Horizontal AI platforms are winning the headlines. Vertical AI agents are winning the contracts.
Harvey, the legal AI agent, hit $50M ARR without selling a single seat to a non-lawyer. Abridge, the clinical documentation agent, is deployed in more than 50 major health systems. Glean, the enterprise search agent, crossed $100M ARR serving knowledge workers inside companies — not across them.
Meanwhile, Microsoft Copilot — the most aggressively distributed horizontal AI product in history — is showing early adoption data that is underwhelming relative to its install base. Enterprises are paying for vertical AI agents because they solve specific, measurable problems. Horizontal tools require enterprises to figure out the problem themselves.
What “Vertical AI Agent” Actually Means
A vertical AI agent is purpose-built to own a workflow inside a specific industry. It is not a ChatGPT wrapper with a different system prompt. The defining characteristics are:
Domain-specific training data
Proprietary case law, clinical notes, financial filings, engineering specs — data the foundation models were never trained on or cannot access
Workflow ownership
The agent operates inside the existing system of record (EHR, DMS, trading terminal) rather than requiring a context switch to a chat interface
Feedback loop accumulation
Every use generates labeled data (attorney edits, physician corrections, audit overrides) that improves the agent and cannot be easily transferred to a competitor
Compliance and audit trails
Regulated industries require explainability and documentation that horizontal platforms do not natively provide — verticals build this in
The Retention Gap: Vertical AI Agents vs Horizontal Platforms
Net revenue retention is the cleanest signal for workflow ownership. When an AI product is embedded in the actual job, expansion is automatic. When it requires deliberate user initiative, expansion stalls.
| Product type | Typical NRR | Why |
|---|---|---|
| Vertical AI agent (workflow-embedded) | 130–160% | Expansion follows headcount + usage grows as agent improves |
| Horizontal AI copilot (general-purpose) | 100–115% | Usage is opt-in; expansion requires user champions |
| Traditional SaaS (non-AI) | 105–125% | Expansion from new modules and seat additions |
| AI API (developer tools) | 120–140% | Usage-based billing benefits from product growth, but elastic churn |
Sources: Bessemer Venture Partners State of the Cloud 2025, OpenView SaaS Benchmarks 2025, company disclosures.
Where Vertical AI Agents Are Already Winning
The most defensible vertical AI companies share a pattern: they entered at a workflow that was too complex for horizontal tools to automate, accumulated proprietary feedback data, and then expanded into adjacent workflows. This is the opposite of the “build a horizontal platform and add verticals later” strategy most AI labs are pursuing.
Trained on proprietary case law and internal firm documents. 50+ AmLaw 100 firms as customers. Attorney edits create feedback loops unavailable to general models.
Ambient clinical documentation embedded in Epic EHR. Physician corrections and clinical note structure create training data that cannot leave the EHR environment.
Indexes company-specific documents, Slack, email, and code. Knowledge graph is proprietary to each company and cannot be replicated by a horizontal ChatGPT integration.
Trained on financial documents, 10-Ks, earnings transcripts, and investment memos. Analyst corrections create feedback loops in a domain where accuracy is legally material.
Why Horizontal Platforms Are Structurally Disadvantaged
OpenAI, Anthropic, and Google are building the intelligence layer. Vertical agents are building the workflow layer. These are not the same business, and they are not in direct competition — yet.
The disadvantage for horizontal platforms is structural. A general-purpose agent trained on internet-scale data cannot perform at the accuracy level required for legal filings, clinical documentation, or financial audit work. The gap is not a model size problem — it is a proprietary data problem. The training signals that make a vertical agent accurate (attorney edits, physician corrections, audit outcomes) are locked inside regulated enterprise environments that the foundation model companies cannot access.
This is why Microsoft Copilot — which has distribution through every Microsoft 365 seat — is struggling to show the usage numbers that justify its $30/user/month premium. The intelligence is there. The workflow ownership is not.
What vertical agents get right
- ✓ Pre-trained on proprietary domain data
- ✓ Embedded in existing system of record
- ✓ Feedback loops that compound over time
- ✓ Built-in compliance and audit trail
- ✓ Pricing justified by measurable ROI
What horizontal platforms miss
- ✕ Generic training data for specialized domains
- ✕ No workflow ownership — bolt-on not embedded
- ✕ User-driven adoption, not system-driven
- ✕ Compliance and audit requires custom work
- ✕ Value proposition is vague at the task level
What This Means for Investors
The market is pricing vertical AI at a premium — and correctly so. Harvey is valued at a higher multiple than most horizontal AI tools because its proprietary data moat is real, its NRR is above 140%, and its TAM is the $1T+ legal services market rather than the enterprise software market as a whole.
At Value Add VC's AI valuations dashboard, you can see that the highest-valued private AI companies are disproportionately vertical: Harvey, Abridge, Hebbia, Coframe. The horizontal infrastructure plays (Mistral, Cohere) trade at lower revenue multiples despite larger fundraises, because the commodity risk from foundation model providers is existential.
The investment thesis is straightforward: back vertical AI agents that own the feedback loop. The companies accumulating domain-specific labeled data at scale will command the pricing power and retention that justify premium multiples. The companies that are simply routing calls to foundation model APIs with a different system prompt will get commoditized the moment OpenAI or Anthropic enters their vertical directly.
The AI platform wars are not being won at the intelligence layer.
They are being won at the workflow layer — by the agents that own the feedback loop inside the work itself.
Track valuations for vertical and horizontal AI companies on the AI Valuations Dashboard at Value Add VC. See where enterprise AI spending is concentrated on the AI Spending tracker. Originally published in the Trace Cohen newsletter.