---
title: "A shelf of tools is not a foundation — AI point solutions and roofing"
description: "Why single-task AI tools for roofers are a dead end for builders, buyers and investors. An interactive essay with 24 graded claims and sourced Census data."
url: "https://foundationproject.net/research/ai-point-solutions-roofing"
author: "Foundation Projects"
published: "2026-10-17"
modified: "2026-10-17"
language: "en-US"
image: "https://foundationproject.net/images/research/a-shelf-is-not-a-foundation-share.jpg"
keywords: ["AI point solutions","vertical AI","roofing software","roofing contractors","AI assistants","Model Context Protocol","software moats","venture capital","NAICS 238160"]
---

AI, roofing, and who owns the outcome

# A shelf of tools is not a *foundation.*

AI point solutions for roofers are a dead end: for the software companies that build them, the roofers who buy them, and the investors who fund them.

A research essay from Foundation Projects · October 2026

The short version

## A good way in. A bad place to *stay.*

AI point solutions for roofers look like a business and behave like a feature. They are a dead end for three different people, for three different reasons.

**The software company**

**No moat.** A wedge earns an introduction. The general-purpose assistants keep absorbing the features point solutions were built on, and what a tool can charge is a sliver of the job it touches.

**The roofer**

**No foundation.** A tool for every step is not a system. Each one keeps its own copy of the customer and hands the integration work back to the owner.

**The investor**

**A low ceiling.** A seat-based wedge tops out at what a company spends on software. The capital that agrees with this paper is buying and running service businesses, not funding another seat.

What lasts is owning the outcome: the workflows, the data they leave behind, and the accountability for the result, with people handling what the system escalates.

How sure are we? The argument started from a video essay. We turned it into 24 checkable claims and graded each one against primary sources. 4 held as stated, 12 held in part, 8 could not be sourced, and none was contradicted outright. This paper builds on the first two groups only. The ledger is in the [methods](https://foundationproject.net/research/ai-point-solutions-roofing#methods), and you can download it.

**Our stake.** Foundation Projects is building a founder-aligned platform for roofing companies, so we are not neutral on this question. That is why the paper shows its sources, grades its own claims, and marks every assumption.

## The wedge was a *good idea.*

For three years a common AI startup pitch has run on one move. Find one expensive, well-defined workflow. Add AI. Sell the return. While models could not do much alone, the product was everything built around them, and a narrow tool with a clear payback was an honest way into an industry.

Roofing got its version. In our own experience, the wedges pitched to us most often are an AI voice receptionist, an insurance supplement writer and an instant-quote tool. They are real tasks with real money behind them, and some of them work.

Twenty identical workflow cards stand for twenty single-task tools. One expensive workflow is picked, AI is added, and the card is sold on its return. Three roofing examples are marked: an AI voice receptionist, an insurance supplement writer, and an instant-quote tool.

In September 2023 [Sequoia described the first wave](https://www.sequoiacap.com/article/generative-ai-act-two/): every company was an “AI copilot,” its inbox was full of “undifferentiated pitches,” and user retention was “terrible.” It called what should come next an Act Two of applications built from the customer back and end to end, and it named Harvey, the legal AI company, as an early example.

That is the part of the story the pitch decks keep. The part they skip is what happens to a wedge once the thing it was built around stops being scarce.

## What the assistants *took in.*

Every point solution rests on a capability: read a document, answer a call, connect to a system. Those capabilities keep moving into the general-purpose assistants. The dates are public. Five of the six entries below come from the companies’ own pages, and the sixth is an outside essay.

1. November 2023

   ### GPTs and file search

   Custom assistants from instructions, knowledge and skills. No code.

   > You can also attach files to let ChatGPT search PDFs and other document types.

   > OpenAI, introducing GPTs

   Built in

2. November 2024

   ### Model Context Protocol

   One open protocol to connect assistants to business systems.

   > Every new data source requires its own custom implementation.

   > Anthropic, on the problem MCP was built to solve

   Standardized

3. October 2025

   ### Agent Skills

   Folders of instructions, scripts and resources the assistant loads when relevant. Shipped for Excel, PowerPoint, Word and fillable PDFs.

   > Skills teaches Claude how to work with Box content.

   > Box, quoted in Anthropic’s announcement

   Packaged

4. By October 2026

   ### Plugins, inside the assistant

   A page of business tools, from finance and sales to customer support and small business, to connect to ChatGPT and Codex.

   > Connect the apps you love and the data you rely on right into ChatGPT with plugins.

   > OpenAI, on its plugins page

   Inside

5. March 2026

   ### Open frontier models

   A coalition builds open models that developers and organizations can specialize for their industries and domains.

   > Nemotron models, which have been downloaded more than 45 million times from Hugging Face.

   > NVIDIA, on its open models

   Open

6. April 2026

   ### The question

   An essay on Harvey, a leading vertical AI company in law, voices the objection that a general model can already do the work, and calls it “a serious argument.”

   > Why pay for Harvey when the general model can already do the work?

   > Objection voiced in an essay from Harvard Law School’s Center on the Legal Profession

   Asked

A timeline of six entries from 2023 to 2026, five from company pages and one from an outside essay. Four show a capability that point solutions sold moving into the general-purpose assistants, one shows open models arriving, and one asks what the specialist is for.

Read in order, the timeline does not say every tool disappears. It says the thing a tool was first valued for stops being scarce. File search over PDFs arrived inside the assistant in 2023. In 2024 a shared protocol was launched to replace one-off connectors. In 2025 Anthropic packaged skills for spreadsheets, slides, documents and forms. By October 2026 OpenAI’s plugins page was inviting business users to connect their apps “right into ChatGPT,” so the shelf is moving inside the assistant. And an essay on Harvey, a leading vertical AI company in law, from Harvard Law School’s Center on the Legal Profession sets out the objection that the general model can already do the work. Its authors call it “a serious argument” and say they “genuinely don’t know the answer” to whether vertical legal AI can hold its position.

Apply the timeline to the three roofing wedges and the pattern is specific.

**The voice receptionist.** In August 2025 [OpenAI’s realtime voice model added phone calling over SIP](https://openai.com/index/introducing-gpt-realtime/). Its [guide](https://developers.openai.com/api/docs/guides/voice-sip?voice-api=realtime) now says, “With SIP and the Realtime API you can direct incoming phone calls to the API.” In September 2026 its newest voice model went [generally available](https://developers.openai.com/api/docs/changelog) at $0.05 a minute for the voice session, billed per second, with model and tool usage charged separately. The platform supplies the pipe. It leaves the rest to the builder: “Your backend still owns the incoming-call decision, session configuration, authorization, and business logic.”

**The supplement writer.** Reading a carrier’s estimate against a contractor’s and drafting a response is document work. Anthropic’s training site does not describe a claims product, but it does tell users to [“Upload any set of competing proposals—software vendors, insurance quotes, contractor bids—and get clear analysis to inform your decision.”](https://academy.claude.com/use-cases/compare-and-analyze-competing-options) What we did not find in a general assistant is the price data. Verisk’s [license terms](https://www.verisk.com/privacy-policies/xactware-eula/) say, “We own all Price Data provided with the Services.” In May 2026 it announced [connectors](https://www.verisk.com/company/newsroom/verisk-brings-its-trusted-analytics-and-generative-ai-capabilities-directly-into-anthropics-claude/) that bring its analytics into Claude. Its [XactRestore connector page](https://www.verisk.com/resources/verisk-xactrestore-connector-for-claude/) lists roofing contractors among its users and offers to have Claude “propose structured roofing line items and quantities aligned with Verisk price lists and best practices.” It says you need “an active XactRestore or XactRemodel subscription and XactAI enabled for your organization.”

