RevTripwire

1. One-liner

Fires the moment a drawing revision lands on work your crew already installed, while the claim is still alive.

2. Trend signal — why now?

Three things moved in the last twelve months, and they moved in the same direction.

Outdated drawings are the single biggest documented cause of rework. Per CII’s analysis, 31% of documentation-related rework traces to outdated drawings or specifications in the field. Rework overall runs 4–12% of project value; the 2024 Autodesk/FMI report puts the ceiling at 20% of total project cost. FMI pegs the annual US bleed from rework, data-hunting and communication breakdown at $177 billion. This is not a soft problem. It is the most expensive recurring failure in the industry and it has a named root cause.

The volume of revisions has outrun human tracking. From a UK document-control practitioner: “On a typical commercial fit-out, it’s not unusual for a set of 200 drawings to generate 600+ revisions over the course of the project.” And the distribution mechanism is still email: “The architect issues 15 revised drawings by email to 8 recipients. That’s 120 individual files scattered across 8 inboxes.” Nobody reads 600 revisions. They skim, and they miss.

Long-context inference stopped being the bottleneck in 2026. Thirteen models now ship 1M+ token windows, and Anthropic dropped the long-context surcharge that used to double input pricing above 200K tokens. Filling a full 1M window costs $0.14 on DeepSeek V4 Flash. Holding an entire drawing set plus a revision history plus a sub’s installed-work log in one comparison pass went from prohibitively expensive to a rounding error inside eighteen months. That is the unlock.

Money is following it. LightTable raised a $22M Series A in 2026 for an AI preconstruction platform that reads drawings to flag design errors. iFieldSmart and Beam AI both ship revision-comparison-by-discipline. Structured AI is building QC agents for technical drawings. The category is validated and funded — but every one of those products points at the general contractor, upstream of the moment where the money is actually lost.

Provenance:

• Signal 1 (Demand): 31% of documentation-related rework comes from outdated drawings or specs in the field; rework is 4–12% of project cost; £18,000 rip-out documented from a crew building Rev B when Rev D had been issued three weeks earlier — https://www.constructionai.io/blog/managing-drawing-revisions-on-site and https://reworkcost.com/cost-of-rework-in-construction — 2026-08-13
• Signal 2 (Feasibility): 1M+ token context windows across 13 models and Anthropic’s removal of the long-context surcharge in 2026 make whole-drawing-set-plus-history comparison economically trivial ($0.14 per full 1M window on DeepSeek V4 Flash) — https://www.morphllm.com/llm-context-window-comparison — 2026-08-13
• Signal 3 (Economic): LightTable $22M Series A for AI drawing review; iFieldSmart, Beam AI, Drawer AI, Structured AI all shipping revision/change intelligence — all aimed at GCs and preconstruction, none at the installing sub — https://bricks-bytes.com/funding-ma/latest-construction-technology-funding-rounds-3rd-aug-2026/ and https://www.ifieldsmart.ai/skills/change-management-ai/ — 2026-08-13 Category: Tech-unlock

3. The opportunity

Every existing revision tool answers “what changed?” That is now a commodity — Bluebeam has done PDF overlay compare for a decade at $260/user/year, and the AI tools do it by discipline.

Nobody answers the question that actually decides who eats the cost: “did it change after we already built it?”

That distinction is worth real money, and it is contractual, not cosmetic. The industry rule is explicit: an ASI is only valid before the relevant work has begun. If the work is already underway when the revision lands, the change requires a change order or construction change directive instead — the ASI must be rescinded and a change order issued to add the cost and time. But the burden of raising this falls entirely on the sub, and there is a fuse on it: most contracts require written notice within 7–14 days of identifying a potential change, and missing that deadline invalidates the claim.

So the same physical event — architect moves a duct penetration — has two completely different financial outcomes:

• Sub notices within the notice window, with dated proof the work was already in: GC or owner pays for the rip-out via change order.
• Sub notices late, or just quietly complies: the sub eats it. Worse — if the subcontractor begins to enact these changes, they have given up their right to ask for further compensation or time. Complying is itself the waiver.

