Thirteen cards at Pass 1 fidelity, completing the cast. Each card is in its 10-moat-final state; the matrices in chapters 08 and 13 carry the analytical diff.
Archetype 2 · Vertical workflow co-pilot · Evaluator Power
HARVEY + SIERRA
The Calibrated Refusal Bundle
Primary · Evaluator PowerSecondary · Data FlywheelTertiary · Agentic Lock-in
Calibrated Refusal Quality
8.5
Accountability Surface
8.8
Domain-Data Depth
8.2
Outcome-Pricing Maturity
7.0
Workflow Embed Depth
7.5
Distribution Reach
5.5
Capital Reserves
8.0
Special move — Pricing the “no.” Harvey's value to AmLaw 100 firms is not the contract draft; it is the calibrated “this clause is non-standard, here is why, do not sign.” Sierra's value to enterprise CX is not the auto-resolve; it is the calibrated escalation when the agent should hand off. Both companies' Series-stage valuations (Harvey ~$8B, Sierra ~$10B / $150M ARR per The Information, 2025–2026) bake in outcome-bearing pricing surfaces that classical vertical SaaS could not reach. The bet is that the institutional accountability surface — license-bonded, contractually liable for being wrong — deepens with deployment in a way no foundation-model wrapper can reproduce.
Opponent
H1 (12mo)
H2 (3yr)
Mechanism / flip-condition
Foundation labs (Anthropic / OpenAI direct legal)
WIN
TIE
Foundation labs ship general legal capability; Harvey/Sierra retain calibration depth and accountability surface in regulated workflows. Tie at H2 if foundation labs build outcome-bearing partnerships with insurers/firms.
Microsoft Copilot Legal / Salesforce Einstein CX
TIE
TIE
Distribution incumbent matches feature surface; Harvey/Sierra hold via depth + accountability. Resolves to whose calibration is deeper at the daily-default position.
Capital + brand premium hold the AmLaw 100 / Fortune 500 accounts; peer competition resolves on whoever wins the largest enterprise calibration corpus first.
Bundling Harvey + Sierra is the right Pass-2 framing because both run the same Evaluator-Power mechanism in different verticals; the per-vertical battles will be unbundled in Pass 2b industry war games. Calibrated-refusal stat is the load-bearing dimension; Pilot and Trunk Tools cards carry the same dimension.
Archetype 3 · AI-native UX rebuild of an embedded workflow
Special move — UX so good the switch pays for itself in weeks. Cursor's attack on the procedural switching cost of leaving a JetBrains/VS Code workflow is mechanical: be 10× better at the core code-edit loop, and re-training amortizes inside a single sprint. By Q1 2026, Cursor reached $2B ARR — the fastest-scaling developer-tools product in history (per The Information). The contested question, and the reason the “Switching defender” tag is in tension with the “Switching attacker” primary moat, is whether Cursor can defend its own procedural switching cost against GitHub Copilot once Copilot closes the UX gap. The bet that turns this from arbitrage to moat is the Agentic Lock-in dimension: project-context memory, repo understanding, and learned tool wiring that Copilot enters at zero on. The honest answer is that the moat is Agentic Lock-in becoming a moat, not yet one.
Opponent
H1 (12mo)
H2 (3yr)
Mechanism / flip-condition
GitHub Copilot (MS distribution)
WIN
TIE
Cursor wins H1 on UX + iteration speed. H2 ties because Copilot closes the UX gap and brings distribution; outcome turns on whether Cursor's agentic context outpaces Copilot's distribution-amplified data.
Foundation labs (Claude Code, OpenAI Codex direct)
WIN
TIE
Frontier labs ship coding capability; Cursor retains workflow embedding. Ties if foundation labs ship competitive editor surfaces.
Replit / design-to-code collapsers
WIN
WIN
Different ICP and stack; Replit attacks the no-code edge of the same workflow.
The Cursor matchup is the cleanest test of whether 10×-UX-as-attack converts to durable moat. The Pass 2 stance: it converts only if the Agentic Lock-in dimension matures faster than the foundation-layer absorption. Per Pass 1.5 Q1, Cursor stays as its own AI-native archetype.
Primary · Agentic Lock-inSecondary · Data FlywheelTertiary · Evaluator Power (Decagon CX side)
Tool-Graph Memory Depth
7.8
Specialist-Node Depth (per company)
8.0
Escalation Calibration
7.5
Federated-Network Position
7.2
MCP-Portability Risk
6.5
Capital Reserves
7.5
Distribution (vs Salesforce)
4.0
Special move — Win the node, not the integration. The Pass 1.5 reframe is load-bearing here: these are not verticalized end-to-end stacks. They are specialist nodes (Decagon = CX agent depth; Crescendo = scaled service handoff; Clay = sales-data orchestration node) in a federated network bridged by translator/coordination layers built on MCP-style protocols and cheap inference. The translator layer commoditizes; the moat sits in the calibration of each node. The competitive question shifts from “can Decagon beat Salesforce end-to-end” to “can Decagon stay deeper at the escalation-calibration node than the surrounding network can route around.” That is a Data-Flywheel + Evaluator-Power question, not a Switching-Costs question.
Opponent
H1 (12mo)
H2 (3yr)
Mechanism / flip-condition
Salesforce Einstein / Now Assist (incumbent suites)
TIE
TIE
Distribution incumbents bundle agent capability into the suite; specialist nodes hold via depth where the work matters most. Ties on the daily-default position per node.
MCP-protocol portability maturation
WIN
TIE
H2 risk: protocol portability commoditizes the runtime. Holds for nodes if calibration depth outpaces protocol commoditization.
Foundation labs direct (Anthropic agent SDK; OpenAI Operators)
WIN
TIE
Foundation labs ship general agent capability; specialist nodes hold via vertical-specific calibration corpus.
The federated framing materially tightens the H2 call. Per the codesign-thesis claim #5, the integration layer is not where the value accrues; the node is. Pass 2b industry war games will unbundle Decagon, Crescendo, and Clay if the per-vertical analysis warrants.
