The general-purpose model narrative has dominated AI headlines for three years. Bigger, more capable, more general — that's the story the frontier labs tell, and they're not wrong that scale produces impressive results. But impressive results on benchmark suites are different from the output quality a clinical team trusts to inform a treatment decision, or that a risk desk trusts to flag a compliance breach. We've consistently backed vertical AI, and the thesis keeps getting stronger.
Why General Loses in Production
General-purpose models optimize for average performance across a huge range of tasks. That's what you want when you don't know what the task will be. But in a hospital, the task space is bounded. The vocabulary is bounded. The failure modes are known and documented. The regulatory framework specifies what the output can and can't contain.
A model trained broadly on internet text is also trained on a lot of noise, contradictory information, and outdated knowledge. For consumer applications where the cost of an error is low and users self-correct easily, that's fine. For enterprise applications where errors have legal, financial, or medical consequences, the noise in a general-purpose model's knowledge is a liability, not a feature.
We've seen this play out repeatedly in our portfolio diligence. One company we evaluated had benchmarked their domain-specific model against the leading general-purpose API for a contract analysis task. Their model was 34% more accurate on clause identification and 61% more accurate on jurisdiction-specific legal standards. It was also 5x cheaper to serve because it was a fraction of the size. That's not a marginal advantage. That's a different product.
Where Vertical Models Win
The domains where vertical AI consistently outperforms general models share a set of characteristics:
- Specialized vocabulary and notation. Medical, legal, financial, and scientific domains have terminology, abbreviations, and contextual conventions that general training data underrepresents. A model that has deeply internalized that vocabulary makes fewer errors on the core task.
- Structured output requirements. Many enterprise tasks require outputs in specific formats — FHIR-compliant clinical notes, structured SEC disclosure language, standardized contract clauses. A vertical model can be trained to produce these formats reliably. Getting a general model to do it consistently requires elaborate prompting and still fails at edge cases.
- Domain-specific reasoning chains. Evaluating a medical imaging finding, reviewing a patent claim, or modeling credit risk follows domain-specific logic. Models trained on domain-specific reasoning examples develop more reliable chains of inference for those tasks.
- Regulatory compliance as a training objective. This is underappreciated. The compliance requirements in a regulated industry can be baked directly into a vertical model's training — as hard constraints on outputs, as evaluation metrics, as negative examples. That's structurally different from prompting a general model to comply.
The Moat Question
The standard objection to vertical AI is: doesn't a more capable general model eventually catch up? Our view is that this question assumes a static target. Vertical models don't stand still — they continue training on domain data, benefiting from domain feedback loops that general models can't access. Clinical notes, legal rulings, financial filings — the organizations generating that data have reasons to keep it proprietary.
The companies that own proprietary domain data pipelines, or that have built tight feedback loops with domain experts who label outputs, are building a compounding advantage. The general model can improve generally. The vertical model improves specifically on the exact distribution of tasks the customer cares about. Those are not converging trajectories.
There's also a distribution advantage. Vertical AI companies sell to buyers who understand the domain. The sales conversation is about accuracy on specific tasks, compliance with specific regulations, and integration with specific workflows. That's a different conversation than "our model scores well on MMLU." It's a shorter sales cycle with a clearer ROI.
Our Investment Criteria
When we evaluate a vertical AI company, we look for three things beyond the standard team and market assessment:
Data advantage that's real and growing. Not a claim about proprietary data — evidence of an actual pipeline. Partnerships with domain data generators, feedback loops with domain experts, or historical archives that competitors can't easily access.
Evaluation infrastructure specific to the domain. A vertical AI company that can't tell us their task-specific eval suite in detail is not as far along as they think they are. The best teams have invested as much in how they measure their model as in how they train it.
A buyer relationship that creates switching cost. Vertical AI companies that get deeply integrated into a workflow — not just as an API endpoint but as a tool that domain experts actively shape through feedback — create retention that's hard to break. We look for early evidence of that kind of workflow integration.
The Broader Thesis
We think the market for vertical AI will follow the pattern of vertical SaaS. A wave of general-purpose tools gets built. Then domain-specific variants emerge that win in specific industries because they understand the workflow, the vocabulary, and the compliance requirements. The general tool stays relevant for general tasks. The vertical tool dominates for the specific ones.
The difference from vertical SaaS is speed. The tooling for building vertical AI — fine-tuning frameworks, adapter-based specialization, domain evaluation suites — has matured to the point where a competent ML team can build a differentiated vertical model in months, not years. The window for building that lead is open, but it closes as the tooling commoditizes further.
Building in vertical AI for healthcare, legal, finance, or developer tooling? We invest early in this space.