**The instant quote.** Google’s [Solar API returns the dimensions of a building](https://developers.google.com/maps/documentation/solar/building-insights), but its [terms](https://cloud.google.com/maps-platform/terms/maps-service-terms) list energy-system uses, not roofing quotes. Aerial imagery is a different matter. In a disclosure about the data it trains its AI on, EagleView states that [“All aerial imagery is collected and owned by Eagleview,”](https://www.eagleview.com/california-data-disclosure) and in April 2026 it announced its [own agentic AI product](https://www.globenewswire.com/news-release/2026/04/21/3277997/0/en/eagleview-launches-eagleview-horizon-the-agentic-ai-engine-powered-by-25-years-of-verified-property-intelligence.html), due by invitation from June.

Put the three together and a roofing point solution sits between two owners. Above it, the assistants ship the capability. Below it, the data owners keep their data under their own terms and ship their own AI. In between sits the point solution: a workflow and a screen. Among the vendor pages we opened, none shows OpenAI, Anthropic or Google selling a roofing receptionist, supplement writer or quote tool. That absence is the gap point solutions fill. In our view it is a gap in packaging, not in capability, and packaging is the easiest part to copy.

The model layer is moving too. In March 2026 [NVIDIA launched a coalition](https://investor.nvidia.com/news/press-release-details/2026/NVIDIA-Launches-Nemotron-Coalition-of-Leading-Global-AI-Labs-to-Advance-Open-Frontier-Models/default.aspx) to build open frontier models that developers and organizations can specialize for their industries and domains. That is one more reason not to build a business that exists only because the rented model is weak.

## The vertical was a *label.*

Harvey is the test case. It picked a vertical, law, and a first customer, law firms, and it is one of the examples Sequoia named in 2023. When its founders started, they [made “the explicit decision to focus our attention on law firms.”](https://www.harvey.ai/blog/helping-law-firms-and-companies-collaborate-at-scale) Three and a half years later they report more than 1,000 customers in 60 countries, including half of the Am Law 100, and over 500 in-house legal teams.

The same post says, “Nothing about our commitment to law firms is changing.” It also says Harvey “is proud to serve both law firms and enterprises,” and that “law firms still represent the largest portion of our customer base.” The mix is moving. Its chief executive told [TechCrunch](https://techcrunch.com/2025/11/14/inside-harvey-how-a-first-year-legal-associate-built-one-of-silicon-valleys-hottest-startups/) in November 2025 that about 4% of revenue came from corporates at the start of the year and 33% did by then. In July 2026 Harvey said it [already works with more than 125 asset management firms](https://www.harvey.ai/blog/y-combinator-backed-benchmark-joins-harvey).

Left: Harvey’s founders wrote in March 2026, “Nothing about our commitment to law firms is changing.” Right: in an October 2026 demo for a roofing platform that is not a law firm, Harvey offered the product for about $20,000 a year with no usage charges. A first-hand account of one demo, not independently verified.

Here is what we saw. In October 2026 we sat through a Harvey demo. We are a roofing platform, not a law firm. Harvey offered to sell us the product anyway: about $20,000 a year, with no usage charges.

Harvey has no public price list that we could find. Its chief operating officer has [described the model in public](https://legaltechnology.com/harveys-15-5bn-valuation-coo-katie-burke-on-what-now-and-what-next/) as seats, “with the option to do consumption for folks who are asking us for it.” A flat annual fee with no usage charge fits the seat half of that description. One quote does not make a pattern, and we do not read much into the price. We read this: a product sold as vertical was offered to an outsider on terms that look like a software license.

This is not Harvey becoming a law firm. Nothing it has said in public points that way, and nothing in the demo did. The point is smaller and more useful. A vertical label describes where a vendor started. It does not describe where the vendor will sell. If the leader’s product is available to a roofing platform at a flat annual price, the thing a roofing point solution has to defend is not its vertical.

Treat the demo as a first-hand account, not a measurement. It is one demo and one quote, there is no public price list to check it against, and terms vary by buyer.

## The seat is a *sliver.*

A point solution is paid for out of a company’s software budget, and software spending is small beside the service work it supports. [General Catalyst](https://www.generalcatalyst.com/stories/the-future-of-services), an investor that buys and runs service companies, puts the service industries it covers at “$6T+ annually” against a “$370B software market.” Until now, it writes, technology for these businesses was “limited to SaaS products that charged license fees per user.”

Roofing has its own measurement, and it is smaller still. In the [2022 Economic Census](https://www2.census.gov/programs-surveys/economic-census/data/2022/sector23/EC2223BASIC.zip), roofing contractors with employees reported $69.9 billion in receipts. They paid $13.5 billion in payroll and expensed $154 million of software purchases, which is 0.22% of receipts, or 0.33% counting purchased computer services. Payroll was about 87 times the software line.

Share of receipts, U.S. roofing contractors with employees, 2022

- Payroll**19.3%**
- Software and computer services**0.33%**
- Software alone**0.22%**

Census Economic Census 2022, NAICS 238160. Payroll counts paid employees only, so it leaves out labor a firm subcontracts. Shares are derived; the figures and arithmetic are in the methods.

Read the number with care. It describes 2022, before most roofing AI tools existed. It counts expensed purchases only, so capitalized software is out, and the Census marks part of the roofing row as imputed. Spending has surely grown since. The shape probably has not: people are the cost stack, and software is a thin line beside them. A [2020 survey of construction firms](https://www.agc.org/sites/default/files/Files/Communications/2020_Outlook_Survey_National.pdf) asked what share of revenue goes to IT. It drew 956 responses in total, though the count varied by question. Sixty-five percent said under 2%, 15% said 2% or more, and 20% did not know.

Law has detailed software-spending data from one vendor, and it points the same way. [Clio’s 2024 report](https://www.clio.com/about/press/clio-latest-legal-trends-report/) puts software at 0.58% of expenses for solo practitioners and between 1.37% and 1.77% for firms with employees. Its [2025 report for solo and small firms](https://www.clio.com/blog/solo-small-law-firms-highlights-2025-legal-trends/) says the average law firm spends “just under 2% of its overall expenses on software.” A popular version of this argument says software takes 2% of a firm’s revenue. That is the wrong denominator for law, because expenses are not revenue. For roofing the measured share is between a tenth and a sixth of it. What survives is the order of magnitude: low single digits at the most.

## The seat and the job (explorer)

On the page, three sliders set a roofing company’s revenue, what it spends on software as a share of revenue, and how many tools that budget is spread across. A 100-square picture shows one year of the company’s revenue, a square for each percent. The starting revenue and software share are the measured averages for U.S. roofing firms with employees. The starting tool count is an illustrative assumption, not a measurement. The arithmetic is multiplication, not a forecast.

| Input | Starting value |
| --- | --- |
| Company revenue | $2,900,000 a year |
| Software, as a share of revenue | 0.33% |
| Tools the budget is spread across | 8 (assumed) |

| Result at those values | |
| --- | --- |
| All software, every tool | $9,570 a year |
| One tool, on average | $1,196 a year |
| One point of margin | $29,000 a year |
| Customers a single-tool vendor needs for $10 million a year | 8,360: 35% of all roofing firms with employees |
| Customers needed for $100 million a year | 83,595: 3.5 times the number of roofing firms with employees |

Move the sliders and the pattern holds. At the measured share, spread across an assumed eight tools, the average roofing company spends $1,196 a year on any one of them. A vendor selling one tool at that price needs 8,360 customers to reach $10 million a year: 35% of all roofing firms with employees. To reach $100 million it needs 83,595: 3.5 times the number of roofing firms with employees.

At 2%, six times the share measured with computer services, $10 million takes 5.7% of the industry and $100 million takes 57% of it. Venture-scale revenue from one roofing tool needs most of the industry on generous assumptions and more than all of it on measured ones. That is why every one of these wedges is pitched as the first step into something wider.

