SB StartupBasket
All ideas
76 /100 GO Medium complexity

SealedSet — AI-use attestation trail for design firms

Proves which drawings your people authored, so an AI exclusion can't void the claim.

InsurTech SaaS US Compliance-drivenSMBAI-agentSolo-builderProfessionalServices
— views
Evaluation Scores
76/100

GO

Overall Score

17
Problem
12
Demand
11
Build
12
Distrib.
12
Revenue
7
Time
5
Defense

SealedSet

1. One-liner

Proves which drawings your people authored, so an AI exclusion can’t void the claim.

2. Trend signal — why now?

On 1 January 2026 the insurance industry did something it almost never does: it wrote a brand-new exclusion for a technology that half its policyholders were already using every day.

Verisk/ISO released CG 40 47 01 26 — Generative Artificial Intelligence Exclusion, plus two siblings (CG 40 48, Coverage B only; CG 35 08, products & completed operations). The operative language excludes bodily injury, property damage and personal/advertising injury “arising out of, or attributable to, generative artificial intelligence.” ISO forms underpin roughly 82% of global P&C policies. By April 2026, W.R. Berkley, Chubb, Travelers, Berkshire Hathaway and Cincinnati Financial had filed to adopt these endorsements or their own proprietary AI exclusion language, and state regulators had approved more than 80% of submitted filings.

The professional liability side moved even faster and hits design firms harder. Berkley filed an “absolute” AI exclusion for E&O that names ChatGPT, Bard, Midjourney and DALL-E explicitly. Philadelphia Insurance and Hamilton Select have both excluded AI-related claims from E&O products. AIG and Great American are in the filing queue.

Now put that against adoption. AIA research (March 2025) found roughly one-third of architecture firms use AI tools daily — 61% of large firms, 27% of small ones. Autodesk, Bluebeam and Microsoft have shipped generative features straight into the tools these firms already run. Nobody opted in to “using generative AI”; it arrived in a product update.

The two curves are crossing. Adoption is ambient and undocumented. The exclusion is broad and now standard. And “arising out of” is one of the widest causal standards in insurance law — a human reviewing AI-generated content does not automatically shield the claim if AI sits anywhere in the causal chain.

The brokers have already priced the consequence. From an A&E renewal advisory: firms entering renewal without documentation stop getting “average single-digit increases” and start getting targeted ones — “question marks pay the spread.” Underwriters now ask, in writing, “how the firm uses AI, who reviews its outputs, and what the documented verification process looks like.”

Almost no 12-person structural engineering firm can answer that question with anything but a shrug.

Provenance:

3. The opportunity

Everyone selling into this space is selling policy — a written AI governance framework, a usage policy PDF, a training module. Credo AI, Holistic AI and Trustible sell enterprise AI governance at $25K–$200K/yr with no published SMB pricing. Law firms sell the governance memo. The AIA Trust publishes guidance telling firms to “adopt a written AI governance policy.”

A policy document is not evidence. When a slab cracks in 2029 and the carrier’s coverage counsel reaches for CG 40 47, the question is not “did you have a policy?” It is “prove this specific sheet was not produced by, or attributable to, generative AI.” A PDF written in 2026 saying “we require licensed review of AI output” does nothing. What survives a coverage fight is a per-deliverable, contemporaneous, tamper-evident record: this sheet set, this revision, these tools were active in the authoring environment, this licensed professional reviewed and sealed it, on this date.

That artifact does not exist anywhere in the small-firm A/E stack. The gap is precisely the shape my catalog keeps rewarding — the vendors sell the score, nobody sells the per-line-item evidence the counterparty will attack.

The disruption angle is not against a software incumbent. It is against the status quo of “we’ll figure it out if a claim comes.” That status quo now costs money at renewal, before any claim, because the underwriter prices the question mark. That’s the wedge: this isn’t a bet on a hypothetical 2031 lawsuit, it’s a line item on the invoice the principal signs this quarter.

4. Target market

Primary customer: Principal / managing partner / risk lead at a US architecture or engineering firm with 3–50 employees — structural, MEP, civil, and small full-service architecture practices. Revenue $500K–$12M. They carry $1M/$2M professional liability, renew annually, and the principal personally signs the renewal application.

