03 · Cast list + battle cards — Strategic Moats in the Age of AI

Cast — organized by attack archetype

Pass 1 organizes the cast around attack vectors, not company count. Where multiple companies share an archetype, they're bundled into one battle card representing the archetype, with named exemplars inside. The 5 cards below (NetSuite, Intuit, Glean, Suffolk, Speckle) are format proofs — Pass 2 fills in the remaining archetype + incumbent cards (Autodesk most critically) at the same fidelity.

AI-native horizontal — 4 attack archetypes

Archetype 1 · Horizontal enterprise knowledge / retrieval layer. Canonical: Glean ($7.2B / $200M ARR, Series F June 2025). Variant: Hebbia. Attack vector: index-everything-with-permission-aware retrieval, building a data flywheel of organizational context. Battle card below.
Archetype 2 · Vertical workflow co-pilot for high-stakes professional services. Harvey (legal, $8B val) + Sierra (CX, $10B / $150M ARR), bundled. Attack vector: domain-specific data + evals + workflow integration that compounds with usage. Primary moat under Pass 1.5 reframe: Evaluator Judgment Power — Harvey's value is the calibrated "no" on a contract clause; Sierra's is the calibrated escalation when the agent shouldn't auto-resolve. Brand-as-Trust + outcomes-based pricing are downstream consequences of that judgment-bearing surface.
Archetype 3 · AI-native UX rebuild of an embedded workflow. Canonical: Cursor ($2B ARR by Feb 2026, fastest dev-tools scale ever). Attack vector: be so much better at the core loop that procedural switching costs collapse. Tests whether 10x-UX is durable or commoditizable by GitHub Copilot's distribution.
Archetype 4 · Generative agentic execution layer — node in a federated agent-translator network. Decagon + Crescendo + Clay, bundled. Attack vector: agents that act on behalf of the user, accumulating workflow lock-in via tool-graph memory, escalation patterns, and learned outcome calibration. Pass 1.5 reframe: these are not vertically integrated end-to-end stacks — they're specialist nodes in a federated network of agent-translators. The moat sits in the node (depth at one job, evaluator quality, escalation calibration), not the integration. Translator/coordination layers commoditize on top of MCP-style protocols. Re: codesign-thesis claim #5.

AEC general contractors — 3 incumbents

Suffolk Construction — most explicit tech bet (Suffolk Technologies fund, ~50 portfolio companies, $110M Fund I closed 2023). The "leaning into the disruption" archetype. Battle card below.
Turner Construction — Turner Ventures CVC; doubled data-center revenue in 2025 on AI-infrastructure projects. The "scale + bonding capacity + AI buildout tailwind" archetype.
DPR Construction — data-and-AI as a stated core focus area. The "process power + employee-ownership culture" archetype. Tests whether private-company patience is itself a moat.

AEC software — 1 contrast incumbent + 4 attack archetypes

Autodesk (contrast probe) — the canonical "looks moated, structurally exposed" incumbent. Helmer-strong on Switching Costs today (Revit + file-format lock-in) but the EULA / contract-killer moves are tells. Generative design + open-format AI tools attack the file-format moat directly; AI-native re-platforming is structurally hard given pricing-model + cannibalization risk. The cautionary archetype against which the four attack archetypes below are war-gamed. Battle card to be built in Pass 2.
Archetype A · Contract / commercial-relationship disruption. Canonical: Higharc. Attack vector: redefine the homebuilder ↔ buyer ↔ contractor commercial structure, not just the design tool. Counter-Positioning at the business-model layer.
Archetype B · API / data-gravity protocol — project-graph as system of record. Canonical: Speckle. Adjacent: open-format generative tools (Snaptrude). Attack vector: not file-format translation. The mechanism is to make the project graph (rooms, walls, systems, versions, contributors, comments) the live system of record across tools — Revit, Rhino, Grasshopper, AutoCAD, Civil 3D, Unreal, Unity — so the file format degrades to a serialization detail. The lock-in surface that's actually being attacked is API access, data terms, and whose graph the project commits to. Re: codesign-thesis claim #2 — "file formats are no longer the moat. APIs, data gravity, and platform terms are." Battle card below.
Archetype C · Document / spec / risk AI agent — Evaluator-Power candidate. Trunk Tools + Document Crunch, bundled. Attack vector: AI agents parsing specs, RFIs, contracts, schedule — automating the document-bound knowledge work without touching the design tool. Pass 1.5 note: this archetype's moat is increasingly Evaluator Judgment Power — TrunkSubmittal's value isn't the 72% of submittals it processed at Gilbane, it's the 72% it correctly refused as non-compliant. Pass 2 will produce a standalone Trunk Tools card with an Evaluator-Power stat dimension; the archetype-bundled card will retain the legacy framing for now.
Archetype D · Field-capture data flywheel. Buildots + OpenSpace, bundled. Attack vector: jobsite computer vision + sensor data compounding into a proprietary construction-progress dataset. The only AEC archetype that builds a Data Flywheel moat from physical-world capture — orthogonal to the other three.