Harvey describes its own move beyond the seat. “Right now it’s mostly seats,” its chief executive told [TechCrunch](https://techcrunch.com/2025/11/14/inside-harvey-how-a-first-year-legal-associate-built-one-of-silicon-valleys-hottest-startups/) in November 2025, “but we’re moving to more outcome-based pricing as the workflows get more complex.” He expected the product to stay “a productivity suite sold seat-based” for “at least the next year or two,” with consumption-based workflows arriving “slowly over time.”

## Three dead ends, *one cause.*

The cause is the same each time: the seat is thin and the wedge gets absorbed. What differs is who is left holding it.

### For the software company: *no moat.*

[Sequoia’s 2023 warning](https://www.sequoiacap.com/article/generative-ai-act-two/) was about data: “the ‘data moats’ are on shaky ground: the data that application companies generate does not create an insurmountable moat, and the next generations of foundation models may very well obliterate any data moats that startups generate.” What it saw building more durable advantage was different: “workflows and user networks.”

[Andreessen Horowitz](https://a16z.com/the-new-business-of-ai-and-how-its-different-from-traditional-software/) made the margin version of the point in 2020. AI companies’ gross margins often sat in the “50-60% range,” below the 60-80%+ of comparable software, with “weaker defensive moats due to the commoditization of AI models and challenges with data network effects.” Some companies that planned to sell pure software were “increasingly bringing a services capability in-house.” That was 2020, and the costs it named were cloud infrastructure and human support, not tokens. The direction is what matters.

So a wedge buys an introduction. What a vendor needs after the introduction is one of three things: a place in the customer’s workflow that is expensive to leave, a network around it, or an outcome it stands behind. A tool that does one step well and holds none of them is the first thing the next assistant release makes optional.

One caution runs the other way. Sequoia’s durable moats were workflows and networks, not data. A roofing system where the owner actually runs the business, the CRM and field system of record, is not a point solution in the sense this paper means. It is a candidate foundation, and the two we looked at are already opening themselves to the assistants. [Contractors Cloud](https://api.contractorscloud.com/docs/guides/connect-your-ai-agent) publishes a guide for connecting Claude, ChatGPT or another AI agent to a customer’s account through MCP. Jobber, a general field-service suite, sells [an AI receptionist](https://www.getjobber.com/features/ai-receptionist/) inside its product and in September 2026 [opened its data to ChatGPT and Claude](https://www.prnewswire.com/news-releases/jobber-launches-mcp-bringing-home-and-commercial-service-data-and-workflows-directly-into-chatgpt-and-claude-302890385.html).

### For the roofer: *no foundation.*

Follow one job from the first call to the last review. In the point-solution pitch each step gets its own tool, and each tool keeps its own copy of the homeowner. Anthropic stated the underlying problem in 2024, about assistants and the systems around them: [“Every new data source requires its own custom implementation.”](https://www.anthropic.com/news/model-context-protocol) An owner who buys a tool per step becomes the integrator.

1. Lead
2. Inspection
3. Estimate
4. Insurance claim
5. Contract
6. Production
7. Materials
8. Invoice
9. Review

**One job record**The same customer. The same data.

One roofing job passes through nine steps, from the first call to the review. In the point-solution pitch each step gets its own tool, and each tool keeps its own copy of the customer. The gaps between tools are handoffs. In the second half the tools connect to a single job record. The tools shown are illustrative, not a count of what any roofer uses.

The bill moves the wrong way as well. In law, Clio found software spending growing about 20% a year since 2013 while revenue grew 9%: a stack of subscriptions that outgrows the business it serves. We found no roofing equivalent, and we do not claim one.

Even a bargain is one workflow. The offer above, an assistant with no usage meter for about $20,000 a year, covers one of the many workflows a roofing company runs, and tools sold one step at a time do not share what they know unless someone connects them.

Someone also answers when the tool is wrong. A receptionist that answers a call can record it. [Florida requires every party’s consent](https://www.leg.state.fl.us/statutes/index.cfm?App_mode=Display_Statute&URL=0900-0999/0934/Sections/0934.03.html) to record a call; [Texas requires one party’s](https://tcss.legis.texas.gov/resources/PE/htm/PE.16.htm). On insurance claims, [Texas bars a contractor from acting as a public adjuster](https://tcss.legis.texas.gov/resources/IN/htm/IN.4102.htm) on its own jobs, and [Florida bars a licensed contractor from performing public adjuster services](https://www.leg.state.fl.us/statutes/index.cfm?App_mode=Display_Statute&URL=0600-0699/0626/Sections/0626.854.html) unless it is also licensed as one. Neither adjuster statute mentions software, and OpenAI’s launch post tells developers they “must also make it clear to end users when they’re interacting with AI, unless it’s already obvious from the context.” In our reading, that duty lands on whoever runs the phone line, whatever tool answers it.

Four questions to put to any tool before you buy it:

- Does it write to your job record, or keep its own?
- What happens to your data and your workflows if you leave?
- Who answers when it is wrong, and what does that cost you?
- What does it do that the assistant you already pay for will not do next year?

### For the investor: *a low ceiling.*

About 24,000 roofing firms in the U.S. have employees ([Economic Census, 2022](https://www2.census.gov/programs-surveys/economic-census/data/2022/sector23/EC2223BASIC.zip)). Of the 25,500 or so roofing establishments with paid employees, 90% have fewer than 20 employees, and all of them together employ about 215,000 people ([County Business Patterns, 2023](https://www2.census.gov/programs-surveys/cbp/datasets/2023/cbp23us.zip)). That is the pool of roofing businesses with employees, before you subtract the ones that will never buy a tool. Roofers with no employees are counted by the Census only within a wider group of trades, so they are not in these figures.

A wedge is pitched as the first step into something wider, and that is the honest version of the pitch. The money that agrees with this paper is not funding the wedge. [General Catalyst](https://www.generalcatalyst.com/stories/the-future-of-services) writes that “technology builders capture even more value by directly owning and operating the service businesses that deliver these transformative capabilities to end customers.” It raised $1.5 billion in 2024 for that strategy, according to [PitchBook](https://pitchbook.com/news/articles/general-catalysts-6-3b-amex-deal-puts-its-ai-roll-up-strategy-on-display). One of its companies, Eudia, bought Johnson Hana, a legal delivery team of 300 or more people. In General Catalyst’s words, that structure “enables Eudia to move beyond point solutions and deliver complete legal outcomes as a subscription-based service.”

Long Lake, an AI holding company backed by General Catalyst among others, [completed its $6.3 billion purchase of American Express Global Business Travel](https://www.amexglobalbusinesstravel.com/press-releases/long-lake-worlds-first-ai-holding-company-completes-acquisition-amex-gbt/) on 29 September 2026.

None of that proves the strategy works. Closing a purchase is not the same as showing the model pays. None of the sources we opened reports software-like margins from a completed roll-up, and the operating figures in General Catalyst’s essay are its own, which we did not audit. What the deals do show is where the capital is pointed: at owning the service, not at selling it a seat.

The question for an investor in roofing AI is therefore not whether the tools work. It is who keeps what they create.

Twenty identical workflow cards stand above a floor line. Row by row, bottom row first, they tip over and fall through it into the dark. One card, the one that was marked with the gold tick, falls last and lands with its tick still on. The page reads: a wedge is a feature on someone else’s roadmap. Tools fall. Outcomes stay.

## What would change *our mind.*

This paper could be wrong, and here is how. A point solution survives when it has something an assistant cannot copy. There are four things that count.

### Data the assistants cannot get

Roofing has real examples. Verisk’s [license terms](https://www.verisk.com/privacy-policies/xactware-eula/) say, “We own all Price Data provided with the Services,” and EagleView says it collects and owns the aerial imagery it trains on. Both have announced their own AI and open it to outside AI systems on their own terms. Verisk’s terms go further: unless a particular use is authorized, a licensee “must not … utilize artificial intelligence, machine learning, statistical learning, neural networks, large language models, generative artificial intelligence” on its content. In our view, a tool built on a dataset it owns is not a wrapper. In these two cases, the tool is the data owner’s own product.