Why they buy: Three converging pressures, in their language:

  • The renewal application now has an AI section and they have to answer it truthfully with no records to back the answer.
  • Their E&O endorsement schedule may already carry an AI exclusion they haven’t read — the advisory literature is explicit that firms “assumed E&O would roll over with identical language” and got caught.
  • Their staff are using Copilot, Revit generative tools and ChatGPT daily whether or not the firm has a policy, and the principal knows it.

The purchase is not “AI governance.” It is “I need to be able to answer the underwriter’s question and not be the question mark.”

Rough TAM reasoning: There are more than 19,000 architecture firms in the US. Firms of 1–9 staff are 75.2% and 10–49 staff are 18.5% — roughly 93.7% of the market, ~17,800 firms, sit under 50 employees. Monograph’s 2026 benchmarks put 90% of all A&E firms in the 1–50 band, which extends the pool well past architecture alone into structural/MEP/civil engineering practices. Realistic serviceable pool for a bootstrapped play: 20,000–35,000 US firms. I need ~400 of them to hit $1M ARR.

Why now for them: Their renewal date. Every firm has one, they arrive continuously through the year, and 2026 is the first cycle where the AI question is on the form. This is a rolling deadline across the entire market — not a single cliff that passes.

5. Product sketch (MVP)

  • Tool inventory sweep — connects to the firm’s Microsoft 365 / Google Workspace admin and Autodesk account, and reports which generative-AI features and licences are actually active per user. Most principals genuinely do not know. This is the “oh no” moment that sells the product.
  • Per-deliverable authoring record — for each sheet set / calc package / spec section issued, captures who authored it, what AI-capable tools were active in that workflow, what was human-originated, and what was AI-assisted-then-reviewed.
  • Licensed review attestation — the professional of record signs off per deliverable, on the record, with a timestamp. One click from the issue workflow, not a separate chore.
  • Tamper-evident sealing — each record is hash-sealed at issue time so its date can’t be quietly rewritten three years later when a claim lands. Contemporaneous beats reconstructed, every time.
  • Renewal answer pack — one PDF, generated on demand, that answers the underwriter’s AI section: tool inventory, written usage policy, verification workflow, and the count of sealed deliverables in the period. This is the deliverable the broker actually wants.
  • Claim retrieval — given a project and a date, pull every sealed record for the deliverables in scope, in an exhibit-ready format for coverage counsel.
  • Exclusion watch — parses the firm’s uploaded policy endorsement schedule and flags whether CG 40 47 / CG 40 48 / CG 35 08 or a carrier-proprietary AI exclusion is attached, and to which coverage part.
  • Staff AI usage policy — generated, versioned, acknowledged by each employee, with the acknowledgement log attached to the renewal pack.

6. AI angle — what’s load-bearing

Two places, and neither is decorative.

Classifying the deliverable. The hard problem is not storing a record, it’s determining what actually happened. A sheet set arrives as a PDF or an exported model. Something has to read the drawing set, the revision clouds, the file metadata and the tool-telemetry signals and produce a defensible classification: human-authored, AI-assisted-and-reviewed, or AI-originated. Doing this by hand per sheet is exactly the friction that makes firms not do it. A vision-plus-metadata model doing it in seconds per sheet is what makes the workflow survive contact with a busy 12-person firm.

Parsing the policy. Endorsement schedules are dense, inconsistent across carriers, and full of proprietary language that isn’t the ISO form. Extracting “is there an AI exclusion here, on which coverage part, with what carve-backs” from a 60-page policy PDF is a genuine document-understanding task, and it’s the thing that produces the first jaw-drop in the sales conversation.

Remove the AI and this becomes a manual logging chore that firms abandon in week three — which is precisely why the artifact doesn’t exist today despite everyone knowing they should keep it. The AI is what makes the record get created at all.

There’s a pleasing irony that the product uses AI to prove you didn’t use AI. I’d lead with it in the marketing.

7. Localization angle

N/A — this is a US-first play. The wedge is ISO form language and US carrier filing behaviour, which is a US-specific artifact. The UK/EU/Australian professional indemnity markets are moving in the same direction but on different forms, different timelines and through different brokers. Expansion path exists (Canada is close to a copy-paste; UK PI follows via Lloyd’s syndicate wordings) but it’s a year-two question, not a wedge.

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

Pricing: Tiered by headcount, sold annually.

  • Solo / 1–4 staff: $99/mo ($1,188/yr)
  • 5–15 staff: $249/mo ($2,988/yr)
  • 16–50 staff: $499/mo ($5,988/yr)
  • Renewal Pack one-off: $499 — for firms who want the underwriter answer now and will subscribe later. This is the door-opener.

