{ "type": "document", "pos": { "line": 1, "column": 1 }, "endLine": 208, "meta": { "filename": "/home/runner/work/noma/noma/examples/research-thesis.noma", "title": "Vertical AI Agents — Investment Thesis", "author": "ferax564", "date": "2026-05-09T00:00:00.000Z", "tags": [ "ai", "vertical-saas", "thesis", "research" ] }, "children": [ { "type": "frontmatter", "data": { "title": "Vertical AI Agents — Investment Thesis", "author": "ferax564", "date": "2026-05-09T00:00:00.000Z", "tags": [ "ai", "vertical-saas", "thesis", "research" ] }, "raw": "title: Vertical AI Agents — Investment Thesis\nauthor: ferax564\ndate: 2026-05-09\ntags: [ai, vertical-saas, thesis, research]", "pos": { "line": 1, "column": 1 }, "endLine": 6 }, { "type": "section", "id": "vertical-ai-agents-investment-thesis", "level": 1, "title": "Vertical AI Agents — Investment Thesis", "children": [ { "type": "directive", "name": "summary", "attrs": {}, "children": [ { "type": "paragraph", "content": "Horizontal LLM platforms commoditize fast. The durable value sits in\n**vertical agents** that own a workflow end-to-end inside a single\ndomain — legal review, claims processing, financial-statement audit,\nclinical documentation. This thesis lays out the structural reasons,\nthe supporting evidence, the leading indicators to watch, and the\ndisqualifying conditions that would invalidate it.", "pos": { "line": 11, "column": 1 }, "endLine": 16 } ], "pos": { "line": 10, "column": 1 }, "endLine": 17, "body": "Horizontal LLM platforms commoditize fast. The durable value sits in\n**vertical agents** that own a workflow end-to-end inside a single\ndomain — legal review, claims processing, financial-statement audit,\nclinical documentation. This thesis lays out the structural reasons,\nthe supporting evidence, the leading indicators to watch, and the\ndisqualifying conditions that would invalidate it." }, { "type": "section", "id": "thesis-at-a-glance", "level": 2, "title": "Thesis at a glance", "children": [ { "type": "directive", "name": "grid", "attrs": { "columns": 2 }, "children": [ { "type": "directive", "name": "card", "attrs": { "title": "Bull Case", "icon": "up" }, "children": [ { "type": "paragraph", "content": "Domain-specific agents win on data depth, workflow integration, and\nliability ownership. Each vertical can support 1–3 category leaders\nwith $500M+ ARR within 5 years.", "pos": { "line": 23, "column": 1 }, "endLine": 25 } ], "pos": { "line": 22, "column": 1 }, "endLine": 26, "body": "Domain-specific agents win on data depth, workflow integration, and\nliability ownership. Each vertical can support 1–3 category leaders\nwith $500M+ ARR within 5 years." }, { "type": "directive", "name": "card", "attrs": { "title": "Bear Case", "icon": "down" }, "children": [ { "type": "paragraph", "content": "Frontier model gains compress the gap. A horizontal model with strong\ntool use plus a thin vertical wrapper captures most of the value.\nVertical agents become features, not companies.", "pos": { "line": 29, "column": 1 }, "endLine": 31 } ], "pos": { "line": 28, "column": 1 }, "endLine": 32, "body": "Frontier model gains compress the gap. A horizontal model with strong\ntool use plus a thin vertical wrapper captures most of the value.