### A place in the workflow that is expensive to leave

Sequoia’s 2023 view was that [workflows and user networks](https://www.sequoiacap.com/article/generative-ai-act-two/) hold up better than data. A system of record where the owner runs the business is that kind of place. The contractor systems named above sit there, and they are the ones opening themselves to the assistants.

### A rule that requires it

Texas and Florida decide who may speak for a policyholder on a claim. Neither statute mentions software. In our reading, rules like these protect the licensed person, not the point solution that drafts the letter.

### An outcome it stands behind

A vendor paid on the result rather than the seat is aligned with its customer in a way a license cannot be. Harvey’s chief executive told [TechCrunch](https://techcrunch.com/2025/11/14/inside-harvey-how-a-first-year-legal-associate-built-one-of-silicon-valleys-hottest-startups/) that Harvey is moving toward outcome-based pricing, starting with “very small things,” while the core product stays seat-based for now.

Ask which of the four a given tool has. If the answer is none, the argument in this paper applies to it. If the answer is one or more, it does not, and we would want to hear about it.

We turned the framing this paper started from into 24 claims and graded each against primary sources: 4 supported, 12 partial, 8 unsupported, 0 contradicted. The conclusion survives. Several of the essay’s numbers do not. The full ledger is in the methods.

## What *lasts.*

Three things compound when a model improves instead of replacing the product around it: the workflows a company runs, the data they leave behind, and the accountability for the result. A foundation is the three together, owned by someone who answers for the outcome.

Owning the outcome does not mean the people doing the work become a cost line. In a roofing company, owners and crews are the ones who handle what the system escalates. This paper is about who owns the system, not about doing without them.

A stack drawn from the bottom up: the data, then all the workflows that run on it, then an operating layer above them. A bracket marks the whole stack as the outcome, answered for by one party, with owners and crews handling what the system escalates. A tool earns a sliver of each job. Whoever owns the outcome answers for it, and keeps what compounds.

## Next steps

If you run a roofing company and want to think this through, we are glad to talk.

[Start a private conversation](https://foundationproject.net/schedule)

This paper is research and opinion from an interested party. It is not investment, legal or insurance advice, and it is not an offer of securities.

Research notes

## Methods, sources and the ledger

Everything the paper asserts is listed here with where it came from and what it does not show. Dates are when pages were read: 17 October 2026.

### How we tested the argument

The argument began as a 15-minute video essay by Devin Kearns, [“Vertical AI is DEAD”](https://youtu.be/Wc4C7iz3cA0), published 30 September 2026. We cite it as the source of the framing, not as evidence. We listed the checkable claims in it (24), then opened primary sources for each: company announcements, product documentation, investor publications, government data and statute text.

Each claim got one of four grades.

**Supported**

A primary source states the claim, or a close version of it.

**Partial**

The direction is real. The wording, number or subject is off.

**Unsupported**

No source we opened establishes it.

**Contradicted**

A primary source says the opposite.

Where our working notes used a different label, such as “too strong” or “arithmetic on an assumption,” we mapped it to the nearest grade and kept the label in the note. The result: 4 supported, 12 partial, 8 unsupported, 0 contradicted. The paper uses supported claims, and partial claims only in their corrected form. Unsupported and contradicted claims are not used.

Our working audit changed in three ways before publication. It said OpenAI’s 2023 plugins page announces that plugins were deprecated. The page today says only that the launch post is outdated, and OpenAI’s current plugins page presents plugins for ChatGPT, so we dropped the claim and used what the pages say. It called the Amex GBT purchase pending, and that purchase has since closed. And a separate fact-check of this paper against its sources moved four grades down. Two claims first graded contradicted became unsupported, because Harvey’s post is about its customers and Sequoia never discusses distribution as a moat. Two graded supported became partial, because the sources describe parts of the claim and none states it. The same check corrected several paraphrases that went further than their sources.