AEC software contrast probe · uneven-decay incumbent (Revit moat erodes in simple-building markets, holds in commercial / healthcare / institutional)
AUTODESK
Sharp at the Edges, Intact in the Middle, Squeezed in Both
Dollar-Weighted Middle (structurally intact at H2)
8.2
Accountability-Stamping Layer (untouched at H2)
8.8
Special move — The cannibalization trap, fortified at the wrong perimeter, decaying unevenly. Autodesk's December 2025 tiered APS pricing, the EULA changes documented by AEC Magazine, and the transaction-model shift are all tells: incumbents now defend the API perimeter, not the file. But the perimeter under attack has moved past where Autodesk is defending, and the moat itself erodes unevenly across building types. Sharp decay in simple-building markets — tract single-family (Higharc already there), multifamily wood-frame, suburban warehouse, dark-store retail, modular — where generative BIM clears the production-ready bar by H2 because the constraint manifold is small and liability is pre-allocated to the developer/builder. Structurally intact in the dollar-weighted middle — commercial, healthcare, lab, institutional, complex residential, infrastructure — because cross-discipline coupling, jurisdiction-specific code interpretation, and licensed-professional accountability do not close in 3 years. The structural-exposure argument therefore pivots: it is not "generative BIM displaces Revit," it is "the API perimeter and the accountability-stamping layer are what is actually being defended," and Autodesk is fortifying the API perimeter (APS pricing) while leaving the accountability-stamping layer mostly unowned. Per chapter 13's loud claim 3, the bind tightens with every dollar of AI investment that lands on top of seat licensing. The deeper diagnostic — why the data flywheel claim is conditional, not realized — lives in chapter 10 section 6; the full segment-tagged strategy menu (data layer, design tools, construction, Tandem, adjacent) lives in the autodesk.json card.
Opponent
H1 (12mo)
H2 (3yr)
Mechanism / flip-condition
Speckle (project-graph attacker)
TIE
LOSE *
Asterisk (paper-wide #3): counter-positioning at the API perimeter under AI pressure. Autodesk wins H2 only by shipping a credibly free APS tier or acquiring Speckle ($300M–$1.5B) before mindshare escape velocity. The fight here is at the API perimeter, not at the generative-BIM displacement layer — the perimeter is where the dollar-weighted middle's defensibility actually lives.
Sharp decay at the simple-building edge: per-home pricing structurally cannibalizes seat licensing, and the typology (tract single-family) is exactly where generative BIM has already cleared the production-ready bar in 2026. Autodesk loses the residential-design slice (~2–5% of AEC segment revenue, LOW–MEDIUM strategic importance) and cannot acquire without rebuilding the seat-pricing model. Holds the dollar-weighted middle.
Snaptrude (concept-to-schematic generative)
TIE
TIE
In simple-building markets (multifamily wood-frame, warehouse, dark-store), Snaptrude can extend toward production-ready and chip residential / light-commercial seats. In the dollar-weighted middle, it stays concept-only because cross-discipline coupling, evaluation closure, and accountability transfer do not close in 3 years. Most likely outcome: Autodesk acquires Snaptrude ($200M–$700M, Spacemaker precedent) to neutralize the concept-side threat. The Revit documentation seat is unaffected.
Trunk Tools / Document Crunch (Evaluator-Power on layer-3)
TIE
TIE
Different layer (document-bound, not design-bound) — and per Priya's decomposition, layer-3 (accountability/stamping) is exactly the layer Autodesk does not own and is structurally untouched by generative-BIM at H2. Autodesk does not directly compete; it does control the data pipe Trunk needs through Procore-API risk and ACC integration.
Buildots / OpenSpace (field-capture flywheel)
WIN
TIE
Field-capture compounds independently of Autodesk's design seat; PlanGrid→Autodesk acquisition precedent ($875M, 2018) suggests the standard playbook is acquire-and-bundle. Convergence at H2 only if the flywheel reaches escape velocity into design feedback before Autodesk decides to acquire.
The structural-exposure flip per Priya's decomposition: if any insurer underwrites E&O on AI-assisted code-compliance review by 2026–2028, the layer-3 accountability moat moves from "owned by the licensed professional via Autodesk-stamped Revit" to "owned by the insurer-underwritten Evaluator-Power vendor." Autodesk loses the structurally-intact-middle's defensibility because the layer that holds the middle is the accountability layer, not the seat license. Inflection signal #1 in the tradeoff drawer.
Tradeoff drawer — how the Revit moat actually erodes through H2
The H2 verdict is uneven decay, not binary collapse. By H2 (3 years), generative text/sketch → production-ready BIM clears the bar in simple-building markets (tract single-family already there in 2026; plausibly multifamily wood-frame, suburban warehouse, dark-store retail, modular by 2029) where the constraint manifold is small, cross-discipline coupling is light, and liability is pre-allocated to the developer/builder. It does not clear the bar in the dollar-weighted middle (commercial, healthcare, lab, institutional, complex residential, infrastructure) because cross-discipline coupling, evaluation closure, accountability transfer, and tool-graph maturity do not close in 3 years. (Priya's full analysis: priya-handoff-autodesk-genai-bim.)
Evaluator Power decomposes into three layers that erode at different rates:
Layer 1 — Mechanical evaluation (clash detection, code-rule passes, schedule reconciliation). Closing fast; commoditized by H2. Solibri-class problems get solved by 2027–2028. Not defensible.
Layer 2 — Engineering judgment (does this column grid work for this site, will this MEP layout be installable, is this detail constructible). Contested by H2. Domain-specific models compete with senior engineers' tacit knowledge. Defensible through H2 for whoever owns the cross-firm decision corpus — which is why the HarveyCo-AEC archetype matters.
Layer 3 — Accountability and stamping (does a licensed professional sign, does insurance underwrite). Essentially untouched at H2. Institutional change, not technical change. The structurally durable layer.
The structural-exposure argument pivots: not "generative BIM displaces Revit," but "the API perimeter and the accountability-stamping layer is what is actually being defended." Autodesk's structural exposure is on layer-2 (engineering judgment, where domain models compete with cross-firm corpora), not on layer-1 (which commoditizes anyway). The Revit moat erodes unevenly: sharp at simple-building markets, structurally intact in the dollar-weighted middle, with the middle itself being squeezed harder than Autodesk's pricing model can absorb.
Inflection signal #1 — insurer underwriting: whether any insurer (Travelers, Berkley, Beazley) underwrites E&O on AI-authored or AI-assisted design output by 2026, even narrowly. This is the load-bearing forward signal. If yes, the layer-3 accountability moat moves out from the licensed-professional-and-Autodesk-stamped-Revit pairing toward an insurer-underwritten Evaluator-Power vendor — and the dollar-weighted middle stops being structurally intact. Autodesk loses the layer that actually holds the middle. If no through 2028, the simple-vs-complex split hardens and Autodesk holds the middle through H2.
What Autodesk should do, per the framework: ship the credibly open project-graph layer voluntarily at margin cost, and re-anchor monetization on the layer-3 accountability-stamping surface (Revit-as-stamped-artifact, plug-in ecosystem, plan-reviewer familiarity, audit logs). The pricing-model fracture is real but smaller than the cannibalization the seat-only path delivers over 5 years. They probably will not. The counter-positioning trap (chapter 03; chapter 13's loud claim 3) tightens with every dollar of AI investment that lands on top of the seat license.