Why this price clears: small A/E firms already pay $1,700–$2,900/yr (solo) and $3,000–$15,000/yr (1–5 professionals) for professional liability alone. A $249/mo tool that measurably moves them out of the “question mark” bucket is 8–10% of the premium it’s protecting. That’s the same ratio firms already accept for QA/QC tooling. And the buyer is the person who signs both invoices.

ACV: ~$2,500 blended, assuming the customer base skews to the 5–15 band where AI adoption is real but there’s no in-house risk manager.

Rough math to $1M ARR: 400 firms × $2,500 = $1.0M. Against a serviceable pool of 20,000+ firms, that’s a 2% penetration requirement. Comfortable.

Rough math to $5M ARR: 1,600 firms at a $3,100 blended ACV — requires the engineering disciplines (structural/MEP/civil) to convert as well as architecture, plus the broker channel producing at scale, plus a $99/seat-month expansion on larger firms. Roughly 6% penetration of the sub-50 pool. Achievable but it needs the channel to work, not just direct sales.

Expansion path: headcount growth is the natural escalator (firms move tiers). Beyond that: per-project archival retention (E&O tails run 7–10 years — charge for the long hold), multi-entity for firms with several licensed entities, and a broker-facing portal where an agency manages the renewal packs for its whole A/E book.

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

This market has an unusually clean channel, because a third party is already paid to worry about exactly this problem.

  1. Wholesale brokers and A/E specialty agencies (the main channel). A few dozen US agencies specialise in architects & engineers professional liability and each carries a book of 200–2,000 small firms. They are actively publishing content about AI exclusions right now — Risk Specialty Group, A&E-focused agencies and program managers have all put out 2026 renewal guidance. They have a commercial problem: their client is about to become a question mark and get a targeted increase, and the broker takes the blame. Give them a co-branded Renewal Pack they can run across their book for free, take the subscription upsell. Target: sign 5 agencies, each producing 10–20 firms. That’s the first 100 on its own.

  2. Renewal-date cold outreach. State licensing boards publish licensed architecture and engineering firm registries (not just individuals) — Texas TBAE, California, New York, Florida all list firm name, address and principal. Scrape 3,000 firms in the 3–50 band across 5 states. The email is not a pitch, it’s a question: “Does your E&O policy carry endorsement CG 40 47? Here’s how to check in 60 seconds.” Half of them will not know the answer and will go look. Reply-rate on a specific, checkable, scary-if-true question in a professional market runs well ahead of generic SaaS cold email; I’d underwrite 4–6% reply and 1.5% close to the $499 pack, converting a third of those to subscriptions.

  3. AIA component chapters and CE credit. Architects need continuing education hours. AIA local components are perpetually hunting for CE-approved lunch sessions and this topic is genuinely urgent. Build one AIA-approved CE session — “AI Exclusions and Your 2026 E&O Renewal” — and run it through 15–20 component chapters and state engineering societies (NSPE, ACEC). 30–60 principals per room, all of them exactly the buyer, all of them getting the scare and the fix in the same hour. This is the highest-converting channel in the mix and it costs travel money.

  4. The endorsement-check lead magnet. Free tool: upload your policy PDF, get back a plain-English answer on whether an AI exclusion is attached and to which coverage part. No signup for the answer, signup for the archive. This is a genuinely useful thing that costs cents to run, and it self-qualifies — everyone who uploads a policy is in-market, and the ones who come back with an exclusion attached are the hottest leads in the funnel.

  5. Coverage-counsel referral. Construction-defect and coverage attorneys are the ones who will actually litigate the first CG 40 47 denial. They have a strong professional interest in their A/E clients keeping contemporaneous records. Warm-intro loop with 10–15 firms in the construction bar.

10. Build complexity — justification

Medium. The core — record capture, hash-sealing, attestation workflow, PDF generation — is standard web stack and takes weeks, not months. The complexity sits in two places: the tool-inventory integrations (M365 Graph, Google Workspace admin, Autodesk Platform Services all have per-user licence and activity APIs, but three separate auth flows and permission models), and the deliverable classification, which needs a real vision-plus-metadata pipeline and careful tuning to avoid confidently mislabelling a sheet set. Policy-PDF parsing across inconsistent carrier wordings is the third chunk of genuine work.