Accounting — 4 cast members

NetSuite (Oracle) — the slow-but-bigger complexity incumbent. Multi-jurisdiction compliance, ERP-grade switching costs, Oracle channel. Battle card below.
Intuit / QuickBooks — 80% SMB share. GenOS, fine-tuned financial LLMs, Enterprise Suite push (the offensive move at NetSuite). Battle card below.
Pilot — managed bookkeeping with AI augmentation. Pass 1.5 note: Pilot is the cleanest accounting-side test of Evaluator Judgment Power — the customer is paying for accountability-bearing review, not for the AI doing the books. Tests services-led-growth as a moat-building tactic vs. pure-software attackers, but reframed under the Evaluator-Power lens: the pricing surface (managed-service margin) is the moat dimension being defended.
Brex + Ramp + Puzzle (treated as one battle-card "stack") — the AI-native, integrated SMB finance fabric. Tests whether bundled flywheel + counter-positioning beats Intuit's brand + distribution.

Cross-cutting test cases — 2

Figma — the most public test case of "is multiplayer collaboration a moat against AI-native attackers." Stock down ~14% on AI-moat concerns (April 2026). Tests Network + Switching against capital-rich incumbents (Adobe), bundled free attackers (Google Stitch), and design-to-code collapsers (Cursor / Replit).
Plaid — the canonical Cornered-Resource + Switching test case. Bank integrations are a regulatory + relational moat; AI doesn't directly attack them. Useful contrast for the "what does AI not compress" argument.

Cast count under archetype framing: 4 horizontal-attack archetypes + 4 AEC-attack archetypes + 3 GC incumbents + 4 accounting cast + 1 Autodesk contrast + 2 cross-cutting = 14 battle cards in Pass 2 (down from 17 by bundling). Tighter, more analytically MECE, fewer orphan companies.

Working battle cards (5)

Format proofs. Once Dania approves the visual + structural format, Pass 2 builds the remaining 12.

Slower-bigger complexity incumbent

NETSUITE

The Complexity Whale

Primary · Switching Secondary · Cornered Resource Tertiary · Scale

Switching Gravity

9.5

Compliance Surface

9.2

Distribution Reach

8.0

Capital Reserves

9.8

Speed of Iteration

3.5

Data Depth

6.2

Brand Power

6.5

Special move — Regulatory Surface Drag. Multi-entity, multi-jurisdiction compliance (revenue recognition, statutory reporting, tax engines across 200+ jurisdictions) is not just a feature — it's a moat that compounds with regulatory complexity. Every new IFRS update, every new e-invoicing mandate, deepens the moat against AI-native attackers who'd have to re-build all of it.

Opponent	H1 (12mo)	H2 (3yr)	Tradeoff
Intuit Enterprise Suite	WIN	TIE	Mid-market complexity ceiling
Brex+Ramp+Puzzle stack	WIN	WIN	Different segment
AI-native ERP greenfield (theoretical)	WIN	TIE *	Asterisk — switching-cost vs AI-rebuilt-UX

Asterisk · Tradeoff — NetSuite vs. AI-native ERP greenfield at H2 (one of 4 paper-wide)

Why this is one of the 4 load-bearing tradeoffs: this is the clean test of switching-cost-as-moat vs. AI-rebuilt-UX-as-attack. NetSuite's moat (multi-entity, multi-jurisdiction, ERP-grade migration cost) is the textbook deep-switching-cost incumbency. The attack vector is an AI-native rebuild that collapses re-training friction enough that greenfield economics shift the buying criteria.