The gap between those two outcomes is decided by whether a busy PM juggling five jobs happened to open the right PDF in the right week. That’s the arbitrage. The incumbents are all standing upstream of it: I checked iFieldSmart’s change-management product directly — it targets “General Contractors, VDC/BIM Teams, Estimators, Preconstruction Teams,” it does not track whether work was already installed, it does not generate notice letters or change order claims, and it’s currently a waitlist.

The GC has no incentive to build this. It is a tool for extracting money from the GC. That’s why it doesn’t exist yet.

4. Target market

• Primary customer: Owner or head of project management at a US specialty trade subcontractor (NAICS 238) — electrical, mechanical, plumbing, fire protection — doing $5M–$50M annual revenue, running 5–25 concurrent commercial projects, with 2–6 project managers. Not residential. Not the GC.
• Why they buy: Because they are already losing these arguments and they know it. Contractors absorb rework cost when the fault is theirs, but “if the cause of the rework is on the owner’s side, such as design errors, owner-directed changes, or differing site conditions, contractors can typically recover rework costs through change orders or claims.” The recovery path exists. Subs miss it because they find out too late and can’t prove installation date. One documented example: a partition built to Rev B when Rev D had been out three weeks — “Cost to rip out and rebuild? £18,000. Time lost? A week. Root cause? A PDF in someone’s email inbox that nobody knew about.”
• Rough TAM reasoning: ~489,000 specialty trade contractor businesses in the US as of 2020 Census data; NAICS 238 is roughly 60% of all construction-sector establishments. The addressable slice — commercial subs large enough to run formal drawing sets and dedicated PMs, but too small for enterprise document control — is realistically 20,000–40,000 firms. At $600–1,500/month that is a $150M–$700M ceiling. Far more than I need for a $5M ARR business.
• Why now for them: Change orders run 5–10% of contract value on commercial projects, individual orders reaching six figures for design errors found mid-construction. Meanwhile subcontractor markups on change work are frequently capped at 5% by the GC, so subs cannot make margin back on volume — they have to win the entitlement argument instead. And in California, SB 440 took effect January 1, 2026, creating a mandatory statutory process with enforceable deadlines and remedies for change-order payment disputes on private projects. A statutory deadline regime rewards exactly the sub who can produce a dated, evidenced notice fast. The subs who can’t will lose more than they used to.

5. Product sketch (MVP)

• Watches the drawing source you already use — connects to the GC’s Procore/ACC/SharePoint plan room, or just an email address you forward issuances to. No new place to check.
• Detects what changed, scoped to your trade only — not 600 revisions, the 40 that touch your work. Everything else is suppressed.
• Crosses each change against your installed-work log — pulls dated field evidence (daily reports, timestamped/GPS photos, foreman check-offs) and answers the only question that matters: was this area already built when the revision dropped?
• Fires a tripwire alert on collision — push to the PM the same day a revision lands on installed work, with the revision, the prior rev, and the field photo proving install date side by side.
• Drafts the notice letter inside the contract’s window — pre-filled with the contract’s own notice clause, the revision ID and issue date, the installation date, and the affected area. PM reviews, signs, sends. This is what converts a detection into money.
• Runs a live claim register — every open collision, notice sent date, days remaining on the notice clock, GC response, resolution value.
• Ships a dispute packet on demand — one PDF per claim with the full evidence chain, formatted for the GC’s project executive or, if it goes there, counsel.
• Weekly digest of near-misses — revisions landing on work scheduled to start within 10 days, so the crew can stop before installing the wrong thing at all.

6. AI angle — what’s load-bearing

Remove the AI and this product cannot exist. Three jobs are genuinely model work:

Trade-scoped semantic diff. Comparing Rev D to Rev B as pixels is a solved 2010 problem and produces noise — a retitled sheet or shifted north arrow lights up the overlay. What matters is whether the electrical rough-in in Grid C4-C7 moved. That requires reading the drawing as a technical document, understanding which discipline and which physical scope each change touches, and discarding the rest. The 2026 long-context economics are what let a single pass hold the full set, the revision history and the trade scope simultaneously.

Reconciling messy field evidence against drawing geometry. Field records are a foreman’s phrase — “pulled feeders east wall, second floor” — plus a photo with GPS and timestamp. Mapping that to a grid reference on a revised sheet, confidently enough to assert an installation date, is fuzzy natural-language-to-spatial matching. No rules engine does this.