Segment-by-segment strategy menu — the 21 moves available, structured by where the war actually lives:
Data layer. Open the project graph at margin cost (free APS read/write tier, drops own AI-capex pressure, neutralizes loud claim 3). Ship AEC Data Model API geometry plus Civil 3D plus Forma natives by 2027 to remove the strongest reason ISVs route through Speckle. Add an opt-in customer-data contribution tier (the Adobe Firefly precedent) to resolve the corpus-rights ambiguity that the December 2025 ToS revision only half-addressed — this is the single move that flips flywheel_threshold_status from "below" to "approaching." Tier developer access by AI use case (internal / commercial / training SKUs) to legalize and price the December 2025 ambiguity rather than litigate it.
Design tools. Forma Connected Client (already executed AU 2025) keeps Revit as desktop satellite to the cloud. Productize Neural CAD for Buildings into a shipping Forma feature by mid-2027 — the foundation-model moat startups cannot match on training data plus compute. Acquire Snaptrude ($150–500M plausible band) to neutralize the most credible "Figma for BIM" attacker. Persona-based unbundling of Revit per Bunszel's AU 2025 telegraph — raises ARPU even as the monolith fragments.
Construction. Bundle Forma Data Management free (already executed March 2026) to strip Procore's CDE upsell. Wire the Rhumbix → Payapps → Forma Build cost spine to collapse the timecard-to-payment cycle Procore cannot match without owning payments. Acquire a small AI-agent vendor as Datagrid counter ($50–200M). Acquire OpenSpace or Buildots to close the visual-intelligence gap that is currently partner-only.
Tandem. Roll out Tandem Insights AI to FM teams as the wedge that converts the consultancy line item to seat-licensed standard. Tuck-in Invicara/Twinview or Cohesion ($100–300M) to leapfrog Bentley iTwin's lead in non-BIM-native asset capture. Eaton plus Forma plus Tandem co-marketed bundle into data centers.
Adjacent. Industrialized-construction crossover via Fusion plus Informed Design plus Forma Building Design — a category attackers cannot credibly enter without manufacturing CAD heritage. Acquire an IC kit-of-parts platform (Cuby/Assembly OSM scale, ~$100–400M) to ship a "Manufacturable BIM" certification mark. Deepen the World Labs integration (built on the $200M February 2026 stake) as the spatial-AI hedge against external foundation-model breakthrough.
Insurer-underwritten accountability stack remains the highest-leverage cross-segment defensive move — partner with Travelers/Berkley/Beazley to underwrite E&O on Revit-stamped, AI-assisted design output, capturing the layer-3 accountability moat before HarveyCo-AEC does. The unifying diagnostic: the load-bearing Autodesk moat is licensed-professional accountability plus cross-discipline coupling, productized into a metered API perimeter and a foundation-model defense layer. The flywheel-shaped tailpiece is conditional on resolving customer-data training rights at scale — until then, Autodesk is building the artifact, not the mechanism.
AEC software contrast probe · the canonical “active moat-construction via terms-of-use + consumption pricing + agentic infrastructure” case in AI-era construction software
AI Capex Pressure (partly absorbed by token surface)
7.0
Capital Reserves ($591M cash, $215M FCF)
7.5
Special move — The default-GC-stack moat-construction trinity. Procore is the system-of-record for construction at top-200 US GCs with 95% gross retention — Switching Costs (Helmer #4) is the durable floor. The H1 amplifier is the September 2025 → April 2026 moat-construction trinity: (a) Sep 30 2025 Developer Policy bans bulk data export including LLM training, (b) Managed Marketplace certifies and gates third-party data access, (c) March 2026 Agentic API meters paid access via consumption pricing. Together they convert customer data into a Cornered Resource (Helmer #6) by terms-of-use — a 21st-century update to Helmer's classical examples (De-Beers-class exclusive supply replaced with platform terms-of-use as the gating mechanism). Unlike Autodesk's December 2025 APS metering (which fortified the OLD perimeter — file-based Revit), Procore's trinity is purpose-built for the AI era: it gates the data the agents need. Loud Claim 3 is partly resolved by the Agentic API consumption pricing — it carves out a NEW pricing surface (token billing) that absorbs AI capex instead of layering it on the seat license, which Autodesk has not pulled off. The CEO transition is the underplayed catalyst: Ajei Gopal (ex-Ansys CEO 2017–2025, tripled revenue to $35B Synopsys exit) Nov 10 2025; Walt Hearn (ex-Ansys WW Sales) CRO + Rachel Pyles CFO Apr 1 2026 — the Ansys public-software-CEO playbook installed at Procore. Datagrid acquired Jan 16 2026 for $168M cash (disclosed only in Q1 10-Q purchase-price allocation, never press-released; goodwill $114.9M) is the agentic backbone — ironic note for the moats paper, Carl Bass (ex-Autodesk CEO) was a Datagrid angel and Autodesk participated in a $22M Datagrid round, so the agentic core Procore now wields against Autodesk and Trunk Tools was partly Autodesk-funded. The Trunk Tools dispute (API revoked Sep 2025, Marketplace denied Oct 2025) is the visible enforcement case — selectively-targeted not industry-wide, the policy is a standing weapon Procore can deploy without public escalation. The deeper diagnostics: why the data flywheel is below threshold not above lives in chapter 10 section 6; why Agent Studio adoption is the canonical H1 test for the Agentic Workflow Lock-in archetype lives in chapter 11; why Sep 2025 Developer Policy is the canonical 21st-century cornered-resource case lives in chapter 6; the full segment-tagged strategy menu (data layer, agentic infrastructure, construction financial OS, vertical expansion, BIM extension, cross-segment) lives in the procore.json card.
Opponent
H1 (12mo)
H2 (3yr)
Mechanism / flip-condition
Autodesk (paired AEC platform-incumbent)
WIN
TIE *
Asterisk (paper-wide #4): the Procore-vs-Autodesk H2 endgame has genuine 3-body complexity — Procore independent vs. PE-take-private vs. Microsoft acquisition. H1 verdict YES per opportunity-thesis.md: Procore shipping faster on agentic AI (Datagrid integrated, Agentic API GA late March 2026, Agent Studio in development) than Autodesk on Neural CAD (production-ready mid-2027 per Autodesk A2). Resolves Autodesk Critical Assumption A3. H2 verdict AMBIGUOUS: most likely parallel platforms with détente at the bundling layer + cold war at the data perimeter (Procore Sep 2025 + Autodesk APS Dec 2025 = industry-wide pattern). Forma Data Management bundled FREE attacks Procore CDE upsell (Risk 3) but Procore Pay + Levelset financial-workflow lane is genuinely differentiated.