\nVertical agents become features, not companies." } ], "pos": { "line": 21, "column": 1 }, "endLine": 33 } ], "pos": { "line": 19, "column": 1 }, "endLine": 33 }, { "type": "section", "id": "core-claims", "level": 2, "title": "Core claims", "children": [ { "type": "directive", "name": "claim", "attrs": { "id": "claim-data-moat", "confidence": 0.78 }, "children": [ { "type": "paragraph", "content": "Vertical agents accumulate proprietary workflow data — corrections,\nedge-case patterns, customer-specific schemas — that horizontal agents\ncannot replicate by scaling base-model capability alone.", "pos": { "line": 38, "column": 1 }, "endLine": 40 } ], "pos": { "line": 37, "column": 1 }, "endLine": 41, "id": "claim-data-moat", "body": "Vertical agents accumulate proprietary workflow data — corrections,\nedge-case patterns, customer-specific schemas — that horizontal agents\ncannot replicate by scaling base-model capability alone." }, { "type": "directive", "name": "evidence", "attrs": { "for": "claim-data-moat", "source": "harvey-public-disclosures-2026q1" }, "children": [ { "type": "paragraph", "content": "Harvey reports that 71% of model improvements in the last 12 months\ncame from fine-tuning on proprietary corrections collected in customer\ndeployments — not from base-model upgrades.", "pos": { "line": 44, "column": 1 }, "endLine": 46 } ], "pos": { "line": 43, "column": 1 }, "endLine": 47, "body": "Harvey reports that 71% of model improvements in the last 12 months\ncame from fine-tuning on proprietary corrections collected in customer\ndeployments — not from base-model upgrades." }, { "type": "directive", "name": "evidence", "attrs": { "for": "claim-data-moat", "source": "ambience-customer-case-2026" }, "children": [ { "type": "paragraph", "content": "Ambience Healthcare's clinical-documentation agent improved acceptance\nrates from 62% to 89% after twelve months in production at a single\nhospital network — the gain was specific to that network's documentation\nconventions and did not transfer to the open-source baseline.", "pos": { "line": 50, "column": 1 }, "endLine": 53 } ], "pos": { "line": 49, "column": 1 }, "endLine": 54, "body": "Ambience Healthcare's clinical-documentation agent improved acceptance\nrates from 62% to 89% after twelve months in production at a single\nhospital network — the gain was specific to that network's documentation\nconventions and did not transfer to the open-source baseline." }, { "type": "directive", "name": "claim", "attrs": { "id": "claim-liability-moat", "confidence": 0.71 }, "children": [ { "type": "paragraph", "content": "In regulated verticals, the agent vendor must own legal liability for\nagent output. This forces a stack of guarantees (audit logs, escalation,\nhuman-in-loop sign-off, SOC 2/HIPAA) that takes years to build and is\nadversarial to horizontal generalists.", "pos": { "line": 57, "column": 1 }, "endLine": 60 } ], "pos": { "line": 56, "column": 1 }, "endLine": 61, "id": "claim-liability-moat", "body": "In regulated verticals, the agent vendor must own legal liability for\nagent output. This forces a stack of guarantees (audit logs, escalation,\nhuman-in-loop sign-off, SOC 2/HIPAA) that takes years to build and is\nadversarial to horizontal generalists." }, { "type": "directive", "name": "evidence", "attrs": { "for": "claim-liability-moat", "source": "legal-tech-procurement-survey-2026" }, "children": [ { "type": "paragraph", "content": "89% of in-house legal teams surveyed in Q1 2026 said \"vendor accepts\nindemnification for agent output\" was a hard requirement for production\ndeployment. Only 4 vendors met the bar; all 4 are vertical specialists.", "pos": { "line": 64, "column": 1 }, "endLine": 66 } ], "pos": { "line": 63, "column": 1 }, "endLine": 67, "body": "89% of in-house legal teams surveyed in Q1 2026 said \"vendor accepts\nindemnification for agent output\" was a hard requirement for production\ndeployment. Only 4 vendors met the bar; all 4 are vertical specialists." }, { "type": "directive", "name": "claim", "attrs": { "id": "claim-workflow-stickiness", "confidence": 0.74 }, "children": [ { "type": "paragraph", "content": "The integration surface — clearinghouses, EHRs, court filing systems,\npractice-management software — is a structural moat, not a feature gap.