# Category Theses — AI Startup Intelligence & Investment Memo Pipeline Written 2026-09-02. These are the scoring yardstick every company gets measured against for "thesis fit." Revisit quarterly as the market moves. --- ## 1. Agent Infrastructure **The real trend.** Agents moved from the read path (analyzing data) to the write path (executing business operations) in 2025-2026, and that shift created a genuine infrastructure gap: durable execution/orchestration (keeping a multi-step agent run alive across failures, retries, human-in-the-loop pauses), observability/evaluation (what did the agent actually decide, and was it right), and gateways/proxies (routing, cost control, guardrails across many models and tools). Funding data backs this up — infrastructure is one of the most heavily funded sub-categories in enterprise AI through 2026, and capital is concentrating hard: the top 10 deals captured ~78% of category funding by mid-2026, up from 70% in 2025. **What's overhyped.** Generic "agent frameworks" and "build your own agent" platforms with no opinion on a specific failure mode. Anything that's really just a chat-with-your-docs wrapper rebranded as "agentic." Multi-agent orchestration frameworks competing purely on developer ergonomics rather than solving a production reliability problem — there are already too many of these and most won't survive consolidation. **The wedge that actually wins.** Solving one gnarly, concrete production problem that the foundation model labs won't solve themselves because it's not their business: durable state across long-running agent workflows, ground-truth evaluation of agent decisions (not just output quality), or cost/latency control at scale. The winners look boring — they're closer to "systems software for a new runtime" than "AI product." Distribution via developer adoption (open source core, paid cloud) beats top-down enterprise sales in this layer, at least pre-Series B. **The 18-month view.** Expect consolidation: today's fragmented landscape (execution engines, eval tools, gateways, memory layers as separate companies) narrows as the biggest players in each sub-category either get acquired by cloud/model vendors trying to own the full stack, or expand horizontally into adjacent infra layers. Durable execution and evaluation are the two sub-categories most likely to matter in 3 years; generic "agent builder" platforms are the most likely to get squeezed out by the labs shipping better agent SDKs directly. **Thesis-fit scoring cues.** High fit: solves a specific, named production failure mode; has real usage/logs proving the problem is real (not hypothetical); founding team has operated agents in production before (not just built demos). Low fit: horizontal "no-code agent builder," thin wrapper on an LLM API with no defensible technical moat, competing on UI polish alone. --- ## 2. Vertical SaaS, AI-Native **The real trend.** AI agents are moving vertical SaaS from "software that supports a human's decision" to "software that autonomously executes the domain-specific work itself." The market is large and compounding fast — the AI agents market overall is projected past $10.9B in 2026 growing ~46%/year through 2030, but vertical/industry-specific agents are growing faster still, around 63%/year. The reason vertical wins: foundation model labs don't have permissioned access to a hospital's clinical notes, a law firm's case archive, or an insurer's claims data — vertical AI-native companies do, because they sit inside the workflow and often own the data relationship. Proof points are already large: Abridge (ambient clinical AI) at a $5.3B valuation with a $316M extension in April 2026; Sierra (customer support agents) at $15.8B; Harvey and Legora in legal both crossing $100M+ ARR at extreme speed. **What's overhyped.** "AI-native" as a label slapped on a copilot feature bolted onto an existing SaaS product roadmap, with no actual change to the software's operating model. Point solutions that automate a single narrow task (e.g., "AI note-taker for X industry") without a path to owning more of the workflow — these get commoditized fast once a platform player adds the same feature. **The wedge that actually wins.** Owning a full end-to-end workflow in a regulated or data-rich vertical (healthcare, legal, construction, insurance, financial services back-office) where compliance requirements and proprietary data access are the moat, not the model. The best companies replace software-that-assists with software-that-does, and they compound advantage by accumulating proprietary outcome data that improves the product and is hard for a new entrant (or a horizontal platform) to replicate. Speed to meaningful ARR (Harvey/Legora/Sierra all crossed $100M ARR unusually fast) is itself a signal of real product-market fit in this category, not just hype. **The 18-month view.** The frontier moves from "impressive demo, ambient AI note-taker" to "software that takes real autonomous actions with accountability" — expect the next wave of breakout vertical AI companies to be measured on actions completed and liability assumed, not just time saved. Categories still described as having "open windows" going into H2 2026 (construction, parts of insurance, parts of finance ops) are worth extra sourcing attention precisely because the obvious verticals (legal, healthcare, customer support) are already getting crowded and richly valued. **Thesis-fit scoring cues.** High fit: proprietary data or workflow access a horizontal platform can't easily replicate; evidence of autonomous execution (not just assistive copilot); fast ARR growth relative to time since launch. Low fit: feature-level AI bolt-on to existing software category; thin data moat; crowded vertical (legal/healthcare-general) without a differentiated sub-niche or wedge. --- ## 3. AI-Native Fintech **The real trend.** Fintech funding is up sharply (~23% YoY in H1 2026, ~$28.6B globally) even as deal *count* fell over 25% — investors are writing fewer, much larger checks, concentrated on AI. AI-native fintech deals specifically jumped from 13 to 50 year-on-year in H1 2026 (+285%), with funding more than tripling. The framing that matters: AI-native fintechs aren't "old fintech plus a chatbot" — they're structurally different businesses built around compliance moats, proprietary data, and AI-native architecture from day one. Agentic payments and AI-native trading/brokerage infrastructure (e.g., Alpaca's "agent-first brokerage" raise) are a distinct emerging sub-thread worth tracking separately from AI-native underwriting/risk (e.g., Taktile) and AI-native back-office/credit infra (9fin, Flutterwave-scale infra plays). **What's overhyped.** "AI-powered" personal finance apps and budgeting tools with a thin LLM chat layer over existing rails — this category is crowded, low-moat, and easily replicated by incumbents. Generic "AI copilot for your bank" enterprise tools that don't touch the actual decision (underwriting, risk, execution) and are really just support-ticket deflection with fintech branding. **The wedge that actually wins.** AI embedded directly into a regulated decision point — underwriting, credit risk, fraud/compliance, trade execution — where the AI-native architecture lets a company underwrite/price/execute faster and more accurately than incumbents built on legacy risk models, and where regulatory/compliance sophistication is itself part of the moat (hard to copy quickly, expensive to get wrong). Agentic payments and agent-first execution/brokerage infrastructure is the newest and least-crowded sub-thread — worth disproportionate sourcing attention given the early but fast-growing deal data. **The 18-month view.** Expect continued check-size concentration (mega-rounds to category leaders, fewer seed bets), and a widening gap between AI-native companies built with compliance/data moats from inception vs. incumbents retrofitting AI onto legacy risk infrastructure. Stablecoins and agentic payments are the areas most likely to see a new category leader emerge in this window rather than get won by an incumbent. **Thesis-fit scoring cues.** High fit: AI sits at a regulated decision point (underwriting/risk/execution/compliance), not just a chat UI; founding team has real fintech/regulatory operating experience; proprietary data feeding the model that a bank can't easily buy. Low fit: consumer budgeting/chat app with thin fintech rails underneath; no clear regulatory or data moat; feature parity achievable by an incumbent bank's product team in a quarter. --- *Sources consulted: Buttondown AI Funding newsletter (Jul 2026), aifunding.me AI Agent Funding tracker, gravity.fast AI Agent Funding Tracker Q3 2026, presenc.ai Agent Infrastructure Landscape (May 2026), saasmag.com Vertical AI Agents piece, Crunchbase News fintech H1 2026 report, Forbes coverage of AI-native fintech (Aug 2026).*