Reading the subcontract to find its own notice clause. Every contract buries its notice window somewhere different — 7 days, 14, “promptly,” or the AIA/ConsensusDocs boilerplate. The product ingests the executed subcontract, extracts the governing clause, and sets the clock and the letter language from it. This is the piece that makes the output legally usable rather than merely informative.

The honest limit: the model proposes, the PM disposes. Every notice goes out under human review. A false collision costs the sub credibility with the GC, so the product is tuned to under-fire on ambiguity and flag uncertainty explicitly rather than auto-send.

7. Localization angle

N/A — this is a US-first play. The wedge is specifically the American subcontract notice-clause regime (AIA/ConsensusDocs written-notice requirements, ASI mechanics, and California SB 440’s new statutory process). That legal specificity is the moat; genericizing it across geographies would dissolve it. UK/Ireland is the obvious second market — same document-control failure mode, and the £18,000 example above is British — but it needs a separate contract-clause corpus (JCT/NEC), which is a v2 problem, not a launch problem.

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

• Pricing: $800/month base for up to 10 active projects, $60/project/month beyond that. Annual prepay discount 15%. No per-seat pricing — seat counts punish the exact behaviour I want (everyone in the office watching the tripwire).
• ACV: ~$11,500 blended, assuming most customers run 10–20 projects.
• Rough math to $1M ARR: 87 customers at $11.5K. For a market of 20,000+ qualified firms that is 0.4% penetration. Achievable.
• Rough math to $5M ARR: ~430 customers, or ~250 customers plus a contingency tier (see below) on recovered claims. Requires moving beyond electrical/mechanical into fire protection, drywall and glazing, plus a second GC-platform integration.
• Expansion path: ACV grows three ways — project count as the customer wins bigger work; additional trades within multi-trade contractors; and the real upside, a contingency tier at 8–12% of recovered claim value for firms that would rather pay on success than subscribe. Lease-audit firms already run this model at 3–9 months per engagement, so the construction buyer understands contingency pricing. I’d hold contingency back until the detection accuracy is proven, because it puts the product’s revenue directly on the line for false positives.

The unit economics are comfortable. Inference is the main variable cost and it’s now trivial — a full drawing set comparison is cents, and even at hundreds of comparisons per customer per month, COGS stays under 5% of revenue. This is a >90% gross margin business.

The ROI pitch writes itself and it is not a stretch: one recovered $18,000 rip-out pays for eighteen months of subscription. A single avoided rework on a single project justifies the year.

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

• Run the free retro-audit as the entire top of funnel. Ask a prospect for one finished project’s drawing set and daily reports. Run the collision detection backwards and hand them a report: “here are the four revisions that landed on work you’d already installed, here’s what you didn’t claim, here’s roughly what it cost you.” This is not a demo, it’s a bill for money they already lost. It is the highest-conviction sales artifact I can imagine in this category, and it costs me inference pennies to produce. Target: 30 audits, expect 8–12 closes.
• NECA’s 119 local chapters, worked one at a time. Chapters publish member directories, are independently staffed and run their own programming. The pitch to a chapter manager is a 20-minute lunch-and-learn titled “The change orders you’re not filing” — with the retro-audit as the live segment. Ten chapters, ~10 qualified firms each engaged, at a 15% close rate is 15 customers. Repeat with MCAA for mechanical. This is a warm, gatekept channel where a directory and a credible talk beat any ad spend.
• Target the PMs, not the owners, on LinkedIn. Project managers at $5M–$50M specialty subs are a precisely filterable LinkedIn audience by title + company size + NAICS. They feel this pain personally — it is their job that gets blamed when the crew builds to Rev B. Cold outreach with a one-minute screen recording of a real (anonymized) collision detection. Expect 3–5% reply on a well-cut video, 500 targeted messages a month.
• Construction claims consultants and construction attorneys as referral partners. These people are already paid to fight change-order disputes after they’ve gone bad. RevTripwire makes their clients’ cases stronger and their engagements cleaner. Revenue share or straight referral. Twenty partners each sending two clients a year is 40 customers with near-zero CAC.
• Publish the collision data as content. Once there are 50 customers, an annual “how many revisions hit installed work” benchmark report by trade is a genuinely novel dataset that nobody else can produce. That is press and inbound, not blog filler.