### The claim ledger

*The 24 claims, how each held up, and what the sources show.*

| Claim | Grade | What the sources show | Sources |
| --- | --- | --- | --- |
| AI for an industry started as point solutions: find one expensive workflow, add AI. | Partial | Sequoia documented the copilot glut in 2023 and named the turn to end-to-end products. No source counts how many startups followed this exact playbook. | [Sequoia Capital](https://www.sequoiacap.com/article/generative-ai-act-two/) |
| Companies that start as point solutions expand into the operating system for the industry. | Partial | Harvard reads Harvey as “a platform, infrastructure, and ecosystem play,” so the expansion is real. It also says the founders resisted the advice to pick a narrow use case, so Harvey did not start as a point solution. No source we opened counts other companies that did. | [Harvey](https://www.harvey.ai/blog/helping-law-firms-and-companies-collaborate-at-scale), [Harvard Law School Center on the Legal Profession](https://clp.law.harvard.edu/ai-and-the-legal-profession/essays/harveys-strategic-evolution.html) |
| Venture investors are still underwriting the old playbook and do not know the ground. | Unsupported | Sequoia called the 2023 “feeding frenzy” unsustainable and the copilot pitches “undifferentiated,” so the critique is old. No source we opened measures what investors believe in 2026. | [Sequoia Capital](https://www.sequoiacap.com/article/generative-ai-act-two/), [General Catalyst](https://www.generalcatalyst.com/stories/the-future-of-services) |
| General-purpose assistants already do much of what point solutions sold. | Supported | File search and custom assistants shipped inside ChatGPT, a shared protocol was launched to replace one-off connectors, skills shipped inside Claude, and OpenAI’s plugins page says to connect business apps “right into ChatGPT with plugins.” | [OpenAI](https://openai.com/index/introducing-gpts), [OpenAI Help Center](https://help.openai.com/en/articles/8555545-file-uploads-faq), [Anthropic](https://www.anthropic.com/news/model-context-protocol), [Anthropic](https://www.anthropic.com/news/skills), [OpenAI](https://openai.com/business/plugins/) |
| Customers want one internal foundation, not a shelf of tools. | Partial | Neither source measures what buyers want. Anthropic describes one protocol in place of a custom connector for each source, which is a vendor’s case for its own product. We found no independent buyer survey. | [Anthropic](https://www.anthropic.com/news/model-context-protocol), [Harvey](https://www.harvey.ai/blog/helping-law-firms-and-companies-collaborate-at-scale) |
| Nvidia will undercut the frontier labs with in-house hardware and compute-only pricing. | Unsupported | NVIDIA’s pages describe open models. They do not describe an on-premises box or a price comparison. The pricing claim is a prediction. | [NVIDIA](https://investor.nvidia.com/news/press-release-details/2026/NVIDIA-Launches-Nemotron-Coalition-of-Leading-Global-AI-Labs-to-Advance-Open-Frontier-Models/default.aspx), [NVIDIA Blog](https://blogs.nvidia.com/blog/ai-future-open-and-proprietary/) |
| Nvidia is building open models that companies can specialize. | Supported | The Nemotron Coalition announcement and NVIDIA’s blog both say so. | [NVIDIA](https://investor.nvidia.com/news/press-release-details/2026/NVIDIA-Launches-Nemotron-Coalition-of-Leading-Global-AI-Labs-to-Advance-Open-Frontier-Models/default.aspx), [NVIDIA Blog](https://blogs.nvidia.com/blog/ai-future-open-and-proprietary/) |
| The wrapper around the model was the product because the model could not do the job alone. | Partial | The pages describe pieces of this. OpenAI called plugins the “eyes and ears” of a language model, Sequoia wrote that the next wave uses models as “a piece of a more comprehensive solution,” and Anthropic said even the best models are “constrained by their isolation from data.” None says the wrapper was the product because the model could not do the job alone. That is our synthesis. | [OpenAI](https://openai.com/index/chatgpt-plugins), [Sequoia Capital](https://www.sequoiacap.com/article/generative-ai-act-two/), [Anthropic](https://www.anthropic.com/news/model-context-protocol) |
| Ontology, context engine and company brain are a real product category. | Partial | Palantir describes its Ontology as “an operational layer for the organization,” and PitchBook says Eudia’s main product is a “Company Brain.” We found no source for “context engine,” and Palantir’s page is undated, so how long the category has existed is not shown. | [Palantir](https://www.palantir.com/docs/foundry/ontology/overview), [PitchBook](https://pitchbook.com/news/articles/general-catalysts-6-3b-amex-deal-puts-its-ai-roll-up-strategy-on-display) |
| Saying those words gets a startup lumped in with every other pitch. | Unsupported | An anecdote. No source we opened measures it. | None found |
| Software is about 2% of a law firm’s revenue. | Partial | Clio’s 2024 report measures software at 0.58% to 1.77% of expenses by firm size, and its 2025 report for solo and small firms says the average law firm spends “just under 2% of its overall expenses on software.” Expenses are not revenue. The order of magnitude holds. | [Clio](https://www.clio.com/about/press/clio-latest-legal-trends-report/), [Clio](https://www.clio.com/blog/solo-small-law-firms-highlights-2025-legal-trends/) |
| Becoming the service captures the budget that software never touches. | Supported | General Catalyst states it as its thesis and is funding it. | [General Catalyst](https://www.generalcatalyst.com/stories/the-future-of-services) |
| Harvey is walking the path to becoming a law firm. | Unsupported | No source we opened shows Harvey operating as a law firm. Its March 2026 post says its commitment to law firms is not changing and that it serves both law firms and enterprises, and Harvard reads it as a platform, infrastructure and ecosystem play. | [Harvey](https://www.harvey.ai/blog/helping-law-firms-and-companies-collaborate-at-scale), [Harvard Law School Center on the Legal Profession](https://clp.law.harvard.edu/ai-and-the-legal-profession/essays/harveys-strategic-evolution.html) |
| A $1M case yields $20,000 of software, so a vendor needs 50 firms to earn $1M. | Partial | The arithmetic is right if software is 2% of revenue. That share is an assumption, not a measurement. | [Clio](https://www.clio.com/about/press/clio-latest-legal-trends-report/) |
| Selling prompts, custom GPTs, workflow automation and skills is a real sequence. | Partial | A fair product history. The dating of the later steps is loose. | [OpenAI](https://openai.com/index/introducing-gpts), [Anthropic](https://www.anthropic.com/news/skills), [n8n](https://n8n.io/ai/) |
| Each earlier era of AI tools is now obsolete. | Partial | Too strong. OpenAI’s 2023 plugins post is marked outdated, but plugins are a current ChatGPT feature again. The GPTs launch post now points readers to workspace agents and ChatGPT Work as “the current experience,” so that era has a successor, but no page we opened says GPTs ended. n8n still sells workflow automation as a living product. | [OpenAI](https://openai.com/index/chatgpt-plugins), [OpenAI](https://openai.com/business/plugins/), [OpenAI](https://openai.com/index/introducing-gpts), [n8n](https://n8n.io/ai/) |
| Vertical products that hard-code business logic are being eaten by the general assistants. | Partial | The mechanism is published: Sequoia on weak data moats, a16z on weak defenses. Extinction is not shown. | [Sequoia Capital](https://www.sequoiacap.com/article/generative-ai-act-two/), [Andreessen Horowitz](https://a16z.com/the-new-business-of-ai-and-how-its-different-from-traditional-software/) |
| Buyers ask to be taught Claude instead of buying the custom build. | Unsupported | A field report. We found no survey, earnings call or win-loss note. | None found |
| AI roll-ups and owning the service are an active capital strategy. | Supported | General Catalyst, Eudia and Long Lake describe it and put money into it. | [General Catalyst](https://www.generalcatalyst.com/stories/the-future-of-services), [PitchBook](https://pitchbook.com/news/articles/general-catalysts-6-3b-amex-deal-puts-its-ai-roll-up-strategy-on-display) |
| That strategy has already worked as a completed model. | Unsupported | The Amex GBT purchase closed on 29 September 2026, but closing a purchase is not showing the model works. None of the sources we opened reports software-like margins from a completed roll-up, and the portfolio operating figures are the investor’s own. | [American Express Global Business Travel and Long Lake](https://www.amexglobalbusinesstravel.com/press-releases/long-lake-worlds-first-ai-holding-company-completes-acquisition-amex-gbt/), [PitchBook](https://pitchbook.com/news/articles/general-catalysts-6-3b-amex-deal-puts-its-ai-roll-up-strategy-on-display), [General Catalyst](https://www.generalcatalyst.com/stories/the-future-of-services) |
| AI-native businesses are measurably much cheaper than incumbents. | Partial | Pricing pressure is documented in law: flat fees, exposed hourly work. We found no price index of AI-native firms against incumbents. | [Clio](https://www.clio.com/about/press/clio-latest-legal-trends-report/), [General Catalyst](https://www.generalcatalyst.com/stories/the-future-of-services) |
| Distribution is not a moat. | Unsupported | Sequoia does not call distribution a moat or say it is not one. It says generative AI’s biggest problem “is not finding use cases or demand or distribution, it is proving value,” and that “workflows and user networks seem to be creating more durable sources of competitive advantage.” Neither speaks to the claim. | [Sequoia Capital](https://www.sequoiacap.com/article/generative-ai-act-two/) |
| A reckoning over token costs is already in the numbers. | Partial | The best source is a 2020 a16z warning about AI gross margins, which names cloud infrastructure and human support as the costs, not tokens. We found no source on token prices either way. | [Andreessen Horowitz](https://a16z.com/the-new-business-of-ai-and-how-its-different-from-traditional-software/) |
| Decentralized networks are the next enterprise structure. | Unsupported | Nothing to cite. The author describes an unfinished article. | None found |

### The roofing figures

Roofing contractors are NAICS 238160. Every figure is for businesses with paid employees. The Census does not publish nonemployer roofing businesses at this level (the nearest category, 2381, covers roofing and seven other trades). The Census publishes dollar items in thousands; values here are whole dollars. The Economic Census row for roofing carries imputation flags: receipts 10% to under 20% imputed, employment 20% to under 30%. Software is expensed purchases only, so capitalized software is not counted. The data describe 2022 or 2023, before most roofing AI tools existed. As a cross-check, the Census Bureau’s [Statistics of U.S. Businesses](https://www2.census.gov/programs-surveys/susb/tables/2022/us_state_naics_detailedsizes_2022.txt) for 2022 counts 24,044 roofing firms with employees, $68.8 billion in receipts and $13.3 billion in payroll, within 2% of the Economic Census row.