Two people, 12–16 weeks to a credible v1. The Renewal Pack alone — inventory sweep + policy parse + generated PDF — is shippable in 6 weeks and is independently sellable at $499, which means revenue well before the full product exists. That sequencing is what makes this fundable out of pocket.

11. Gating checklist

GatePass?Note
Legal in target market✅Record-keeping and attestation tooling. Not insurance advice, not legal advice, and the product must be explicit about that line — it produces the firm’s own records, it doesn’t opine on coverage.
Ethical — no harm / dark patterns✅The product helps firms tell the truth accurately. The one ethical tripwire is that it must never help a firm misrepresent AI use to an underwriter — records are hash-sealed at creation precisely so they can’t be retro-edited. That constraint is also the moat.
Market exists (evidence above)✅ISO forms live since Jan 2026, multiple carriers excluding, ~17,800 sub-50-staff firms, underwriters asking the question on renewal applications.
1–5 person team can build this✅Two people, 12–16 weeks.
Launchable with <$50K / ₹40L✅~$15–25K: two founders’ time, API costs, CE course accreditation, travel to chapter events.

All five pass.

12. Feasibility score

AxisWeightScoreNotes
Problem intensity2017/20Hair-on-fire at renewal, which every firm hits annually. Loses points because between renewals it can drop to background noise — the pain is sharp but periodic rather than daily. The claim-denial scenario is severe but rare and distant; the renewal pricing hit is the real recurring driver.
Demand evidence1512/15Strong structural evidence: named ISO forms, named carriers filing, dated effective date, AIA adoption stats, broker advisories explicitly pricing undocumented firms higher. Docked 3 because I could not source verbatim complaints from A/E principals themselves — the signal is all intermediary (brokers, attorneys, trade press), not raw customer voice. That’s the gap to close in week one.
Build feasibility1511/15Standard stack for the record layer, but three OAuth integrations plus a vision classification pipeline plus policy parsing puts it past the “solo in 6 weeks” band. 12–16 weeks for a pair.
Distribution clarity1512/15Broker channel and AIA CE are both named, reachable and already primed by the topic. Firm registries are public and scrapable. Docked because the broker channel is a partnership motion with real conversion uncertainty — agencies are slow, and I’m assuming they’ll co-brand rather than build.
Revenue mechanics1512/15Pricing is a small fraction of a premium the buyer already pays, 400 customers for $1M is 2% penetration, and the $499 pack de-risks the funnel. Docked because renewal-driven products have a churn question — does a firm keep paying in month 13 once the renewal is behind them?
Time to first revenue107/10The $499 Renewal Pack can sell from week 6–8. Full subscription revenue lags. Not a same-week close, but not a long enterprise cycle either.
Defensibility105/10Honest assessment: copyable. The record format isn’t proprietary and a competitor could ship a similar tool in a quarter. What accrues is the sealed archive itself — a firm with three years of contemporaneous sealed records cannot migrate that history to a competitor without losing the contemporaneity that gives it evidentiary value. That’s real lock-in but it takes years to matter. Broker relationships are the faster moat.
Total10076/100

13. Qualitative modifiers

Founder-fit tags

technical-heavy · sales-heavy

Technical because the classification pipeline and the integrations are the product. Sales-heavy because the broker channel and the AIA chapter circuit are both relationship motions — this does not sell itself through a signup form. A domain advisor from the A/E professional liability world is close to mandatory; a founder who has never sat in a renewal meeting will get the artifact subtly wrong.

Key assumptions to validate (3–5)

  1. Assumption: A/E principals feel enough pain at renewal to pay, rather than shrugging and accepting the increase. How to test: 25 calls with principals at 5–30 staff firms who renewed in the last 90 days. Ask what the AI section of their application asked and what they answered. If more than half say “there wasn’t one” or “I don’t remember,” the timing is early.
  2. Assumption: Brokers will co-brand and distribute rather than ignore it or build it. How to test: pitch 8 A/E specialty agencies with a free Renewal Pack for 10 of their clients. Two signed pilots = channel is real.
  3. Assumption: The sealed record is actually what a carrier or coverage counsel would find persuasive. How to test: 5 conversations with construction/coverage attorneys and, if reachable, one A/E underwriter. Show the artifact, ask what they’d attack. This assumption is load-bearing for the whole thesis — if the answer is “we’d never look at that,” the product is theatre.
  4. Assumption: Firms will actually complete the attestation step per deliverable rather than abandoning it. How to test: 10-firm pilot, measure sealed-deliverable rate in weeks 4–8, not week 1. Below 60% of issued deliverables and the workflow is too heavy.
  5. Assumption: Tool inventory can be read accurately enough via admin APIs to be credible. How to test: run the sweep on 5 friendly firms and have their IT confirm the output. Errors here destroy trust instantly.