Trunk Tools (the squeeze, working as designed)
WIN
WIN
Sep 2025 Developer Policy + Oct 2025 Marketplace denial = pre-emptive moat denial. Selectively revoke API access; deny Marketplace status; force workarounds (PDF export, customer-side ingestion via Egnyte/Box). Trunk Tools' H1 outlook materially worse than April 2026 research suggested — must convert Gilbane / Suffolk / DPR enterprise references into a non-Procore-dependent business model before Q4 2026 or be acquired by a Procore competitor (Autodesk / Microsoft) at distressed prices. Procore does not buy because (a) hostile relationship after API revocation, (b) Datagrid renders Trunk Tools redundant, (c) acquiring would validate "custom LLMs > Procore-bundled AI" thesis.
Primepoint (drawing-intelligence knowledge graph)
WIN
TIE *
Procore has the integration + Datagrid backbone; Primepoint has the AI-native team (Bourdev ex-FAIR, Palo ex-Trello) + drawing-intelligence knowledge graph. If Procore parity ships in H1, Primepoint compresses; if not, compounds past threshold. Most likely H2 outcome: Procore acquires at $40–80M (Datagrid 4.4×-raised pattern). Friendly mechanics — Primepoint already integrates Procore + ACC.
Trimble (post-Document Crunch + Neuron Factory)
WIN
TIE
Trimble bought Document Crunch Apr 2026 + invested in Neuron Factory (knowledge graph) — positioning Trimble Construction One as “third-platform” answer to Procore Datagrid. H1: Trimble distracted by portfolio complexity ($3.59B revenue declining -2.6% FY25). H2: Trimble + Document Crunch + Neuron Factory could combine into credible third platform by mid-2027 but likely lags both Procore and Autodesk in execution velocity.
Different data modality — reality capture / visual-progress vs. document/financial. Hexagon parent ($15B) consolidated this layer post-Avvir; 60+ territory franchise distribution Procore lacks. Coexistence more than competition near-term; H2 risk if visual-AI + agentic converge in a way Procore can't replicate without acquiring a Buildots/OpenSpace-class player ($300–500M plausible band).
Microsoft (most likely H2 strategic acquirer)
TIE
TIE *
Microsoft is not a direct competitor today — but is the most likely strategic acquirer of Procore at H2 (5–10% probability over 5 years, would pay 7–9× revenue $10–13B for Dynamics 365-AEC vertical). Until then: M365 + Teams + Copilot is the horizontal agent-runtime that Procore's Agentic API must federate with via MCP rather than try to replace. Asterisk because the H2 endgame includes “Microsoft acquires Procore” as a non-trivial branch in the 3-body complexity (paper-wide asterisk #4).
Nemetschek (post-HCSS rollup)
WIN
TIE
Post-HCSS rollup ($2.4B Apr 2026, ~11× revenue — high-water mark for AEC software M&A), Nemetschek is now a fifth major buyer alongside Procore/Autodesk/Trimble/Oracle. Bluebeam in portfolio. More strategically positioned as Snaptrude's better acquirer than as Procore's direct competitor; H2 risk only if rollup-strategy consolidates AI-natives into a third-platform threat.
Higharc (different segment)
TIE
LOSE
Different segment — Higharc is homebuilder configurator (production residential), Procore is GC platform (commercial / institutional). Procore can't follow into per-home pricing without breaking its seat model. No direct competition; Higharc wins H2 by default in its segment.
Tradeoff drawer — the moat-construction trinity, motivation chains, and the Q3 2026 visible test
The H1 (12mo) verdict is binary on Q3 2026 AI revenue rollout. Two scenarios cover 65% of outcomes per the war game in Companies/Procore/competitive-landscape.md: Scenario 1 (35%) — Q3 2026 earnings shows AI revenue contribution >2% of total, Datagrid-derived agents have 200+ enterprise deployments, multiple expands toward 7–8× EV/Rev, stock recovers to $95–120 range. Scenario 2 (30%) — Q3 2026 earnings shows AI still “immaterial” or sub-2%, Agent Studio stuck in beta, multiple stays compressed at 5–5.4×, stock stays in $50–70 range, Procore pivots to operator-CEO playbook of margin extraction. Everything else (Scenarios 3–5: Trunk Tools acquired by Microsoft/Autodesk, training-rights challenge, PE LOI) covers the remaining 35% as H2 modifiers.
Three load-bearing assumptions carry the entire thesis:
PA1 — Customer-data training rights resolve in Procore's favor by Q4 2026. Without favorable training rights, Datagrid has no privileged data flywheel — just an architecture choice. Best guess: 65% probability favorable, most likely a “soft revision” mirroring Autodesk Dec 2025 ToS. Validation: Procore enterprise MSA / DPA leak; AGC/NDIA published position; any GC-coalition pushback by mid-2026.
PA2 — Agent Studio ships AND customers actually build/adopt no-code agents on Procore data. THE H1 agentic lock-in mechanism. If customers don't build, switching costs stay at product-feature level — moderate but not extreme. No-code AI builders have a poor adoption track record in horizontal SaaS (Salesforce Einstein, Microsoft Power Platform, Workday precedents). Best guess: 50% probability of meaningful adoption by Q4 2026, most likely 5–15 enterprise customers building 2–3 agents each by Groundbreak 2026. Validation: Groundbreak 2026 customer demonstrations; Q3 2026 earnings disclosure of customer-built agent counts.
PA3 — Q3 2026 sales rollout produces visible AI revenue (>2% of revenue contribution by Q4 2026). Visible commercial proof point. Q1 2026 was “immaterial.” Walt Hearn's WW-sales playbook from Ansys built for exactly this operationalization. Best guess: AI revenue Q4 2026 ~1.5–3.5% of revenue — 40% probability clearly-validating, 50% ambiguous, 10% clearly-disconfirming.Validation: Q2 2026 earnings (Aug 2026) leading indicators; Q3 2026 earnings (Nov 2026) direct disclosure.
Motivation chain — why the moat-construction trinity exists, and what it actually buys:
Sep 30 2025 Developer Policy bans bulk data export for commercial purposes including LLM training →
Managed Marketplace certifies and gates which third parties get data access (Trunk Tools denied; Marketplace re-vetting through FY26) →
March 2026 Agentic API consumption pricing meters paid access to remaining data →
Datagrid agents + Agent Studio capture workflows on Procore data →
The trinity only compounds if all three pieces hold. Each piece has a real failure mode in the next 12 months: PA1 (training rights), PA2 (Agent Studio adoption), PA3 (Q3 2026 revenue). The moat is contingent.