*Figures read from Census files, and values derived from them.*

| Figure | Value | Year | Where it was read |
| --- | --- | --- | --- |
| Roofing firms with paid employees | 24,082 | 2022 | [U.S. Census Bureau](https://www2.census.gov/programs-surveys/economic-census/data/2022/sector23/EC2223BASIC.zip): EC2223BASIC.dat, GEOTYPE=01, NAICS2022=238160, column FIRM |
| Roofing establishments with paid employees | 25,519 | 2023 | [U.S. Census Bureau](https://www2.census.gov/programs-surveys/cbp/datasets/2023/cbp23us.zip): cbp23us.txt, naics=238160, lfo=-, column est |
| Employees at those establishments (March 12 pay period) | 215,242 | 2023 | [U.S. Census Bureau](https://www2.census.gov/programs-surveys/cbp/datasets/2023/cbp23us.zip): cbp23us.txt, naics=238160, lfo=-, column emp |
| Sales, value of shipments, or revenue | $69,906,202,000 | 2022 | [U.S. Census Bureau](https://www2.census.gov/programs-surveys/economic-census/data/2022/sector23/EC2223BASIC.zip): EC2223BASIC.dat, GEOTYPE=01, NAICS2022=238160, column RCPTOT ($1,000) |
| Annual payroll | $13,458,130,000 | 2022 | [U.S. Census Bureau](https://www2.census.gov/programs-surveys/economic-census/data/2022/sector23/EC2223BASIC.zip): EC2223BASIC.dat, GEOTYPE=01, NAICS2022=238160, column PAYANN ($1,000) |
| Expensed purchases of software | $154,272,000 | 2022 | [U.S. Census Bureau](https://www2.census.gov/programs-surveys/economic-census/data/2022/sector23/EC2223BASIC.zip): EC2223BASIC.dat, GEOTYPE=01, NAICS2022=238160, column PCHEXSO ($1,000) |
| Data processing and other purchased computer services | $74,209,000 | 2022 | [U.S. Census Bureau](https://www2.census.gov/programs-surveys/economic-census/data/2022/sector23/EC2223BASIC.zip): EC2223BASIC.dat, GEOTYPE=01, NAICS2022=238160, column PCHDAPR ($1,000) |
| Establishments with fewer than 20 employees | 23,053 | 2023 | [U.S. Census Bureau](https://www2.census.gov/programs-surveys/cbp/datasets/2023/cbp23us.zip): cbp23us.txt, naics=238160, lfo=-, columns n\<5 + n5\_9 + n10\_19 (16,661 + 3,826 + 2,566) |
| Average receipts per roofing firm with employees | $2,902,840 | derived | receipts ÷ firms = 69,906,202,000 ÷ 24,082 |
| Expensed software purchases as a share of receipts | 0.22% | derived | software ÷ receipts = 154,272,000 ÷ 69,906,202,000 |
| Software plus purchased computer services as a share of receipts | 0.33% | derived | (software + computer services) ÷ receipts = (154,272,000 + 74,209,000) ÷ 69,906,202,000 |
| Payroll as a share of receipts | 19.25% | derived | payroll ÷ receipts = 13,458,130,000 ÷ 69,906,202,000 |
| Payroll as a multiple of expensed software purchases | 87.2× | derived | payroll ÷ software = 13,458,130,000 ÷ 154,272,000 |
| Share of establishments with fewer than 20 employees | 90.34% | derived | 23,053 ÷ 25,519 |

### What the explorer does and does not do

The explorer is multiplication. A company’s software budget is its revenue times the share you set. One tool’s average take is that budget divided by the number of tools. One point of margin is 1% of revenue. The customers a single-tool vendor needs are the revenue goal divided by one tool’s take, rounded up, set against the 24,082 roofing firms with employees.

It is not a forecast and not a model of any real vendor. Real prices vary, a vendor may sell to adjacent trades, and customers rarely pay the average. Its starting revenue and software share are the measured averages above. The starting tool count, eight, is an illustrative assumption: we found no measure of how many tools a roofing company runs, and it sets every per-tool figure. The other presets are assumptions: the high end of what Clio measured for law firms, and a generous 3%.

### A first-hand account

In October 2026 Foundation Projects sat through a demo with Harvey. The company offered the product to us, a roofing platform that is not a law firm, for about $20,000 a year with no usage charges. That is our account of one conversation. We hold no published price to check it against, we have not confirmed it with Harvey, and terms vary by buyer.

What is public points both ways. Harvey’s March 2026 post says its commitment to law firms is not changing, that it serves law firms and enterprises, and that law firms are still its largest customer group. Its chief executive says revenue from corporates grew from about 4% to 33% in 2025 and that pricing is mostly seats today, with outcome-based pricing coming first for very small things. Its chief operating officer says seats, with consumption for those who ask. It has no public price list that we could find.

The wedges we describe as pitched to us most often, an AI voice receptionist, an insurance supplement writer and an instant-quote tool, are likewise our own experience, not a survey.

### What we did not use

- Figures in Harvard’s May 2026 update (valuations, user counts, joint ventures): the primary announcements were not opened.
- Clio’s 2025 solo and small firm report itself: we did not page the report. We used Clio’s own blog summary of it, read through an archived copy, and the 2024 press release for the cohort figures.
- Any SEC filing on NVIDIA open-model spending, any Anthropic or OpenAI services joint venture announcement, a primary source for the term “vibe coding,” and a buyer survey on whether buyers ask to be taught Claude. None was opened, so none is used.
- Newer roofing revenue. The 2024 annual enterprise survey publishes construction only to three-digit industry codes.

### Sources

#### Government data

- [Economic Census 2022, Construction: basic statistics, NAICS 238160 (roofing contractors)](https://www2.census.gov/programs-surveys/economic-census/data/2022/sector23/EC2223BASIC.zip)U.S. Census Bureau, 2022 data**Used for:** Roofing contractors with paid employees: firms, receipts, payroll, and expensed purchases of software and computer services.**Does not show:** Employer establishments only, for 2022, before most roofing AI tools existed. Software is expensed purchases only, so capitalized software is excluded. The row carries imputation flags: receipts 10% to under 20% imputed, employment 20% to under 30%.
- [County Business Patterns 2023, NAICS 238160 (roofing contractors)](https://www2.census.gov/programs-surveys/cbp/datasets/2023/cbp23us.zip)U.S. Census Bureau, 2023 data**Used for:** Roofing establishments with paid employees, their employment, and their size distribution.**Does not show:** An establishment is one physical location. Businesses with no employees are excluded. Employment is the March 12 pay period.
- [Statistics of U.S. Businesses 2022, NAICS 238160 (roofing contractors)](https://www2.census.gov/programs-surveys/susb/tables/2022/us_state_naics_detailedsizes_2022.txt)U.S. Census Bureau, 2022 data**Used for:** A cross-check on the Economic Census row: 24,044 roofing firms with employees, $68.8 billion in receipts and $13.3 billion in payroll.**Does not show:** A firm is all establishments under common ownership in one industry and area. Employer businesses only, and not used for any figure in the paper beyond this check.
- [Texas Insurance Code, chapter 4102 (public adjusters)](https://tcss.legis.texas.gov/resources/IN/htm/IN.4102.htm)Texas Legislative Council, 2026-10**Used for:** Texas barring a contractor from acting as a public adjuster on its own jobs.**Does not show:** Statute text, read October 2026. Not legal advice.
- [Florida Statutes, section 626.854 (public adjusters)](https://www.leg.state.fl.us/statutes/index.cfm?App_mode=Display_Statute&URL=0600-0699/0626/Sections/0626.854.html)Florida Legislature, 2026-10**Used for:** Florida defining a public adjuster as anyone other than a licensed attorney who is paid to prepare or file an insurance claim for an insured, and barring a licensed contractor from performing public adjuster services unless also licensed as one.**Does not show:** Statute text, read October 2026. Not legal advice.
- [Florida Statutes, section 934.03 (interception of communications)](https://www.leg.state.fl.us/statutes/index.cfm?App_mode=Display_Statute&URL=0900-0999/0934/Sections/0934.03.html)Florida Legislature, 2026-10**Used for:** Florida’s all-party consent rule for recording a call.**Does not show:** Statute text, read October 2026. Not legal advice.
- [Texas Penal Code, section 16.02 (unlawful interception)](https://tcss.legis.texas.gov/resources/PE/htm/PE.16.htm)Texas Legislative Council, 2026-10**Used for:** Texas’s one-party consent rule for recording a call.**Does not show:** Statute text, read October 2026. Not legal advice.