Tips toward attacker if: a foundation-model company or a well-capitalized startup ships a credible AI-native ERP with full revenue recognition + multi-entity by mid-2029 (within H2 = 3 years from May 2026), AND can clear SOC 2 / ISO / regional data residency at enterprise grade. Both bars are high. The technical bar is re-implementable; the regulatory bar takes 3+ years of customer deployments to earn — which is exactly the H2 horizon.

Holds for NetSuite if: migration cost stays at $5M–$50M per mid-market customer (current Gartner benchmark range), AND regulatory surface keeps compounding (BEPS Pillar Two, country-specific e-invoicing) faster than AI agents can encode it. Compliance moat is the load-bearing variable — AI commoditizes accounting logic but not jurisdictional coverage.

SMB-flywheel incumbent moving up-market via IES + Anthropic-powered agents

INTUIT

The Counter-Position Move at NetSuite

Primary · Data Flywheel (transaction) Secondary · Network (accountant channel) Tertiary · Counter-Positioning (IES vs NetSuite)

Distribution Reach

9.6

Transaction-Flywheel Velocity

8.8

Accountant-Channel Lock

9.0

IES Counter-Position vs NetSuite

7.2

Adjacency Leverage (Mailchimp / KK)

6.5

Vertical-Edition Depth (mfg)

2.5

TurboTax Disruption Drag

5.5

Capital Reserves

9.2

Special move — The Anthropic-powered agent SDK as IES counter-position. Intuit's offensive move isn't "moving up-market." It's a counter-position at NetSuite priced into the IES architecture: 30-day deploy vs. NetSuite's 18–24 weeks of actual implementation (per Gartner-cited 73% budget-overrun rate), consumer-grade UX vs. NetSuite's shop-floor adoption failure, and the February 2026 multi-year Anthropic partnership letting mid-market customers build no-code/low-code agents on IES via Claude Agent SDK. What makes the agent layer distinctive (not table-stakes): the GenOS strategy rides on Intuit's transaction graph — IES ingests AP, AR, payroll, and payments natively through the Intuit ecosystem (QuickBooks Payments, QuickBooks Payroll, Mailchimp, Credit Karma adjacencies). That data shape is Intuit's, not Anthropic's. NetSuite can sign its own foundation-model partnership; it can't conjure Intuit's SMB transaction flywheel. What's table-stakes: "fine-tuned financial LLM" is no longer a moat claim by itself — Acumatica 2026 R1 shipped AI Studio (no-code LLM-to-screen) in March 2026; the gap window is closing.

Opponent	H1 (12mo)	H2 (3yr)	Mechanism / flip-condition
NetSuite (Oracle pushing up-market)	TIE	WIN	IES grew QBO-Adv+IES +40% YoY and new IES contracts +50% QoQ through Q2 FY26 — the counter-position is working at the segment NetSuite is abandoning. Flips for NetSuite if IES manufacturing edition slips past 2027 (currently 12–18 months of platform engineering away — no BOMs, work orders, WIP, serial/lot today).
Brex+Ramp+Puzzle stack	TIE	LOSE *	Bundled-flywheel + AI-native ledger beats unbundled accountant-channel brand at the startup cohort. Holds for Intuit only if Anthropic agent SDK lands real customer agents in 2026 and the accountant-influenced share (already 33%, up 10pts QoQ) keeps compounding into the lower-mid-market.
Pilot (managed-bookkeeping + Evaluator)	WIN	TIE	Different ICP today, but Pilot is the test case for Evaluator Judgment Power as a pricing surface. If Pilot's accountability-bearing review margin compounds while Intuit's IES tries to sell AI-as-product without absorbing liability, the Evaluator moat tilts to Pilot inside the assisted-bookkeeping segment.
Acumatica (the real benchmark)	TIE	TIE	Acumatica's purpose-built manufacturing depth (native MES, APS, ECC) vs. IES's speed + UX + Anthropic-agent stack. Acumatica 2026 R1 narrowed the AI gap with AI Studio. Resolves cleanly to TIE both horizons until IES ships verticals beyond construction (currently the only shipped IES vertical edition, open-beta since Feb 2026).
IRS Direct File (TurboTax structural drag)	LOSE	LOSE	25 states, 30M+ taxpayers covered. Out of scope for the IES moat thesis but priced into the stock — INTU at $369 (Apr 2026), down 55% from $814 peak, P/E compressed 45x→23x. Career-comp risk, not business risk; IES insulated.