10. Build complexity — justification

Medium. The AI work rides on off-the-shelf long-context models — no custom training, no proprietary corpus needed at launch. The genuine effort is in the integrations (Procore and Autodesk Construction Cloud APIs, plus an email-ingest fallback), the drawing-parsing pipeline, and the field-evidence reconciliation, which is where the accuracy risk concentrates. Two people, 4–5 months to a credible v1 that handles one trade (electrical) on commercial projects. The email-forwarding ingest path is what keeps this out of High complexity — it means v1 doesn’t require any GC to grant platform access, which would otherwise be a brutal chicken-and-egg.

11. Gating checklist

Gate	Pass?	Note
Legal in target market	✅	Helping a party document and assert its own contractual rights. Product drafts, human sends. Careful positioning as documentation software, not legal advice.
Ethical — no harm / dark patterns	✅	Restores an entitlement the contract already grants to the weaker party. Under-fires on ambiguity by design.
Market exists (evidence above)	✅	$177B annual rework, 31% traceable to outdated drawings, funded competitors adjacent, existing paid tools at $260–440/user/yr.
1–5 person team can build this	✅	Two people, 4–5 months.
Launchable with <$50K / ₹40L	✅	Inference is cents per run. Cost is time, not capital.

12. Feasibility score

Axis	Weight	Score	Notes
Problem intensity	20	17/20	Documented, expensive, recurring, with a hard contractual deadline attached. Loses subs real money on most commercial jobs. Not 19–20 because subs have survived on tolerating it — it’s a bleed, not a fire, and the pain is diffuse across projects rather than concentrated in one dreaded moment.
Demand evidence	15	12/15	Strong indirect evidence: hard rework statistics, funded adjacent competitors, existing paid tools in the category, verbatim practitioner accounts. Docked because I could not source subs specifically asking for installed-work collision detection — the demand is inferred from the loss, not from a stated want. Reddit/forum access was blocked during research.
Build feasibility	15	10/15	Off-the-shelf models, but real integration surface and a genuinely hard accuracy problem in field-evidence-to-drawing reconciliation. 4–5 months for a pair, not 6 weeks.
Distribution clarity	15	12/15	The retro-audit is an unusually strong wedge and NECA’s 119 chapters are a named, gatekept, warm channel. Docked because chapter-by-chapter is slow and PM-level outreach doesn’t always reach the budget holder.
Revenue mechanics	15	12/15	Pricing sits comfortably between Contractor Foreman ($49) and Procore (enterprise), ROI is one claim, margins >90%. Docked because $800/mo is above what many mid-size subs currently spend on any single tool, and the contingency tier is unproven here.
Time to first revenue	10	7/10	Retro-audits can pre-sell during the build, but v1 needs 4–5 months before a paying customer gets real value. Revenue in ~6–8 weeks post-launch.
Defensibility	10	5/10	Honest score. The detection is copyable by any of the funded incumbents in a quarter if they decide subs are worth serving. The durable moats are workflow lock-in (the claim register becomes the sub’s system of record), accumulated contract-clause knowledge, and — most importantly — the structural fact that GC-facing incumbents are conflicted about building a tool that extracts money from GCs. That’s a real but not permanent shield.
Total	100	75/100

13. Qualitative modifiers

Founder-fit tags

technical-heavy · domain-expertise-required

This needs someone who can build a real document pipeline and someone who has sat in a change-order fight. Without the domain half, the notice letters will be wrong and the first GC pushback will destroy credibility. Find a former specialty-contractor PM as a co-founder or first hire — not an advisor.

Key assumptions to validate (5)

1. Assumption: Collisions between revisions and already-installed work happen often enough to matter — at least 2–4 per project on typical commercial work. How to test: Run the retro-audit manually on 10 completed projects from 3 friendly subs. Count actual collisions. This is the single most important number in the business and it is measurable before writing any code.
2. Assumption: Subs will hand over drawing sets and daily reports to a stranger for a free audit. How to test: Ask 20. Track how many actually send files versus say yes and ghost. Anything under 40% means the funnel is broken at step one.
3. Assumption: Field evidence is good enough to establish installation dates. Many subs keep sloppy daily reports. How to test: Inspect the daily-report and photo quality in the 10 retro-audit projects. If half can’t support a date claim, the product needs to also drive field capture — a much bigger build.
4. Assumption: Subs will pay $800/mo rather than treat this as a nice-to-have. How to test: Present the retro-audit findings with a price. Measure close rate, not enthusiasm. Also test the contingency tier head-to-head.
5. Assumption: GCs won’t retaliate against subs who systematically file more notices. How to test: Interview 10 subs on whether aggressive claim behaviour has cost them invitations to bid. If it has, positioning shifts hard toward “prevention” (the near-miss digest) rather than “recovery.”