Segment-by-segment strategy menu — the 21 moves available, structured by where the war actually lives:
Data layer (the trinity itself). Sep 30 2025 Developer Policy already enforced. Managed Marketplace certifying third parties through FY26. Agentic API consumption pricing GA late March 2026. The single move that resolves PA1 cleanly is the opt-in customer-data training tier (Adobe Firefly precedent) — flips flywheel_threshold_status from “below” to “approaching.” Without it, the trinity is one lawsuit or industry-coalition statement away from unraveling.
Agentic infrastructure (the H1 bet). Datagrid integration shipped (contract review agent <30 days post-close). Procore Agents (RFI/submittal/compliance/contract-review) GA on Datagrid backbone. Agent Studio is THE PA2 bet — ship AND adoption needed; no-code AI builder track record in horizontal SaaS is poor. Q3 2026 broader sales rollout = the PA3 visible commercial test. MCP-addressable agent runtime federation as the defensive consolation prize if the perimeter cracks.
Construction financial OS (the differentiated lane vs. Autodesk). Procore Pay + Levelset deepening — $1.4T/year flowing through GC AP/AR; even 0.1% fee capture = $1.4B revenue lane eventually. Tuck-in payments / construction-fintech adjacency (insurance-claim automation, surety bonds, construction-specific factoring; $50–200M plausible). Bundle Field Scheduling AI / Safety Hub / Insights into core. Autodesk has no fintech/payments capability; Trimble doesn't either — this is the structural differentiator.
Vertical expansion (the lanes that don't depend on AI thesis crystallizing). FedRAMP Moderate authorization Jan 29 2026 unlocks Federal construction TAM ~$300B+. NVIDIA Omniverse / DSX integration Mar 16 2026 stakes the AI-factory / data-center construction vertical ($600B+ buildout in flight) — NVIDIA itself uses Procore for Vera Rubin DSX construction. Hyperscaler co-sell (Microsoft / Google / Meta / Amazon) operationalizes the partnership; 1–2 named hyperscaler customers by Groundbreak 2026 = K6 trigger.
BIM extension (defending vs. Autodesk Forma). Novorender (Jan 2025, $44.3M) + FlyPaper Sherlock (May 2025) give Procore a credible BIM viewer + clash detection layer. Acquire Primepoint ($40–80M plausible per Datagrid 4.4×-raised pattern) to close the “construction knowledge graph” position vs. Trimble Neuron Factory + CONXAI. Acquire Buildots/OpenSpace/Doxel ($300–500M+) to close the visual-progress / reality-capture data lane — defensive move because if Trimble or Microsoft acquires one, Procore loses the lane.
Cross-segment. Trunk Tools squeeze (already executing as designed — pre-emptive moat denial via Sep 2025 policy). International expansion 15% → 25% by 2027 (UK / Germany / Australia / Middle East priorities). Operator-CEO margin expansion playbook (Ansys precedent) is the floor under Procore valuation regardless of AI moat outcome — FY26 guide 18–18.5% non-GAAP op margin (390–440 bps expansion).
The unifying diagnostic: Procore + Autodesk are paired cases of the same lesson — incumbent moats in AI-era AEC are constructed via terms-of-use + consumption pricing + agentic infrastructure, not assumed via legacy data positions. Autodesk fortified the OLD perimeter (file-based Revit) with December 2025 APS metering; Procore built a NEW perimeter (data-pipe-as-cornered-resource) purpose-built for the AI era. Both face the same load-bearing question: can customer-data training rights resolve favorably under the Sep 2025 / Dec 2025 ToS framing? If yes, both incumbents cross flywheel threshold and earn multiple expansion. If no — or if forced through a soft revision under industry pushback — the moat-construction trinity is exposed as friction without lock-in. For the moats paper, the two cards together carry the AEC platform-incumbent narrative: the anatomy of the play is the contribution, and Q3 2026 is the visible test for both.
AEC archetype A · Contract / commercial-relationship disruption
Special move — Attack the commercial structure, not the design tool. Higharc's attack vector on production homebuilders is a textbook counter-position: redefine the buyer ↔ builder ↔ contractor commercial structure (configurator-driven, options-priced, design-locked at sale) so that the legacy chain — CAD seats + spec-home design firm + change-order revenue — cannot ship Higharc's product without cannibalizing the change-order-revenue line that funds it. Per the Pass V loud claim (b), this is also a moat that strengthens with incumbent AI investment, because every dollar Autodesk spends on generative design lands on top of the legacy seat-pricing model and accelerates the cannibalization Autodesk already cannot afford.
Opponent
H1 (12mo)
H2 (3yr)
Mechanism / flip-condition
Autodesk (design seat in homebuilder)
WIN
WIN
Counter-position bind structural; Autodesk cannot ship Higharc's product without rebuilding the seat-pricing model.
Builder ERPs hold the back-office; Higharc encroaches at the design + sales front. Tie at H1; Higharc wins H2 if it stays deep at the configurator+sales node.
Foundation labs direct (Claude Design / OpenAI canvas)
WIN
WIN
Per codesign-thesis claim #1: frontier canvases cannot reason across physics, code, liability for production homebuilding. Higharc holds the constraint surface.
Higharc commercial-disruption metrics are [Pass 3 deep-dive will sharpen] — current public press is heavy on funding rounds, light on per-builder change-order-revenue impact.
AEC archetype C (elevated) · Document / spec / risk AI agent · Evaluator Power exemplar
TRUNK TOOLS
The 72% Correct Refusal
Primary · Evaluator PowerSecondary · Data Flywheel (custom construction LLM)Tertiary · Switching (Procore-API risk)
Calibrated-Refusal Rate (Gilbane benchmark)
7.2
Accountability Surface
7.0
Custom-LLM Defensibility (claimed)
6.5
Founder-Market Fit
9.2
Enterprise GC Logo Depth
7.8
Procore Platform-Risk Exposure
8.0
Capital Reserves ($70M Series B)
5.5
Outcome-Pricing Maturity
5.0
Special move — The 72% non-compliance flag at Gilbane. TrunkSubmittal's Gilbane deployment reviewed 2,000+ submittals in three months and flagged 72% as non-compliant, shortening submittal cycle times by ~50% (per company press, ENR coverage 2025). The Pass 1.5 Evaluator-Power read is that the value is not the 28% it processed cleanly — it is the 72% it correctly refused. The customer pays for the calibrated “no” on a contract clause, a spec section, an RFI cross-reference. Founder Sarah Buchner's rare construction-plus-data-science background (Stanford GSB; PhD Civil Engineering & Data Science; carpenter at 12) is itself the hardest-to-replicate input. The Insight Partners-led $40M Series B (July 2025) priced this thesis explicitly.