Three deep-dive-grounded specifics this card carries (vs. Pass 1's generic "80% SMB share, GenOS, Enterprise Suite push"): (1) IES Construction Edition is the only shipped IES vertical edition (Feb 11 2026, open beta) — the manufacturing edition is 12–18 months of platform engineering away (no BOMs, work orders, WIP, serial/lot). (2) Counter-position is structural, not aspirational: IES at 30-day deploy beats NetSuite's 73% budget-overrun rate and $5–50M migration cost benchmark. (3) The Anthropic partnership (Feb 24 2026, multi-year, Claude Agent SDK on IES, spring 2026 rollout) is the distinctive piece — but Acumatica 2026 R1 shipped AI Studio in March 2026, so the differentiation window is closing inside H1.

Per asterisk discipline, only the Intuit-vs-Brex/Ramp/Puzzle matchup gets a tradeoff drawer (the bundled-flywheel-vs-accountant-channel close call at H2). Other matchups resolve cleanly. Detailed argument in Chapter 12 prose.

Archetype 1 · Horizontal enterprise knowledge / retrieval (canonical: Glean; variant: Hebbia)

GLEAN

The Permission-Aware Knowledge Graph

Primary · Data Flywheel Secondary · Agentic Lock-in Tertiary · Switching

Data Depth

8.8

Switching Gravity

7.8

Speed of Iteration

9.0

Brand Power

5.5

Distribution Reach

4.8

Capital Reserves

7.5

Counter-Position

6.2

Special move — The 12–18 Month Maturation Curve. Glean's own claim: the knowledge graph takes 12–18 months of real usage to fully mature for a large enterprise. That's not a marketing line — it's a structural escape-velocity threshold. Once mature, a Microsoft Copilot challenger can't catch up by being free; the customer would have to re-mature a competing graph. If true, this is the moat. If the gap closes faster than 18 months, it isn't.

Opponent	H1 (12mo)	H2 (3yr)	Tradeoff
Microsoft Copilot	TIE	TIE *	Asterisk — distribution vs. data flywheel; the central horizontal matchup
Google Gemini for Workspace	WIN	TIE	Google enterprise distribution underbuilt
OpenAI ChatGPT Enterprise	WIN	WIN	Permission model + multi-source retrieval

Asterisk · Tradeoff — Glean vs. Microsoft Copilot at H2 (one of 4 paper-wide)

Why this is one of the 4 load-bearing tradeoffs: this is the clean test of distribution-vs-data-flywheel — the central question of the AI-native horizontal chapter. Microsoft has unmatched distribution (M365 seat base ~400M) and an existential incentive to win enterprise AI. Glean has better retrieval and a head-start on permission-aware knowledge graphs that mature with usage.

Tips toward Microsoft if: Copilot's retrieval quality reaches "good enough" before Glean's knowledge-graph maturation premium becomes indispensable. Specific signal at month 12: Copilot's enterprise-search retrieval quality benchmarks vs. Glean's in the same M365-heavy customer environment. If Microsoft closes the gap to within ~10%, distribution wins.

Tips toward Glean if: the knowledge-graph maturation premium holds — year-2 customer NPS / retention materially higher than Copilot's in the same accounts. Also: how multi-app the customer is. Glean wins more in heterogeneous stacks (M365 + Slack + Salesforce + Notion + Jira); Microsoft wins more in M365-pure stacks. By H2 = 3 years, the tilt likely follows whichever side compounds faster on its native dynamic — Glean's flywheel or Microsoft's seat distribution.

Hand-off flag for Priya: is the permission-aware retrieval layer technically replicable in 18 months by Microsoft, or is there a hard architectural moat there? This is the technical question whose answer changes the H2 call materially.