\nEach integration is custom, slow, and high-trust.", "pos": { "line": 70, "column": 1 }, "endLine": 72 } ], "pos": { "line": 69, "column": 1 }, "endLine": 73, "id": "claim-workflow-stickiness", "body": "The integration surface — clearinghouses, EHRs, court filing systems,\npractice-management software — is a structural moat, not a feature gap.\nEach integration is custom, slow, and high-trust." }, { "type": "directive", "name": "counterevidence", "attrs": { "for": "claim-workflow-stickiness", "source": "model-context-protocol-traction-2026" }, "children": [ { "type": "paragraph", "content": "The Model Context Protocol (MCP) is reducing per-integration cost by an\norder of magnitude in some verticals. If MCP-style adapters become\nuniversal, the integration moat compresses faster than the data moat does.", "pos": { "line": 76, "column": 1 }, "endLine": 78 } ], "pos": { "line": 75, "column": 1 }, "endLine": 79, "body": "The Model Context Protocol (MCP) is reducing per-integration cost by an\norder of magnitude in some verticals. If MCP-style adapters become\nuniversal, the integration moat compresses faster than the data moat does." } ], "pos": { "line": 35, "column": 1 }, "endLine": 79 }, { "type": "section", "id": "risks", "level": 2, "title": "Risks", "children": [ { "type": "directive", "name": "risk", "attrs": { "id": "risk-frontier-leap", "severity": "high", "owner": "ferax564" }, "children": [ { "type": "paragraph", "content": "A frontier-model capability leap (e.g., GPT-6-class reasoning + native\nlong-horizon tool use) could collapse the workflow gap. Most vulnerable\nverticals: those where workflow complexity comes from reasoning chains\nrather than data integration (e.g., research synthesis).", "pos": { "line": 84, "column": 1 }, "endLine": 87 } ], "pos": { "line": 83, "column": 1 }, "endLine": 88, "id": "risk-frontier-leap", "body": "A frontier-model capability leap (e.g., GPT-6-class reasoning + native\nlong-horizon tool use) could collapse the workflow gap. Most vulnerable\nverticals: those where workflow complexity comes from reasoning chains\nrather than data integration (e.g., research synthesis)." }, { "type": "directive", "name": "risk", "attrs": { "id": "risk-platform-shift", "severity": "medium", "owner": "ferax564" }, "children": [ { "type": "paragraph", "content": "If the major model providers ship vertical-agent SDKs with\nrevenue-share models, distribution shifts toward platform-bundled\nofferings. Mitigation: invest in customer ownership of data\n(BYO-storage, on-prem options).", "pos": { "line": 91, "column": 1 }, "endLine": 94 } ], "pos": { "line": 90, "column": 1 }, "endLine": 95, "id": "risk-platform-shift", "body": "If the major model providers ship vertical-agent SDKs with\nrevenue-share models, distribution shifts toward platform-bundled\nofferings. Mitigation: invest in customer ownership of data\n(BYO-storage, on-prem options)." }, { "type": "directive", "name": "risk", "attrs": { "id": "risk-regulatory-freeze", "severity": "medium", "owner": "ferax564" }, "children": [ { "type": "paragraph", "content": "EU AI Act high-risk classification or US sector-specific rules (FDA,\nSEC) could freeze deployments for 12–18 months in the affected\nverticals. This *helps* incumbents and *helps* well-capitalized\nvendors with compliance teams — and disproportionately hurts startups.", "pos": { "line": 