#### Company announcement

- [How Harvey is helping law firms and companies collaborate at scale](https://www.harvey.ai/blog/helping-law-firms-and-companies-collaborate-at-scale)Harvey, 2026-03-13**Used for:** Harvey’s stated focus on law firms, its customer counts, its expansion to in-house teams, its commitment to law firms, and that law firms are still its largest customer group.**Does not show:** Company-reported figures. It says nothing about pricing or about selling outside law.
- [Y Combinator-backed Benchmark joins Harvey](https://www.harvey.ai/blog/y-combinator-backed-benchmark-joins-harvey)Harvey, 2026-07-16**Used for:** Harvey’s statement that it already works with more than 125 asset management firms.**Does not show:** Company-reported. Customer names are Harvey’s own.
- [ChatGPT plugins](https://openai.com/index/chatgpt-plugins)OpenAI, 2023-03-23**Used for:** The 2023 description of plugins as the “eyes and ears” of a language model, and the note on the page today that the launch post is outdated.**Does not show:** The page says only that the post is outdated. It does not say plugins ended: OpenAI’s current plugins page, listed below, presents plugins for ChatGPT.
- [Introducing GPTs](https://openai.com/index/introducing-gpts)OpenAI, 2023-11-06**Used for:** Custom assistants built from instructions, knowledge and skills without code, and file attachment for searching PDFs.**Does not show:** A 2023 launch post. Today it carries a banner: “This post introduced GPTs. For the current experience, explore: Workspace agents, ChatGPT Work.” It does not say how many point solutions GPTs affected.
- [Introducing gpt-realtime and Realtime API updates for production voice agents](https://openai.com/index/introducing-gpt-realtime/)OpenAI, 2025-08-28**Used for:** The Realtime API supporting phone calling through SIP, and the instruction that developers “must also make it clear to end users when they’re interacting with AI, unless it’s already obvious from the context.”**Does not show:** A launch post. It does not describe any roofing product.
- [Introducing the Model Context Protocol](https://www.anthropic.com/news/model-context-protocol)Anthropic, 2024-11-25**Used for:** The problem stated: models isolated from data, and a custom connector for every source. The protocol meant to replace them.**Does not show:** A vendor announcement. Adoption is not measured here.
- [Introducing Agent Skills](https://www.anthropic.com/news/skills)Anthropic, 2025-10-16**Used for:** What Agent Skills are, which skills ship, and which software companies say they build on them. Updated 18 December 2025 with organization-wide management and an open standard.**Does not show:** A vendor announcement. Partner statements are the partners’ own.
- [Verisk brings its trusted analytics and generative AI capabilities directly into Anthropic’s Claude](https://www.verisk.com/company/newsroom/verisk-brings-its-trusted-analytics-and-generative-ai-capabilities-directly-into-anthropics-claude/)Verisk, 2026-05-05**Used for:** Verisk’s insurance analytics reaching Claude through its own connectors.**Does not show:** A vendor announcement. Access requires Verisk subscriptions.
- [Eagleview data disclosure](https://www.eagleview.com/california-data-disclosure)EagleView, 2026-10**Used for:** EagleView stating, in a disclosure about the data it trains its AI on, that “All aerial imagery is collected and owned by Eagleview.”**Does not show:** A California AI training-data disclosure, read October 2026. It describes EagleView’s training data, not the whole market for aerial imagery.
- [Eagleview launches Eagleview Horizon, the agentic AI engine](https://www.globenewswire.com/news-release/2026/04/21/3277997/0/en/eagleview-launches-eagleview-horizon-the-agentic-ai-engine-powered-by-25-years-of-verified-property-intelligence.html)EagleView (GlobeNewswire), 2026-04-21**Used for:** EagleView announcing its own agentic AI product, with Model Context Protocol integrations and agent-to-agent capabilities that “enable external AI systems to directly integrate” its data.**Does not show:** A press release dated 21 April 2026. It says the product will be available by invitation only from 1 June. It does not say how EagleView licenses its data to others.
- [Jobber launches MCP, bringing home and commercial service data and workflows directly into ChatGPT and Claude](https://www.prnewswire.com/news-releases/jobber-launches-mcp-bringing-home-and-commercial-service-data-and-workflows-directly-into-chatgpt-and-claude-302890385.html)Jobber (PR Newswire), 2026-09-25**Used for:** A field-service system opening its data and workflows to ChatGPT and Claude through MCP.**Does not show:** A vendor release. Jobber is a general field-service suite, not a roofing-specific product.
- [NVIDIA Launches Nemotron Coalition of Leading Global AI Labs to Advance Open Frontier Models](https://investor.nvidia.com/news/press-release-details/2026/NVIDIA-Launches-Nemotron-Coalition-of-Leading-Global-AI-Labs-to-Advance-Open-Frontier-Models/default.aspx)NVIDIA, 2026-03-16**Used for:** An open model to be built by the coalition and released so that “developers and organizations worldwide” can “specialize AI for their industries and domains,” and the coalition’s members.**Does not show:** A press release. It says nothing about on-premises pricing or a per-token price comparison.
- [The future of AI is open and proprietary](https://blogs.nvidia.com/blog/ai-future-open-and-proprietary/)NVIDIA Blog, 2026-03-25**Used for:** The download count for Nemotron models, and the case for “systems of models, tuned and specialized for different modalities, domains and organizations.”**Does not show:** Company-reported.
- [Long Lake, the world’s first AI holding company, completes $6.3 Billion Acquisition of American Express Global Business Travel](https://www.amexglobalbusinesstravel.com/press-releases/long-lake-worlds-first-ai-holding-company-completes-acquisition-amex-gbt/)American Express Global Business Travel and Long Lake, 2026-09-29**Used for:** That the acquisition closed on 29 September 2026 at about $6.3 billion, that Long Lake is backed by investors including General Catalyst, and how the deal was financed.**Does not show:** A press release. It says nothing about returns, margins or whether the model works. The SEC Form 8-K for the same date confirms the closing.