Opponent
H1 (12mo)
H2 (3yr)
Mechanism / flip-condition
Procore native AI Copilot + Agent Studio
TIE
TIE
Procore controls the data pipe (API access revoked September 2025; resolution unknown). Trunk Tools holds via depth + custom-LLM if Microsoft partnership provides alternative distribution channel.
Document Crunch (peer in archetype C)
TIE
TIE
Different specialty (Document Crunch heavier on contract review; Trunk on submittals + revisions). Coexistence in distinct sub-segments more likely than displacement.
Foundation labs / fine-tuned GPT-4-class on construction docs
WIN
TIE
Custom-LLM advantage on construction-specific documents; ties at H2 if fine-tuned frontier models close the gap (Buchner's assumption A1 in deep-dive).
Gilbane / Suffolk in-house dev (build vs buy)
WIN
WIN
GCs deploy, do not build. Customer-not-competitor.
The Procore API revocation (September 2025) is the load-bearing risk. Resolution status is unknown as of April 2026; the deep-dive's assumption A5 carries this card's H2 calls. [Pass 3 deep-dive will sharpen] on the resolution and on independent benchmarks of custom-construction-LLM accuracy vs. fine-tuned frontier models.
AEC archetype D · Field-capture data flywheel · thesis-grade not benchmarked
Special move — A flywheel thesis without a threshold benchmark. Per Pass 1 Q3 discipline: Buildots and OpenSpace publish customer counts and case studies, but no credible third-party benchmark exists on (a) how much proprietary data each flywheel actually accumulates per customer per year, and (b) whether that data has decreasing or increasing marginal value past year 1. OpenSpace's 75,000+ project corpus of visual-to-BIM comparisons is the strongest single claim in the archetype; Buildots leans on Skanska / Suffolk / Turner deployment depth. I'd update toward "moat is real" if I saw: a third-party-validated learning curve per customer, year-over-year accuracy improvement attributable to corpus depth, and a credible argument that the jobsite-data corpus is hard to replicate from a frontier vision model. I'd update toward "moat is thin" if I saw: frontier vision models achieving comparable visual-to-BIM accuracy on public construction-imagery corpora.
Opponent
H1 (12mo)
H2 (3yr)
Mechanism / flip-condition
Foundation vision models direct (GPT-4V class on construction)
Procore native field-capture / Autodesk Construction Cloud
WIN
TIE
Distribution incumbents bundle field-capture; Buildots/OpenSpace hold via depth + customer integration cost.
Living-graph competitor (codesign claim #4)
TIE
LOSE
If a "construction-end living-graph" archetype emerges (codesign #4), today's frozen-at-handover field-capture loses to the player who extends data flow into operations. Disconfirming-signal disclosure: this is the most likely failure mode of the archetype as currently scoped.
Per Pass 1 Q3, this card is "thesis-grade, not benchmarked." Marginal-data-decay and flywheel-velocity stats carry "?" suffixes deliberately. [Pass 3 deep-dive will sharpen] if a third-party benchmark on per-customer data accumulation appears.
Special move — Scale leveraged into the AI-infrastructure buildout. Turner doubled data-center revenue in 2025 (per Data Center Dynamics / company disclosures). The AI-infrastructure compression of timelines — hyperscaler $650–690B 2026 capex; Microsoft's ~$80B in unfulfillable Azure orders due to power constraints — is a textbook scale-economies tailwind: the segment most rewards the GC who can mobilize bonding capacity, large-project process, and supply-chain leverage at the largest single-site projects of the era. Turner Ventures (CVC) is the partial hedge against process-power and data-flywheel attack from below; the bet is mostly on capturing more of the buildout than peers, not on being the AI-native GC.
Opponent
H1 (12mo)
H2 (3yr)
Mechanism / flip-condition
Suffolk (tech-leaning peer)
TIE
TIE
Tech bet vs scale bet; resolves to which strategy wins on the H2 buildout cycle.
DPR (process-power peer)
TIE
TIE
Different moat type, different ICP slice; coexistence likely.
AI-native preconstruction startups
WIN
TIE
Estimating commoditizes; bonding capacity and scale hold the mega-project segment.
Owner-led IPD on hyperscaler datacenter portfolios
TIE
LOSE
Hyperscalers with repeat-volume datacenter portfolios are the most likely first wave to attempt owner-led delivery; Turner's scale moat thins to specialty-only on the irregular projects.
Data-center revenue concentration is a tailwind in H1 and a concentration risk if the buildout pace breaks. [Pass 3 deep-dive will sharpen] on Turner's share of hyperscaler datacenter spend vs. peers.
AEC general contractor · process power + employee-owned culture
Special move — Patience as a moat ingredient. DPR's process power runs on the longest possible clock: employee-ownership, lean construction depth across self-perform trades, and culture that does not optimize for quarterly visible results. Per chapter 07, process power compounds on a 5–20 year horizon and decays under public-company quarterly pressure. The honest pressure-test: is the patience itself doing the work, or is it the talent and lean depth, with patience as a permissive condition? The argument that holds: patience is what allows the talent-and-lean-depth to compound without being optimized away. The argument against: every employee-owned firm claims this; not all of them are DPR. The decomposability score (3.8 out of 10) is the AI-era stress-test — if AI decomposes lean-construction know-how into model-trainable patterns, DPR's moat decays even with patience intact. Per Part V's flipped Process-vs-Data-Flywheel cell, this is the variable that matters.
Opponent
H1 (12mo)
H2 (3yr)
Mechanism / flip-condition
Suffolk / Turner peer GCs
TIE
TIE
Different moat profiles; per-niche outcomes vary (DPR strong in healthcare / advanced industrial; peers stronger in commercial / data-center).
AI-native preconstruction + scheduling startups
WIN
TIE
AI commoditizes specific decomposable tasks; ties at H2 only if AI decomposition reaches DPR's self-perform process layer.
DPR's healthcare and lab moat is partially a complexity moat that resists owner-led; hyperscaler datacenter exposure is lower than Turner's.
Private-company-patience as moat is a defensible argument but not a measurable one. [Pass 3 deep-dive will sharpen] with self-perform-share data and lean-deployment metrics relative to peer GCs.