98, "column": 1 }, "endLine": 101 } ], "pos": { "line": 97, "column": 1 }, "endLine": 102, "id": "risk-regulatory-freeze", "body": "EU AI Act high-risk classification or US sector-specific rules (FDA,\nSEC) could freeze deployments for 12–18 months in the affected\nverticals. This *helps* incumbents and *helps* well-capitalized\nvendors with compliance teams — and disproportionately hurts startups." }, { "type": "directive", "name": "risk", "attrs": { "id": "risk-talent-concentration", "severity": "low", "owner": "ferax564" }, "children": [ { "type": "paragraph", "content": "Top vertical talent is concentrated in 4–6 startups per category.\nAcquihire risk is real but bounded; not a thesis-breaker.", "pos": { "line": 105, "column": 1 }, "endLine": 106 } ], "pos": { "line": 104, "column": 1 }, "endLine": 107, "id": "risk-talent-concentration", "body": "Top vertical talent is concentrated in 4–6 startups per category.\nAcquihire risk is real but bounded; not a thesis-breaker." } ], "pos": { "line": 81, "column": 1 }, "endLine": 107 }, { "type": "section", "id": "what-would-invalidate-the-thesis", "level": 2, "title": "What would invalidate the thesis", "children": [ { "type": "directive", "name": "open_question", "attrs": { "id": "invalidator-frontier-tool-use" }, "children": [ { "type": "paragraph", "content": "A frontier model that achieves >85% acceptance on a high-stakes\nvertical workflow (clinical docs, deposition review, SEC filing prep)\nwithout any vertical-specific tuning. If this happens within 18 months,\nthe data moat is weaker than claimed and `confidence` on\n`claim-data-moat` should drop below 0.5.", "pos": { "line": 112, "column": 1 }, "endLine": 116 } ], "pos": { "line": 111, "column": 1 }, "endLine": 117, "id": "invalidator-frontier-tool-use", "body": "A frontier model that achieves >85% acceptance on a high-stakes\nvertical workflow (clinical docs, deposition review, SEC filing prep)\nwithout any vertical-specific tuning. If this happens within 18 months,\nthe data moat is weaker than claimed and `confidence` on\n`claim-data-moat` should drop below 0.5." }, { "type": "directive", "name": "open_question", "attrs": { "id": "invalidator-mcp-universal" }, "children": [ { "type": "paragraph", "content": "Universal MCP-style adapters that reduce vertical integration cost from\nweeks to hours across at least three regulated verticals. If this ships\nbroadly within 12 months, `claim-workflow-stickiness` confidence should\ndrop below 0.5.", "pos": { "line": 120, "column": 1 }, "endLine": 123 } ], "pos": { "line": 119, "column": 1 }, "endLine": 124, "id": "invalidator-mcp-universal", "body": "Universal MCP-style adapters that reduce vertical integration cost from\nweeks to hours across at least three regulated verticals. If this ships\nbroadly within 12 months, `claim-workflow-stickiness` confidence should\ndrop below 0.5." } ], "pos": { "line": 109, "column": 1 }, "endLine": 124 }, { "type": "section", "id": "quantitative-backdrop", "level": 2, "title": "Quantitative backdrop", "children": [ { "type": "directive", "name": "dataset", "attrs": { "id": "vertical-ai-funding" }, "children": [ { "type": "paragraph", "content": "schema:\n vertical: string\n funded_companies: number\n total_funding_usd_m: number\n median_arr_growth_yoy: number\nrows:\n - [legal, 14, 1280, 3.4]\n - [healthcare, 22, 2650, 2.9]\n - [financial-audit, 9, 540, 4.1]\n - [insurance, 11, 720, 3.6]\n - [construction, 6, 290, 2.7]", "pos": { "line": 129, "column": 1 }, "endLine": 139 } ], "pos": { "line": 128, "column": 1 }, "endLine": 140, "id": "vertical-ai-funding", "body": "schema:\n vertical: string\n