Tech-leaning AEC general contractor

SUFFOLK

The GC That Bought a VC Fund

Primary · Cornered Resource (bonding + relationships) Secondary · Process Power Tertiary · Data Flywheel (emerging)

Compliance Surface

8.2

Capital Reserves

7.0

Distribution Reach

6.5

Speed of Iteration

5.5

Data Depth

4.5

Brand Power

6.2

Counter-Position

2.5

Special move — Bonding Capacity as Cornered Resource. A GC can't bid mega-projects without surety bonding capacity in the hundreds of millions to billions. That capacity is awarded based on decades of completed-project history, balance sheet, and surety-relationship trust. AI doesn't move this needle. A new entrant — even a perfectly AI-equipped one — can't bond a $1B hospital project from cold start. This is the most under-discussed AEC moat.

Opponent	H1 (12mo)	H2 (3yr)	Tradeoff
AI-native preconstruction startups	WIN	TIE	Estimating gets commoditized
Turner / DPR (peer GCs)	TIE	TIE	Tech bet vs. scale bet
Owner-led IPD / vertical-integrated developers	TIE	LOSE *	Asterisk — disintermediation by owners with repeat volume

Asterisk · Tradeoff — Suffolk vs. owner-led IPD / vertical-integrated developers at H2 (one of 4 paper-wide)

Why this is one of the 4 load-bearing tradeoffs: the disintermediation bet is the most analytically dangerous attack on the GC moat. At H2 = 3 years, AI lowers the cost of running a construction project for an owner with sufficient repeat volume (REITs, hyperscalers, healthcare systems). If AI lets these owners run their own construction-management with a thin services bench + sub network, the GC gets squeezed. This is the structurally most dangerous attack — and the one GCs are least defended against.

Tips toward owners if: a top-3 hyperscaler successfully runs $500M+ datacenter projects with AI-assisted self-perform construction management by mid-2029, AND surety markets adapt to bond owner-led delivery on those projects. Both are required; today neither is true at scale.

Holds for Suffolk if: bonding capacity remains the chokepoint AND complex multi-trade coordination remains tacit-knowledge-bound. If both hold, GCs become specialists in projects too irregular for owner self-perform — but that's a smaller, lower-margin business than today's GC franchise.

Other Suffolk matchups (preconstruction startups, peer GCs) resolve to clean calls without a tradeoff drawer per asterisk discipline. Argument lives in Chapter 10 prose.

Archetype B · API / data-gravity protocol — project-graph as system of record (canonical: Speckle; adjacent: Snaptrude)

SPECKLE

The Project Graph as System of Record

Primary · Counter-Positioning Secondary · Data Flywheel (project-graph) Tertiary · Network (developer + AEC tool ecosystem)

API / Data-Gravity Pull

7.8

Counter-Position vs Autodesk

8.8

Tool-Ecosystem Coverage

7.5

Speed of Iteration

8.0

Open-Source Trust / Brand

6.2

Capital Reserves

3.5

Enterprise Distribution

3.0

Special move — The Project Graph as System of Record. Speckle's attack vector is not "translate Revit ↔ IFC ↔ DWG." It's to make the live project graph (rooms, walls, systems, versions, contributors, comments) the system of record across every tool the project touches — Revit, Rhino, Grasshopper, AutoCAD, Civil 3D, Unreal, Unity. Once the graph becomes authoritative, the file format collapses to a serialization detail. This reframes the moat-vs-moat story: Autodesk's defensive perimeter is no longer the RVT binary, it's API access and data terms. The clearest tell is Autodesk's tiered APS pricing (December 2025) — incumbents now defend the API perimeter, not the file. Counter-positioning in its purest form: Autodesk literally cannot ship Speckle's product without converting its Revit-license customers into shoppable accounts.

Opponent	H1 (12mo)	H2 (3yr)	Mechanism / flip-condition
Autodesk (whose graph is system of record?)	TIE	WIN *	Asterisk — counter-positioning under AI pressure. Whose graph wins as project system of record? File format is downstream of that decision.
Snaptrude (open-format generative)	WIN	WIN	Different layer (Snaptrude generates inside the graph; Speckle is the graph). Collaborator, not competitor.
Hyperscaler / open-standards push (Nvidia OpenUSD)	TIE	LOSE	Standards arbitrage from above; if OpenUSD becomes the graph format, Speckle becomes a connector layer rather than the system of record.