Accounting · managed-bookkeeping margin as Evaluator-Power exemplar (Pass 1.5 retag)
Special move — The customer pays for the review, not the AI. Pilot is the cleanest accounting-side test of Evaluator Judgment Power as a pricing-surface moat. The customer is paying for accountability-bearing review — the CPA who signs the books, the firm that stands behind the close, the calibrated “this transaction is mis-categorized” on a noisy ledger. The AI doing the books underneath compresses cost; the margin sits with the entity that has the institutional standing to certify. Pass 1.5 retagged Pilot from generic services-led-growth to Evaluator-Power because the pricing surface (managed-service margin) is the moat dimension being defended. Per Part IV chapter 12, this surface composes naturally with a Data Flywheel (multi-customer ledger patterns) and an Agentic Lock-in dimension once Pilot ships AI agents into the bookkeeper's daily loop.
Intuit's distribution dominates H1 in self-serve; Pilot's Evaluator surface holds in segments where the customer wants liability transfer.
Brex+Ramp+Puzzle stack (AI-native unbundled)
TIE
TIE
Different ICP today (AI-native startups vs. assisted-bookkeeping mid-market); the segments overlap as Brex stack moves up.
Foundation labs direct (ChatGPT Bookkeeper)
WIN
TIE
Liability transfer is the moat; foundation labs do not stand behind the books. Holds while liability stays with the certifier.
Big-Four mid-market push
WIN
TIE
Big Four institutional accountability vs. Pilot's tech-leveraged accountability; resolves to whose calibration is deeper at managed-bookkeeping ICP.
Pilot pricing-experiment maturity vs. share-of-savings, metered, insurance-linked precedents per codesign-thesis claim #3 is [Pass 3 deep-dive will sharpen].
Special move — Bundled flywheel as counter-position to seat pricing. The stack's collective attack on Intuit is a textbook counter-position: spend management (Brex/Ramp) plus AI-native ledger (Puzzle) replaces QuickBooks for the AI-native cohort, priced through interchange + spend rebates rather than per-seat. Intuit cannot match without breaking its own QuickBooks economics. Compounding inside the stack is a multi-flywheel: every payment runs through Brex/Ramp; every transaction lands in Puzzle's AI-native ledger; the close-the-books agent inside Puzzle accumulates calibration on those rails. The Pass 1.5 reframe is that the Intuit-vs-stack matchup at H2 is one of the four paper-wide asterisk candidates, but the matchup that lives in the asterisk is Intuit's side (see Intuit card); from the stack's side, this is a clean H2 win in the AI-native segment.
Opponent
H1 (12mo)
H2 (3yr)
Mechanism / flip-condition
Intuit (QuickBooks SMB + IES)
TIE
WIN
Bundled flywheel + AI-native ledger beats accountant-channel brand at the AI-native startup cohort. Holds for Intuit only if accountant-influenced share keeps compounding into lower-mid-market.
NetSuite (Oracle pushing down)
WIN
WIN
Different segment; NetSuite's mid-market complexity cap means the stack holds the SMB AI-native cohort.
Pilot (Evaluator-Power play)
TIE
TIE
Different ICP today; Pilot's liability-transfer surface is what the stack does not offer. Coexistence likely.
Stack-coherence stat (6.5) reflects the structural reality that Brex, Ramp, and Puzzle are three companies, not one. Coordination friction is the load-bearing risk. [Pass 3 deep-dive will sharpen] on Puzzle ARR and stack-customer overlap rates.
Cross-cutting test case · network defender under AI-native attack
Special move — Multiplayer collaboration as network moat under AI pressure. Figma is the most public test case of the central question this paper sets: is multiplayer collaboration a moat against AI-native attackers? The April 2026 ~14% stock drop on AI-moat concerns priced this question explicitly. The structural read from chapters 02 and 11: Figma's direct + indirect network effect is real and dense; the procedural switching cost on designer skill is real but compressible by 10× UX rebuilds (the Cursor pattern); the data-flywheel dimension is shallow because Figma does not aggregate cross-customer design corpora the way OpenAI/Anthropic/Google do for general-purpose generation. The vulnerability is not network density — it is that the network is mostly multiplayer-collaboration value, not data-network value, in a category where AI is moving the work upstream toward generation-from-prompt.
Opponent
H1 (12mo)
H2 (3yr)
Mechanism / flip-condition
Adobe (capital-rich incumbent)
WIN
TIE
Multiplayer network holds H1; Adobe's catch-up depends on closing collaboration UX gap and bundling AI generation aggressively.
Google Stitch (bundled-free attacker)
TIE
TIE
Distribution arbitrage from Google; Figma holds via designer brand and procedural switching, but the seat tier compresses.
Cursor / Replit / design-to-code collapsers
WIN
TIE
Different layer (code-first); the threat is cross-disciplinary collapse of design-to-code that bypasses the designer + Figma seat. Holds at H2 only if design remains a distinct upstream surface.
Frontier labs direct (Anthropic Claude Design, OpenAI canvas)
WIN
TIE
Per codesign claim #1, frontier canvases ship UI design but not coupled-constraint reasoning; Figma holds in the production-grade design layer until canvases get deeper.
The stock's ~14% April 2026 drop is a narrative signal, not a fundamental verdict. [Pass 3 deep-dive will sharpen] on Figma's AI-feature adoption rates and seat-tier compression vs. multiplayer-tier durability.
Cross-cutting test case · "what AI does not compress" foil
Special move — The moat AI does not directly attack. Plaid's primary moat is what regulators, banks, and time built: thousands of bank integrations under continuous compliance, with relational trust accumulated over a decade of institutional negotiations. AI does not move this needle. A frontier model cannot manufacture the bank-relationship credentials, the compliance surface, or the institutional trust that Plaid has bonded. The card's purpose in the cast is contrast: it shows what survives the AI era unchanged. The argument is reinforced when read alongside the Speckle and Cursor cards — the moats most exposed to AI attack are the ones bonded to procedural and analytic work; the moats most resistant are the ones bonded to regulated, relational, embedded structure. The risk axis is not AI; it is open-banking regulation (US Section 1033 implementation) that could dilute the cornered-resource layer if the regulator forces standardized bank API access.
Opponent
H1 (12mo)
H2 (3yr)
Mechanism / flip-condition
AI-native fintech direct (Brex / Ramp banking layer)
WIN
WIN
Plaid sits below them in the stack; they are customers, not competitors.
Same moat type; coverage and developer-side preference vary by ICP.
US Section 1033 / open-banking rule changes
WIN
TIE
If the regulator forces standardized bank APIs, Plaid's coverage moat dilutes; relational moat holds. The H2 risk is regulatory, not AI.
