Alexander Niehenke: Strengthening Vertical AI
Alexander Niehenke believes that the AI industry is shifting from competition based on general models to vertical AI that targets specific industries and has data and process barriers.
Vertical AI is not merely a smaller model; it integrates foundational models with proprietary data, terminology systems, compliance rules, and specific workflows in fields such as healthcare, law, finance, or manufacturing.
General models provide underlying reasoning capabilities, but enterprise clients purchase systems that can directly complete tasks, such as processing medical records, reviewing contracts, detecting fraud, or automating policy processes, rather than just acquiring chat capabilities.
Defensibility comes from proprietary data and action loops: products initially penetrate through document retrieval, voice assistants, or content generation, then connect to enterprise record systems, accumulating new data through ongoing task execution, ultimately competing for system record rights.
Vertical AI can also use smaller or fine-tuned models to reduce inference costs, improve latency, and minimize compliance exposure; the trade-off is a decrease in cross-industry generalization ability and an increased reliance on data quality, integration capability, and industry sales ability.
In market mechanisms, foundational model companies sell computing power, APIs, and general capabilities, while vertical AI companies purchase these underlying capabilities and sell quantifiable results and automated labor to industry clients. Beneficiaries are application companies that have industry distribution, data access, and workflow control; those under pressure are generic AI layers that only provide homogenized interfaces.
Source: Public Information
ABAB AI Insight
The previous success logic of vertical software was established by companies like Veeva and Procore around industry processes to build record systems; the change in the AI era is that products no longer just manage processes but directly handle unstructured work previously done by humans. The path proposed by NEA is to first penetrate clients with voice agents, semantic search, or content generation for free, then connect to existing record systems, and subsequently establish new system records through new data generated by agents executing processes.
Capital is reallocating funds from foundational models to companies that can embed models into high-value industries. Legal AI company Harvey raised $200 million in funding in 2026, with a valuation of $11 billion, backed by investors including GIC, Sequoia, Andreessen Horowitz, and Coatue, with total funding exceeding $1 billion; its expansion targets are intelligent workflows for law firms and corporate legal departments.
Healthcare AI company Abridge represents another capital path: its product automatically generates clinical records for doctors, completing a $300 million Series E funding round in June 2025, with a valuation of $5.3 billion; just four months prior, the company completed a $250 million Series D funding round at a valuation of $2.75 billion. The commonality between law and healthcare is high labor costs, dense unstructured documents, and clear error liability, making it difficult for general chat tools to directly replace vertical delivery systems.
This represents a restructuring of the industry chain. Foundational models are gradually becoming purchasable capability layers, with competitive advantages shifting from "who trains a larger model" to "who has real task entry points, proprietary feedback data, and result accountability." As AI can enter transaction, diagnosis, audit, or compliance processes, the existing SaaS pricing logic based on seats will be pressured by pricing models based on tasks, results, or hours saved.
ABAB News · Cognitive Laws
Models will be commoditized, workflows will not
General intelligence sells capabilities, vertical intelligence sells results
Data determines the upper limit, entry points determine pricing power