Risk flags

1. Relationship risk — the real one. Subs live on repeat invitations from a handful of GCs. A tool that mechanizes claims against those GCs could be seen as a relationship grenade. Mitigation is to lead with the near-miss digest — stopping the rework before it’s built helps everyone and makes the sub look sharp — and treat recovery as the fallback. This risk is the biggest threat to the thesis and assumption 5 tests it directly.
2. Platform dependency. Procore and ACC control the drawing pipe. Either could restrict API access or ship this natively. The email-ingest fallback is the hedge and should stay a first-class path, never a degraded one.
3. Accuracy risk with asymmetric cost. A false collision sent to a GC costs credibility disproportionately — one bad notice can poison the tool’s reputation inside a firm. Under-fire on ambiguity, always human-in-the-loop, and never auto-send.
4. Competitive encroachment. LightTable ($22M raised), iFieldSmart and Beam AI are one product decision away from this. The window is the 12–18 months while they chase GC and preconstruction budgets, which are larger and more obvious. Speed matters more than polish here.
5. Evidence-quality dependency. The product is only as good as the sub’s field records. Customers with weak documentation get weak results and will churn blaming the tool. Qualify hard on daily-report discipline during onboarding, and be willing to turn away firms that don’t have it.

14. Structured verdict

Score:                  75/100
Verdict:                GO
Confidence:             Medium
Best-fit builder:       Technical founder paired with a former specialty-contractor PM
                        who has personally fought a change-order dispute
Time to revenue:        4-5 months to v1, first paying customer ~6-8 weeks after launch;
                        retro-audits can pre-sell during the build
Capital to launch:      $15-25K (two founders' time, inference, integration costs)
Top 3 assumptions to validate first:
  1. Collision frequency — manually retro-audit 10 completed projects across 3 subs,
     count revisions that landed on already-installed work. Need 2+ per project.
  2. Field evidence sufficiency — inspect those same 10 projects' daily reports and
     photos. Can installation dates actually be established?
  3. GC retaliation risk — interview 10 subs on whether systematic claim-filing has
     ever cost them bid invitations.
Kill criteria:
  - Abandon if retro-audits across 10 real projects surface fewer than 1.5 collisions
    per project on average — the pain is too rare to sell against.
  - Abandon if fewer than 3 of 20 subs approached will actually send project files for
    a free audit — the wedge doesn't open.
  - Abandon if >50% of surveyed subs report that filing more notices has damaged or
    would damage GC relationships — the product is unsellable regardless of accuracy.
  - Abandon if a funded incumbent ships sub-side installed-work collision detection
    before v1 launches.

15. Next step — 1-week validation sprint

• Day 1–2: Recruit 3 specialty subs (electrical or mechanical, $5M–$50M revenue) through personal network or a NECA chapter manager. Ask each for 3–4 completed commercial projects: full drawing set with revision history, daily reports, field photos. The ask is easy because the offer is “I’ll tell you what you left on the table.”
• Day 3–4: Manually run the collision analysis on those ~10 projects. No product — long-context model plus a spreadsheet. For each project count: total revisions, revisions touching that trade, revisions landing on already-installed work, and whether a change order was ever filed for it. Estimate the unclaimed dollar value.
• Day 5: Present findings back to each sub with a price on the table. Ask for a signed LOI or a deposit — not “would you use this.”

Falsifiable outcome: Go if the 10 projects average ≥2 collisions per project with a credible unclaimed value ≥$10K per project, AND at least 1 of the 3 subs signs an LOI or pays a deposit at $800/mo. Anything less and the loss is too small or too invisible to build a company on — and I’d rather learn that in week one for the price of a few hours of inference than in month six.