Foundation labs / general AI
WIN
WIN
Foundation labs do not have bank-integration relationships or compliance surface; AI does not compress this moat.
The contrast point is the chapter's purpose; Plaid is in the cast as the foil for "what AI does not compress." [Pass 3 deep-dive will sharpen] on US Section 1033 implementation timing if it shifts the H2 call.
Primary · Counter-PositionSecondary · Agentic Lock-inTertiary · Data Flywheel
The play in one sentence
Capture the unstructured "word-of-mouth" knowledge inside an AEC firm — staff project memory, relationship maps, who-knows-what — turn it into agentic workflows that own the daily BD/marketing/proposal motion, and let consumption pricing pay for itself out of customer wins. The CRM becomes the dumb pipe; Joist becomes the intelligence layer.
Stats (rough, founder-augmented)
AEC logo strength
8.8
Counter-position sharpness
9.0
Founder / domain fit
9.0
Data flywheel quality
5.5
Workflow lock-in (today)
3.8
Workflow lock-in (post-Agent Hub)
8.0
Capital reserves
3.5
Special move — the JCU pricing pivot. Joist is moving from "Procore-playbook" company-size SaaS to consumption-based pricing in Joist Compute Units (JCUs) alongside the Agent Hub launch in July 2026. This is structurally hard for Unanet, Deltek, Salesforce to match — their 80%-margin recurring CRM revenue is the cannibalization bind. Same lineage as Bricsys/BricsCAD's 20-year perpetual-license arbitrage on Autodesk, with sharper teeth: Joist prices at customer outcome (proposal won, contract booked), not at headcount.
Customers worth naming
Mortenson — ENR Top 25 GC. "If we tried to get rid of Joist AI now, we'd have a revolt."
Burns & McDonnell — ENR Top 10 engineering. Largest contract to date ($350K).
Hensel Phelps as strategic backer (via Diverge VC) — ENR Top 10 GC, Rohan's former employer.
Battles (head-to-head)
Opponent
H1 (12mo)
H2 (3yr)
Mechanism & flip
Unanet ProposalAI (CRM-incumbent's bolt-on AI)
WIN
TIE *
H1: AI-native architecture + voice/staff capture is the wedge. Flip at H2: if Unanet ships voice + staff-capture inside ProposalAI, the data-capture mechanism commoditizes. Joist's surviving moat = workflow density only if Agent Hub is entrenched first. Single biggest external watch.
Direct replacement = picking a fight Joist can't win. Buyer pain on consolidation, incumbent API restriction, 5-year build to match commission/contract/regulatory data. Reframe: not "replace," but "system of intelligence above" — let the CRM be the dumb pipe.
H1: AEC knowledge graph + agentic capture beats roll-your-own. H2: as frontier LLMs cheapen and AEC firms build internal AI, "we capture what others don't" advantage shrinks unless workflow density compounds.
Datagrid (Procore-acquired AEC AI, Jan 2026)
WIN
TIE
Different layer today (Datagrid = field/submittals; Joist = BD/proposal). H2 risk: if Procore expands into pursuit/BD via Datagrid, the agentic-construction-platform play eats edges of Joist's TAM.
Single biggest risk — and the inflection signal to watch
Risk: Unanet ProposalAI matures into voice + staff-capture in 12-18 months, before Joist's Agent Hub reaches embedded-at-scale. If that ships, the data-capture moat compresses to commodity. The only surviving moat is workflow density — and Joist needs to have already built it.
Inflection signal: Watch Unanet ProposalAI roadmap. If Unanet ships a voice-meeting-capture or staff-knowledge-extraction agent in 2026-2027 and Joist hasn't crossed the workflow-density threshold first, H2 verdict tips toward TUCK-IN ACQUISITION ($50-150M).
Base case (12-24mo): Series A closes at $30-60M post on the founder-claimed $13.4M revenue narrative. ARR landing $4-8M. Strategic-acquisition interest forms but doesn't pull trigger.
Bull (H2): wins workflow-density race. Category-defining ARR $25-60M. Strategic exit at $300-800M to Unanet/Deltek/Procore — OR independent path to scale comparable to Procore's category leadership in field-tools.
Bear (H2): Unanet's data + bolt-on AI wins. Joist becomes a tuck-in at $50-150M ($8-20M ARR scale). Same destination as the mobile-first AEC consolidation: category leader acquired by data incumbent.
Reference precedents: Cosential→Unanet (2020, est. $80-150M); Pype→Autodesk (2020, est. $40-80M); Datagrid→Procore (Jan 2026). The independent path is Procore itself — different-wedge consolidator.
Pricing-as-counter-positioning — the sharpest move
The JCU consumption pricing pivot is the cleanest counter-positioning move in AEC software 2026. Sits in the same lineage as Bricsys's 20-year perpetual-license arbitrage on Autodesk and Procore's per-customer-construction-volume pricing — both kept those companies independent because the incumbent couldn't follow without breaking its own model.
Joist's per-outcome pricing is structurally inaccessible to Unanet/Deltek/Salesforce because it cannibalizes their 80%-margin seat-licensed CRM base. The cannibalization bind is the moat.
Joist sits squarely on the engine layer (system of intelligence over the BD workflow), not the data-plumbing layer. Real-world test of three Pass 2b synthesis claims: pricing-as-counter-positioning, federated-agent-on-consumption, and engine-layer-holds-against-horizontal. [Pass 3 deep-dive will sharpen] — founder-claimed $13.4M target has three competing readings; ARR validation is the load-bearing question.
Stat-dimension discipline summary
Cards in the Evaluator-Power family (Harvey+Sierra, Trunk Tools, Pilot) carry "Calibrated-Refusal Quality" + "Accountability Surface" stats. Cards in the Data-Flywheel family (Buildots+OpenSpace, Brex+Ramp+Puzzle) carry "Flywheel Velocity" + "Data Compounding" stats. Cards in the Agentic-Lock-in family (Decagon-stack, Cursor) carry "Tool-Graph Memory" + "Workflow Embed Depth" stats. The pattern is intentional: the stat dimensions say something about the moat's mechanism rather than just sketching a brand or distribution score.
Asterisks across the entire paper, Pass 2 final state: 4, all anchored on the Part V flipped cells. (1) Glean vs Microsoft Copilot at H2. (2) NetSuite vs AI-native ERP greenfield at H2. (3) Speckle vs Autodesk reframed at H2. (4) Suffolk vs owner-led IPD at H2. Plus the Priya-pending H2 asterisk on the Autodesk card's generative-design row, which is not a fifth load-bearing asterisk — it is a conditional that flips the Speckle and generative-design rows of one card on her answer.