Venture capital firm Andreessen Horowitz stated that there are still gaps in vertical AI startups: existing software manages system records, large model labs control general access, and vertical companies can reliably focus on completing specific tasks.

It provided four market selection criteria: the work must be repetitive enough to learn, judgment is important with subtle differences to learn, experts can quickly assess outputs and point out areas for improvement, and starting from a single task can lead to taking on more roles. These four criteria together form a learning loop, exchanging for deeper permissions, more job-specific data, and continuous calibration of what constitutes qualified delivery.

On the same day, partner Seema Amble published a lengthy article titled "The Incumbents Have Arrived," solidifying the logic: interfaces and record systems are being separated, and AI-native companies should re-bundle around "jobs"; vertical companies must outperform dedicated agents of incumbents and general agents like Claude in completing cross-system tasks, or they have no reason to exist. The article uses legal tech company Harvey as an example: using synthetic data, public legal materials, and expert data to retrain open-source weights, starting without relying on client data to create courses.

The firm previously described vertical software as "labor becoming software": no longer competing for IT budgets, but for labor expenses on the profit and loss statement; companies in healthcare, legal, and housing sectors achieving over $100 million in recurring revenue within a few years, with finance and accounting following closely. In 2025, the focus will shift from retrieval to reasoning, and in 2026, the proposition will be multi-party collaboration—buyers, sellers, tenants, consultants, and suppliers each have permissions and compliance, only vertical software can transform handover rights from isolated intelligence into accountable processes.

In terms of capital, automation in insurance underwriting and claims like FurtherAI has already secured tens of millions in Series A funding led by them; Hebbia is used to illustrate financial process engineering, Abridge is present in examination rooms, and Harvey handles law firm contract archives. Simultaneously, evaluations like Vals AI have emerged in the industry: cutting-edge models still have over half failures in real financial analysis tasks, leaving gaps that can be filled by vertical suites at the "job level."

In market mechanisms, what is bought is repeatable labor that can be scored by experts, and what is sold is general chat access and traditional seat-based charging software. Funds are being diverted from lab pre-training budgets to industry workflows, evaluation sets, and permission integrations. The beneficiaries are vertical companies that can turn a single task into a data asset, while the pressured parties are incumbents who only manage record systems but fail to make cross-system judgments, and large models that stop at front-end dialogue and cannot enter compliant delivery at the distribution layer.

Supplementary structure: vertical transactions account for about half of the software and AI venture capital deals in North America, but large capital is still concentrated in foundational rounds like OpenAI and Anthropic; the competition at the job level is not about computing power scale, but who first writes the definition of "qualified work" into the loop.

Source: Public Information