funded_companies: number\n total_funding_usd_m: number\n median_arr_growth_yoy: number\nrows:\n - [legal, 14, 1280, 3.4]\n - [healthcare, 22, 2650, 2.9]\n - [financial-audit, 9, 540, 4.1]\n - [insurance, 11, 720, 3.6]\n - [construction, 6, 290, 2.7]" }, { "type": "directive", "name": "plot", "attrs": { "id": "vertical-ai-arr-plot", "type": "bar", "dataset": "vertical-ai-funding", "column": "median_arr_growth_yoy", "xcolumn": "vertical", "title": "Median ARR YoY growth by vertical (2026)" }, "children": [], "pos": { "line": 142, "column": 1 }, "endLine": 143, "id": "vertical-ai-arr-plot" }, { "type": "directive", "name": "plot", "attrs": { "id": "vertical-ai-funding-plot", "type": "line", "dataset": "vertical-ai-funding", "column": "total_funding_usd_m", "xcolumn": "vertical", "title": "Total funding raised by vertical ($M)" }, "children": [], "pos": { "line": 145, "column": 1 }, "endLine": 146, "id": "vertical-ai-funding-plot" } ], "pos": { "line": 126, "column": 1 }, "endLine": 146 }, { "type": "section", "id": "watchlist-positions-not-recommendations", "level": 2, "title": "Watchlist (positions, not recommendations)", "children": [ { "type": "table", "header": [ "Vertical", "Public proxy", "Private leader", "Note" ], "align": [ null, null, null, null ], "rows": [ [ "Legal", "RELX, Thomson Reuters", "Harvey, EvenUp", "Watch incumbents' agent rollouts" ], [ "Clinical docs", "—", "Abridge, Ambience", "Acceptance rate is the leading metric" ], [ "Financial audit", "Intuit, S&P", "Numeric, Trullion", "Audit-trail UX is the moat" ], [ "Insurance claims", "Verisk", "Sixfold, EvolutionIQ", "Loss-ratio impact is the proof point" ] ], "pos": { "line": 150, "column": 1 }, "endLine": 155 } ], "pos": { "line": 148, "column": 1 }, "endLine": 155 }, { "type": "section", "id": "deltas-since-last-update", "level": 2, "title": "Deltas since last update", "children": [ { "type": "directive", "name": "state_change", "attrs": { "block": "claim-data-moat", "attribute": "confidence", "from": 0.72, "to": 0.78, "reason": "Harvey's Q1 disclosure quantified the proprietary-correction loop more concretely than expected", "at": "2026-05-09" }, "children": [], "pos": { "line": 159, "column": 1 }, "endLine": 160 }, { "type": "directive", "name": "state_change", "attrs": { "block": "claim-liability-moat", "attribute": "confidence", "from": 0.65, "to": 0.71, "reason": "legal-tech procurement survey put the indemnification requirement at 89%, vs. 71% the prior survey", "at": "2026-05-09" }, "children": [], "pos": { "line": 162, "column": 1 }, "endLine": 163 } ], "pos": { "line": 157, "column": 1 }, "endLine": 163 }, { "type": "section", "id": "quarterly-review-task", "level": 2, "title": "Quarterly review task", "children": [ { "type": "directive", "name": "agent_task", "attrs": { "id": "quarterly-thesis-review" }, "children": [ { "type": "paragraph", "content": "Every quarter (Q1: Mar 31, Q2: Jun 30, Q3: Sep 30, Q4: Dec 31), walk this\ndocument and:", "pos": { "line": 168, "column": 1 }, "endLine": 169 }, { "type": "list", "ordered": true, "items": [ { "type": "list_item", "content": "For each `claim`, check whether new public evidence supports or", "pos": { "line": 171, "column": 1 }, "endLine": 171 } ], "pos": { "line": 171, "column": 1 }, "endLine": 171 }, { "type": "paragraph", "content": " contradicts. If material, add a fresh `evidence` or `counterevidence`\n block and adjust the `confidence` attribute.", "pos": { "line": 172, "column": 1 }, "endLine": 173 }, { "type": "list", "ordered": true, "items": [ { "type": "list_item", "content": "For each `risk`, check whether the leading indicators have moved.", "pos": { "line": 