Ve Vertical AI software Loading... : Investor Sentiment and Bull/Bear Views

As agents shift from chat to background and asynchronous work, the interface becomes dashboards, queues, and task lists rather than a universal chatbot. Horizontal chat products will become heavy and less delightful, while vertical AI players that understand a specific process can manifest the right buttons, tabs, and workflow names. This favors vertical AI software across legal, security, coding, and other fields; he predicts 90% of enterprise tokens in five years will be for work no human user initiated.

The final evolution is vertical AI: industry-specific models trained on large domain data and delivering results for finance, manufacturing, and other sectors. The key question is which domain-specific AI software winner emerges in each industry.

Sacks says generalized AI models failed in roughly 95% of large-enterprise deployments while vertical applications, domain-specific models and smaller specialized models (SLMs) showed much greater success. The reason is last-mile work: LLMs need connections to enterprise data, very detailed prompting, hallucination validation and iteration, and moving from about 90% to 99% accuracy requires real industry knowledge. He concludes that value will be captured by many vertical applications and specialized models across many different markets rather than by one foundation model eating all the value, and calls that healthy for the ecosystem.

Sacks reads the enterprise survey result as evidence that AI value accrues to vertical applications and smaller specialized models rather than to one general foundation model. Generalized model deployments failed about 95% of the time because LLMs need enterprise context, detailed prompting, hallucination validation and iteration, which he calls the last-mile problems, while vertical applications, vertical models and SLM approaches with a tighter problem and data set showed much greater success. He argues that going from 90% to 99% effectiveness requires industry-specific knowledge, so the outcome is lots of vertical applications and specialized models capturing value across many separate markets instead of a single foundation model eating all the value, which he considers healthy for the ecosystem.