#### Product documentation

- [Plugins: connect tools and apps to ChatGPT and Codex](https://openai.com/business/plugins/)OpenAI, 2026-10**Used for:** That OpenAI presents business tools as plugins that connect “right into ChatGPT,” for use in ChatGPT Work and Codex, filtered by category such as finance, sales, customer support and small business.**Does not show:** A product page, read October 2026. It carries no launch date, and it does not say how many tools are listed or how they sell.
- [File uploads FAQ](https://help.openai.com/en/articles/8555545-file-uploads-faq)OpenAI Help Center, 2026-10**Used for:** The document tasks ChatGPT lists today: summarizing a PDF, extracting quotes, comparing two documents.**Does not show:** A help page, read October 2026. Limits vary by plan.
- [Realtime API with SIP](https://developers.openai.com/api/docs/guides/voice-sip?voice-api=realtime)OpenAI, 2026-10**Used for:** Inbound phone calls directed to the API, and the builder owning the call logic.**Does not show:** Developer documentation, read October 2026.
- [API changelog](https://developers.openai.com/api/docs/changelog)OpenAI, 2026-09-10**Used for:** General availability of GPT-Live 1 and its per-minute voice price.**Does not show:** A changelog entry. Backend model and tool usage is billed separately.
- [Compare and analyze competing options](https://academy.claude.com/use-cases/compare-and-analyze-competing-options)Anthropic Academy, 2026-10**Used for:** Anthropic naming insurance quotes and contractor bids among documents to upload and compare.**Does not show:** A training page, read October 2026. Its worked example is choosing among competing vendor proposals. It is not an insurance-claims product, and applying it to carrier and contractor estimates is our inference.
- [Verisk XactRestore connector for Claude](https://www.verisk.com/resources/verisk-xactrestore-connector-for-claude/)Verisk, 2026-10**Used for:** Estimating tools inside Claude, a section aimed at roofing contractors, and the Verisk subscription and XactAI license it requires.**Does not show:** A vendor product page, read October 2026.
- [Xactware end user license agreement, section 3.6 (artificial intelligence and automation tools restrictions)](https://www.verisk.com/privacy-policies/xactware-eula/)Verisk, 2026-09-14**Used for:** Verisk’s license terms restricting the use of AI tools on Xactware services and content, except for an authorized AI use, and stating “We own all Price Data provided with the Services.”**Does not show:** The public license, as stamped 14 September 2026. We have not seen Verisk’s Services Agreement, which can authorize particular uses (its own Claude connector is one). Not legal advice.
- [Solar API: building insights](https://developers.google.com/maps/documentation/solar/building-insights)Google, 2026-10-16**Used for:** A general platform API returning the dimensions of a building’s roof.**Does not show:** Developer documentation. The terms for using it are listed separately.
- [Google Maps Platform service-specific terms](https://cloud.google.com/maps-platform/terms/maps-service-terms)Google, 2026-10-15**Used for:** The uses Google’s terms list for the Solar API.**Does not show:** Terms as read in October 2026. Not legal advice, and not a statement about any vendor’s license.
- [Jobber Receptionist for home service businesses](https://www.getjobber.com/features/ai-receptionist/)Jobber, 2026-10**Used for:** A general field-service suite selling an AI receptionist that “answers calls and texts 24/7, books work, takes requests, and fields customer questions.”**Does not show:** A vendor product page, read through an archived copy. Jobber is a general field-service suite, not a roofing-specific product.
- [Connect your AI agent](https://api.contractorscloud.com/docs/guides/connect-your-ai-agent)Contractors Cloud, 2026-10**Used for:** A guide for connecting Claude, ChatGPT or another AI agent to a Contractors Cloud account through MCP, with work orders, invoices and roofing projects among its examples.**Does not show:** Developer documentation, read October 2026.
- [Ontology overview](https://www.palantir.com/docs/foundry/ontology/overview)Palantir, 2026-10**Used for:** Palantir describing its Ontology as “an operational layer for the organization.”**Does not show:** Vendor documentation, read October 2026. The page is undated, so it does not show how long the category has existed.
- [AI workflow automation](https://n8n.io/ai/)n8n, 2026-10**Used for:** That at least one earlier category, workflow automation with language-model steps, is still sold as a living product.**Does not show:** A marketing page, read October 2026.

#### Investor publication

- [Generative AI’s Act Two](https://www.sequoiacap.com/article/generative-ai-act-two/)Sequoia Capital, 2023-09-20**Used for:** The 2023 account of the first wave of AI applications, the turn toward end-to-end products, and the view that data moats are weak while workflows and user networks hold up.**Does not show:** An investor’s 2023 view. It predates the 2026 market it is applied to here.
- [The future of services](https://www.generalcatalyst.com/stories/the-future-of-services)General Catalyst, 2025-08-28**Used for:** The services-versus-software comparison, the thesis of owning the service, and the investor’s own examples.**Does not show:** An investor’s thesis and its own operating figures. None of the portfolio numbers were audited here.
- [The new business of AI (and how it’s different from traditional software)](https://a16z.com/the-new-business-of-ai-and-how-its-different-from-traditional-software/)Andreessen Horowitz, 2020-02-16**Used for:** The 2020 margin and moat warning for AI companies, and the observation that some were bringing services in-house.**Does not show:** Written in 2020 about cloud infrastructure and human-support costs, not token prices. Not a 2026 audit.

#### Vendor-reported data

- [Clio’s ninth Legal Trends Report](https://www.clio.com/about/press/clio-latest-legal-trends-report/)Clio, 2024-10-07**Used for:** Software as a share of expenses by firm size, software spending against revenue growth, and the move to flat fees.**Does not show:** Law firms only, and shares of expenses, not revenue. We found no equivalent measurement for roofing contractors.
- [Highlights from the 2025 Legal Trends for Solo and Small Law Firms](https://www.clio.com/blog/solo-small-law-firms-highlights-2025-legal-trends/)Clio, 2025-05-16**Used for:** The average law firm spending “just under 2% of its overall expenses on software,” with solo firms spending about half of that.**Does not show:** A vendor’s blog summary of its own report, read through an archived copy. Law firms only, and a share of expenses, not revenue. We did not page the report itself.
- [Construction hiring and business outlook survey, national results](https://www.agc.org/sites/default/files/Files/Communications/2020_Outlook_Survey_National.pdf)Associated General Contractors of America, 2020 edition**Used for:** How construction firms answered “approximately what percent of your firm’s gross annual revenue is spent on IT?”**Does not show:** A member survey with 956 responses in total (54% general contractors, 35% specialty contractors), not roofing-specific. The survey notes that responses varied by question. The question appears in the 2018, 2019 and 2020 editions we opened.

#### Secondary analysis

- [Inside Harvey: how a first-year legal associate built one of Silicon Valley’s hottest startups](https://techcrunch.com/2025/11/14/inside-harvey-how-a-first-year-legal-associate-built-one-of-silicon-valleys-hottest-startups/)TechCrunch, 2025-11-14**Used for:** Harvey’s chief executive on the share of revenue from corporates, and on pricing: mostly seats now, a seat-based productivity suite for “at least the next year or two,” with outcome-based pricing starting with “very small things.”**Does not show:** Press interview. The figures are the executive’s own.
- [Harvey’s $15.5bn valuation: COO Katie Burke on what now and what next](https://legaltechnology.com/harveys-15-5bn-valuation-coo-katie-burke-on-what-now-and-what-next/)Legal IT Insider, 2026-09-17**Used for:** Harvey’s chief operating officer describing its pricing model in public.**Does not show:** Press interview. Read through an archived copy.
- [Harvey’s strategic evolution](https://clp.law.harvard.edu/ai-and-the-legal-profession/essays/harveys-strategic-evolution.html)Harvard Law School Center on the Legal Profession (Anthea Roberts and David Wilkins), 2026-04-03**Used for:** A two-author reading of Harvey as “a platform, infrastructure, and ecosystem play,” the founders’ choice to start broad instead of narrow, and the objection that a general model can already do the work.**Does not show:** Secondary analysis. The authors write that neither has tried Harvey, call the general-model objection “a serious argument,” and say whether vertical legal AI can hold its position is “one we genuinely don’t know the answer to.” Figures in its May 2026 update (valuations, user counts, joint ventures) are not used because the primary announcements were not opened.
- [General Catalyst’s $6.3B Amex deal puts its AI roll-up strategy on display](https://pitchbook.com/news/articles/general-catalysts-6-3b-amex-deal-puts-its-ai-roll-up-strategy-on-display)PitchBook, 2026-05-06**Used for:** General Catalyst’s $1.5 billion Creation fund in 2024, the agreed Amex GBT deal as of May 2026, and Eudia’s product.**Does not show:** Secondary reporting, published 6 May 2026, before the deal closed. The closing is confirmed in the company’s own release.

#### Commentary

- [Vertical AI is DEAD](https://youtu.be/Wc4C7iz3cA0)Devin Kearns (Custom AI Studio), YouTube, 2026-09-30**Used for:** The framing this paper tests: that point-solution AI built around one expensive workflow is a dead end, and that owning the operating business is the durable play.**Does not show:** A commentator’s argument and predictions, not evidence. We turned it into 24 checkable claims and graded each one.

### Downloads

The ledger and the roofing figures are published as data, so anyone can check them or reuse them.

- [Claim ledger (CSV)](https://foundationproject.net/research/point-solutions/claim-ledger.csv)
- [Claim ledger (JSON)](https://foundationproject.net/research/point-solutions/claim-ledger.json)
- [Roofing figures and derivations (CSV)](https://foundationproject.net/research/point-solutions/roofing-figures.csv)
- [Roofing figures and derivations (JSON)](https://foundationproject.net/research/point-solutions/roofing-figures.json)

The data is licensed under [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/): reuse it for any purpose, with credit to Foundation Projects.