lists (Big “I” chapter directories), named channel (state CE webinars), uncontested search terms. Conversion math is estimated, not tested.
Revenue mechanics	15	11/15	$249/mo is well within agency tech budgets and 350 customers gets to $1M. Risk: small agencies are famously slow software buyers and may treat this as a once-a-year need rather than a subscription.
Time to first revenue	10	8/10	Free-screen-to-paid funnel is short; realistically 6–10 weeks to first paying agency.
Defensibility	10	4/10	Weakest axis, and I won’t dress it up. The detection is copyable. The moat is the curated carrier-wording library, the acknowledgment archive (switching cost grows with every signed record), and association relationships. That’s a 12-month head start, not a wall.
Total	100	76/100

13. Qualitative modifiers

Founder-fit tags

technical-heavy · sales-heavy

Needs someone who can build the document pipeline and is willing to email 3,000 agency principals and run webinars. An insurance-domain advisor is close to mandatory for the form library and for credibility in the association channel.

Key assumptions to validate (3–5)

1. Assumption: Small agency principals feel enough E&O anxiety about AI exclusions to pay ~$249/mo. How to test: 40 phone calls to agency principals from Big “I” directories; ask what they did about CG 40 47 at January renewals. If most say “what’s that?” — that’s a marketing problem, not a demand problem. If they say “not worried” — that’s a kill signal.
2. Assumption: The exclusions are actually appearing at meaningful rates in real renewal packets. How to test: Collect 50 real renewal packets from friendly agencies, measure actual AI-exclusion incidence. Below ~10% and the urgency evaporates.
3. Assumption: Detection can hit acceptable accuracy on carrier-proprietary wording, not just ISO forms. How to test: Build the library against those 50 packets, measure false-negative rate. Above 5% false negatives and the product is a liability rather than a protection.
4. Assumption: Agencies will treat this as a subscription, not a one-time audit. How to test: Watch whether trial users return for a second and third batch of renewals within 60 days.

Risk flags

1. Market timing: This is the central risk. The idea is early — deliberately. If the first big AI-exclusion denial story doesn’t land in the next 12 months, urgency stays theoretical and the sale stays hard. If it lands, this product is in exactly the right place. Directional bet on a dated regulatory change.
2. Incumbent drift downmarket: Nomad Data or Regure could ship a cheap self-serve tier. Their enterprise positioning and sales motion make it unlikely fast, but it’s a 12–18 month exposure.
3. Carrier retreat: Travelers subsidiaries obtained approvals for AI exclusions and then filed notices saying they would not adopt them. If broad adoption reverses, the wedge narrows to the full material-change product sooner than planned. Approval is not adoption — this needs continuous monitoring. (Same shape as the rescinded-order trap: verify adoption, not just filings.)
4. Professional liability of the tool itself: If the product misses an exclusion and an agent relied on it, the agent will point at the vendor. Requires careful positioning as a screening aid, explicit non-advice disclaimers, and the product’s own E&O policy — which, ironically, should be checked for an AI exclusion.

14. Structured verdict

Score:                  76/100
Verdict:                GO
Confidence:             Medium
Best-fit builder:       Technical solo founder with an insurance-agency advisor
Time to revenue:        6–10 weeks
Capital to launch:      $8–10K
Top 3 assumptions to validate first:
  1. AI-exclusion incidence in real renewal packets — collect 50 packets, measure rate
  2. Principal willingness to pay $249/mo — 40 discovery calls from Big "I" directories
  3. False-negative rate on carrier-proprietary (non-ISO) wording — measure against the 50-packet corpus
Kill criteria:
  - Abandon if AI-exclusion incidence in 50 real renewal packets is below 10%
  - Abandon if fewer than 5 of 40 agency principals will commit to a paid pilot
  - Abandon if false-negative rate on non-ISO wording stays above 5% after library curation
  - Abandon if major carriers broadly reverse adoption (watch Travelers-style non-adoption filings)

15. Next step — 1-week validation sprint

• Day 1–2: Pull Big “I” chapter member directories for 5 states. Build a list of 300 agency principals. Send the fear-letter email offering a free screen of one renewal packet. Simultaneously ask 10 friendly agencies for anonymized January–March 2026 renewal packets.
• Day 3–4: Manually screen every packet received. Measure real AI-exclusion incidence. Hand-build the first version of the wording library from what actually shows up — ISO forms and proprietary variants. Run 15 discovery calls with principals who replied.
• Day 5: Decide. Go if: ≥10% of collected packets contain an AI exclusion AND ≥5 of the principals contacted will pay for a pilot before the product exists. No-go if: incidence is negligible or principals uniformly shrug — in which case the real product is the broader material-change screen, and the AI exclusion was just a headline.

The falsifiable number is exclusion incidence in real packets. Everything else is opinion until that’s measured.