174, "column": 1 }, "endLine": 174 } ], "pos": { "line": 174, "column": 1 }, "endLine": 174 }, { "type": "paragraph", "content": " Adjust `severity` if warranted.", "pos": { "line": 175, "column": 1 }, "endLine": 175 }, { "type": "list", "ordered": true, "items": [ { "type": "list_item", "content": "For each `open_question` invalidator, check whether the trigger", "pos": { "line": 176, "column": 1 }, "endLine": 176 } ], "pos": { "line": 176, "column": 1 }, "endLine": 176 }, { "type": "paragraph", "content": " condition has been met. If yes, escalate to a `decision` block\n recommending exit or rebalance.", "pos": { "line": 177, "column": 1 }, "endLine": 178 }, { "type": "list", "ordered": true, "items": [ { "type": "list_item", "content": "Do not delete prior evidence. Append, don't overwrite — the audit", "pos": { "line": 179, "column": 1 }, "endLine": 179 } ], "pos": { "line": 179, "column": 1 }, "endLine": 179 }, { "type": "paragraph", "content": " trail is the value.", "pos": { "line": 180, "column": 1 }, "endLine": 180 } ], "pos": { "line": 167, "column": 1 }, "endLine": 181, "id": "quarterly-thesis-review" } ], "pos": { "line": 165, "column": 1 }, "endLine": 181 }, { "type": "section", "id": "stale-evidence-guard", "level": 2, "title": "Stale-evidence guard", "children": [ { "type": "directive", "name": "agent_task", "attrs": { "id": "stale-evidence-scan" }, "children": [ { "type": "paragraph", "content": "Every two weeks, scan all `evidence` blocks for `source` attributes\nolder than 90 days. Propose (do not apply) a `replace_block` patch for\neach, citing the latest available source.", "pos": { "line": 186, "column": 1 }, "endLine": 188 } ], "pos": { "line": 185, "column": 1 }, "endLine": 189, "id": "stale-evidence-scan", "body": "Every two weeks, scan all `evidence` blocks for `source` attributes\nolder than 90 days. Propose (do not apply) a `replace_block` patch for\neach, citing the latest available source." } ], "pos": { "line": 183, "column": 1 }, "endLine": 189 }, { "type": "section", "id": "export", "level": 2, "title": "Export", "children": [ { "type": "directive", "name": "export_button", "attrs": { "format": "prompt", "target": "document" }, "children": [ { "type": "paragraph", "content": "Label: Copy as second-opinion review prompt", "pos": { "line": 194, "column": 1 }, "endLine": 194 } ], "pos": { "line": 193, "column": 1 }, "endLine": 195, "body": "Label: Copy as second-opinion review prompt" }, { "type": "directive", "name": "export_button", "attrs": { "format": "llm", "target": "document" }, "children": [ { "type": "paragraph", "content": "Label: Copy structured LLM context", "pos": { "line": 198, "column": 1 }, "endLine": 198 } ], "pos": { "line": 197, "column": 1 }, "endLine": 199, "body": "Label: Copy structured LLM context" }, { "type": "directive", "name": "export_button", "attrs": { "format": "markdown", "target": "summary" }, "children": [ { "type": "paragraph", "content": "Label: Copy summary as Markdown", "pos": { "line": 202, "column": 1 }, "endLine": 202 } ], "pos": { "line": 201, "column": 1 }, "endLine": 203, "body": "Label: Copy summary as Markdown" }, { "type": "quote", "content": "A thesis is only useful if you can revisit it. The blocks above are\nstructured so a future-you (or a future agent acting on your behalf)\ncan update only what changed, leave the rest alone, and produce a\nclean Git diff that shows exactly which beliefs moved.", "pos": { "line": 205, "column": 1 }, "endLine": 208 } ], "pos": { "line": 191, "column": 1 }, "endLine": 208 } ], "pos": { "line": 8, "column": 1 }, "endLine": 208, "aliases": [ "research-thesis" ] } ] }