Risk flags

  1. Market timing: The single biggest risk, in both directions. Adoption of AI exclusions is accelerating but not universal — one advisory notes only “a handful so far carry absolute AI exclusions.” If exclusions stall at 20% market penetration, urgency stays low. If the market instead settles on affirmative AI endorsements that restore coverage for disclosed activities, the product pivots (still needs the disclosure evidence) but the fear-based pitch weakens.
  2. Regulatory/standards risk in reverse: if AIA, ACEC or a major carrier publishes a free standard AI-use log template, the differentiated artifact becomes a commodity form overnight. Mitigation is to be the one who offers that template and owns the sealing layer underneath it.
  3. Platform dependency: the tool-inventory sweep depends on M365 Graph, Google Workspace and Autodesk APIs continuing to expose per-user AI feature usage. Autodesk in particular could restrict this. The product degrades to manual declaration without it — still usable, meaningfully weaker.
  4. Evidentiary uncertainty: there is no case law yet on CG 40 47. Nobody knows what a court will accept as proof that a deliverable was not “attributable to” generative AI. The product is a bet on a reasonable answer to an untested question, and the first coverage decisions (2027–2029) could reshape or invalidate the format.
  5. Churn: renewal-triggered buying can mean renewal-triggered cancelling. If firms buy in the month before renewal and churn the month after, ACV collapses. The contemporaneity argument is the retention lever — a gap in the record is worse than no record — but it has to be taught, not assumed.

14. Structured verdict

Score:                  76/100
Verdict:                GO
Confidence:             Medium
Best-fit builder:       Technical pair, one of whom can run a broker/chapter sales motion,
                        with a paid A/E professional-liability advisor on retainer
Time to revenue:        6–8 weeks to first $499 Renewal Pack; 12–16 weeks to subscription
Capital to launch:      $15–25K
Top 3 assumptions to validate first:
  1. Renewal pain is real — 25 principal calls about what their 2026 application asked
  2. Coverage counsel and an underwriter agree the sealed record is persuasive evidence
  3. Two A/E specialty agencies sign co-branded pilots across their books
Kill criteria:
  - Abandon if fewer than 8 of 25 principals surveyed recall an AI question on their
    2026 renewal application — the market isn't feeling it yet, revisit in 12 months
  - Abandon if coverage counsel consensus is that contemporaneous firm-generated records
    would not shift a CG 40 47 coverage dispute — the core artifact is then worthless
  - Abandon if zero of 8 pitched A/E agencies will pilot; direct-only CAC won't support
    a $2,500 ACV at the volume required
  - Abandon if pilot firms seal under 40% of issued deliverables by week 8

15. Next step — 1-week validation sprint

  • Day 1–2: Pull firm registries from Texas TBAE and California, filter to 3–50 staff A/E firms. Book 25 principal calls. One question set: what did your 2026 E&O application ask about AI, what did you answer, and could you back that answer with records? Log verbatim responses — I need the customer voice this proposal is currently missing.
  • Day 3: Talk to 5 construction/coverage attorneys and, if I can get one, an A/E underwriter. Show a mocked-up sealed record for a single sheet set. One question: would this change how you argue a CG 40 47 denial? Push them to attack it.
  • Day 4: Pitch 8 A/E specialty brokers on a free co-branded Renewal Pack for 10 clients each. Measure who takes the meeting and who says yes.
  • Day 5: Decide.

Falsifiable go/no-go: proceed only if ≥12 of 25 principals confirm an AI question appeared on their 2026 renewal application and cannot back their answer with records, ≥3 of 5 attorneys say a contemporaneous sealed record would materially help a coverage dispute, and ≥2 of 8 brokers commit to a pilot. Miss any one of those three and this is a VALIDATE that waits for the 2027 renewal cycle, not a build.

Interested in a detailed proposal?

Get a deep-dive with market research, competitive analysis, and implementation roadmap.

Contact us

info@startupbasket.ai