Asterisk · Tradeoff — Speckle vs. Autodesk at H2 (one of 4 paper-wide)

Why this is one of the 4 load-bearing tradeoffs: this is the clean test of counter-positioning under AI pressure, restated under the Pass 1.5 reframe. The matchup is no longer "open vs proprietary file format" — it's "whose project graph wins as system of record." Autodesk's defensive moves (EULA changes documented by AEC Magazine, transaction-model shift, tiered APS pricing in December 2025) are tells of an API/data-gravity moat under pressure, not a file-format moat. The longer Autodesk relies on controlling the graph through tier-2 API pricing rather than shipping a credibly open project-graph layer, the deeper the counter-position trap becomes. By H2 = 3 years (mid-2029), the project-graph thesis has either won by default (Autodesk fragments under generative-design pressure and agents route around the binary format via Revit MCP servers and ODA SDK) or been absorbed (Autodesk ships a credible open graph at the cost of Revit-license margin).

Tips toward Autodesk if: Autodesk ships a credible open project-graph layer (likely via Forge/APS) that Speckle can't differentiate from and accepts the margin compression that comes with making the graph genuinely open. Plausible only if leadership chooses to cannibalize before being cannibalized — which is exactly the move Helmer's Counter-Positioning predicts incumbents won't make voluntarily.

Holds for Speckle if: AEC continues demanding cross-tool interop faster than Autodesk wants to provide it (the codesign-thesis read: the agent layer is being assembled around Autodesk via Revit MCP servers, ODA SDK, DataDrivenConstruction extraction tools — a "fragments-into-platform" trajectory that breaks API-as-moat inside a 24-month window), and Speckle builds a sustainable commercial layer (managed cloud, enterprise governance) without losing open-source credibility.

Hand-off flag for Priya: does generative design (text/sketch → production-ready BIM) clear the production bar within H2? If yes, the project-graph moat dynamic accelerates — agents need a graph to write into, and Speckle's open one is the path of least resistance. If generative design stalls at concept-design only, Autodesk gets more time and the call gets closer.

Other Speckle matchups (Snaptrude, hyperscaler open-standards push) resolve to clean calls per asterisk discipline. The commercial-execution risk for Speckle is real and lives in the Capital Reserves score (3.5) — not in a separate drawer.

What Pass 2 adds for the remaining cast (under archetype framing): 9 more battle cards at the same fidelity — three horizontal-attack archetype cards (Vertical workflow co-pilot · Harvey + Sierra with explicit Evaluator-Power stat dimension, AI-native UX rebuild · Cursor, Generative agentic execution · Decagon + Crescendo + Clay framed as federated-network nodes), three AEC-attack archetype cards (Contract / commercial disruption · Higharc; Document / spec / risk AI agent — likely standalone Trunk Tools card with Evaluator-Power dimension per Pass 1.5; Field-capture data flywheel · Buildots + OpenSpace), the Autodesk contrast-probe card (load-bearing — the canonical "looks moated, structurally exposed" incumbent, re-aimed as API/data-gravity defender per Pass 1.5), plus Turner, DPR, Pilot with Evaluator-Power dimension, Brex+Ramp+Puzzle stack, Figma, and Plaid. Total: 14 battle cards in Pass 2. On top of that, the 10×10 master matchup matrix (Pass 1.5: was 9×9; the new Evaluator Judgment Power row + column generate the most interesting matchups against Counter-Positioning, Branding, and Process Power) ties all the company-level battles back to the moat-type level.

Asterisks across the entire paper: 4. (1) Glean vs Microsoft Copilot at H2 — distribution-vs-data-flywheel. (2) NetSuite vs AI-native ERP greenfield at H2 — switching-cost-vs-AI-rebuilt-UX. (3) Speckle (file-format archetype) vs Autodesk at H2 — counter-positioning under AI pressure. (4) Suffolk vs owner-led IPD at H2 — disintermediation by repeat-volume owners. Every other matchup resolves to a clean WIN/LOSE/TIE.