What are Vertical AI Agents?
Vertical AI agents are specialized AI systems that deeply integrate into specific industry workflows—like Harvey for legal ($8B), Cursor for coding ($29B), and Abridge for healthcare ($5.3B). Unlike horizontal "do anything" assistants, vertical agents accumulate tribal knowledge, integrate with domain-specific systems (Epic, DMS, IDEs), and deliver Service-as-Software: selling labor outcomes rather than tool access. Enterprise spending on vertical AI ($10.8B) has crossed over to exceed horizontal AI ($8.4B).
The initial hypothesis of the Generative AI era was elegant: build a model capable enough, and it becomes a universal reasoning engine for all corporate tasks. One model to rule them all.
That hypothesis failed in production.
Enterprise spending on Generative AI hit $37 billion in 2025—a 3.2x increase year-over-year. But the allocation tells the real story. While infrastructure spending doubled, application spending exploded by a factor of eight. And within applications, spending on vertical and departmental AI ($10.8B) has crossed over to exceed horizontal AI ($8.4B).
The "do anything" agent was a transitional form. The market has decisively pivoted to vertical agents—specialized systems that don't merely assist but effectively inhabit the workflows of specific industries. This is The Agent Thesis in action: AI that replaces tasks, not tools.
The Horizontal Ceiling
General-purpose agents like ChatGPT Enterprise, Claude, and Microsoft Copilot are remarkable feats of engineering. They can write poetry, debug Python, and summarize financial reports with equal facility. But as enterprises moved from pilots to production in 2024-2025, the friction of generalist models became the primary bottleneck to value realization. This competitive dynamic—where specialized players outmaneuver generalists—is explored through AI Game Theory.
The problem isn't capability—it's context.
A horizontal agent knows the law in theory. It has read every legal textbook published on the internet. But it doesn't know how Allen & Overy drafts a credit agreement. It doesn't understand that "consideration" in a specific M&A context refers to a precise financial structure, not the general English definition.
A horizontal agent can transcribe speech accurately. But it doesn't know that confusing "Vincristine" with "Vinblastine" is a fatal error, or that distinguishing between a patient reporting a symptom and a physician prescribing treatment is the difference between accurate and dangerous documentation.
A horizontal agent can suggest code completions. But it can't spawn a shadow workspace, apply changes, run the linter, detect errors, self-correct, and retry until the code actually compiles—all before showing you the result.
These aren't edge cases. They're the difference between a demo and a deployment.
The Economic Shift: Service-as-Software
The economic engine driving vertical agents is a fundamental transformation in how software creates value.
Traditional SaaS: Sell access to a tool. Per-seat pricing. Salesforce charges for CRM access, but users still perform data entry, analysis, and outreach. Software value is capped by the productivity gain it offers—typically 10-20%.
Vertical AI: Sell the labor itself. Per-outcome pricing. Intercom's Fin AI charges $0.99 per resolved support ticket. If the agent fails to resolve the issue, the customer pays nothing.
This shifts risk from buyer to vendor and aligns incentives completely. More importantly, it expands the addressable market by an order of magnitude.
A law firm might spend $50 million annually on software but $500 million on associate salaries. Traditional SaaS captures 1-5% of an employee's salary equivalent. Vertical AI captures 25-50%. Harvey isn't competing for the software budget—it's competing for the labor budget.
This is the transition from Software-as-a-Service to Service-as-Software.
The Tribal Knowledge Moat
Every industry operates on two layers of information:
- Explicit knowledge: Documented rules, published procedures, official guidelines
- Tacit knowledge: Undocumented intuition, historical context, "the way things are done"
Horizontal models, trained on the open internet, possess the explicit knowledge of the world. They know the law, medicine, and code in theory.
Vertical agents accumulate the tacit knowledge of specific firms and industries. As they execute workflows, they generate proprietary data—failed drug discovery paths, successful legal arguments, resolved medical coding disputes—which feeds back into the model.
This creates a flywheel. More usage generates more proprietary data, which improves the model, which attracts more usage. Generic foundation models can't replicate this without the workflow integration that generates the data in the first place.
The Hollow Firm
The rise of vertical agents supports the "Hollow Firm" hypothesis—a future organizational structure where high-revenue enterprises operate with dramatically leaner headcounts, particularly at the junior level.
In law, finance, and consulting, junior staff spend years doing grunt work (document review, data entry, formatting) to learn the trade. Vertical agents are automating this layer entirely.
The result is a "barbell" workforce: a small number of senior experts leveraging AI agents to do the work of hundreds of juniors. For enterprises, this means massive margin expansion. For vertical AI vendors, it cements their product as mission-critical infrastructure—an operating system for the firm. For the full analysis of what happens when junior roles disappear, see The Hollow Firm 2.0.
The Legal Vertical: Harvey
Harvey has become the category-defining example of vertical AI, reaching an $8 billion valuation with $760 million in funding. This valuation reflects the market's belief that Harvey isn't a software tool but a platform capable of capturing significant percentage of the global legal services market.
For the complete story of how Harvey reached $8B by solving the trust problem with Vault architecture and 0.2% hallucination rates, see our Harvey deep dive. For how they're expanding into workflow automation, see Harvey's Workflow Builder.
The Vault Architecture
The primary objection to Generative AI in law is data privacy. Law firms are legally and ethically bound to protect client confidence. Using a public model creates real risk of data leakage or training contamination.
Harvey solved this with its proprietary "Vault" architecture:
- Firms upload sensitive datasets—decades of M&A deal structures, litigation playbooks, internal memos—into a secure, isolated environment
- Models perform RAG against this isolated Vault without data leaving the firm's perimeter or training the base model for other clients
- The AI drafts documents that mimic the firm's specific style and strategic preferences
This is impossible for a generic model. ChatGPT doesn't know your firm's voice. Harvey does.
Citation-Backed Generation
In law, hallucination is malpractice. Inventing a case or statute can end careers and invalidate work product.
Harvey employs a neuro-symbolic approach: when answering a legal query, it decomposes the query into factual claims and cross-references them against authoritative sources (leveraging partnerships with LexisNexis). The system provides verifiable citations for every claim, allowing lawyers to click through to source text.
Internal evaluations reportedly show this method reduces hallucination rates to approximately 0.2%—far below unaugmented LLMs.
A&O Shearman Deployment
The validation lies in adoption metrics. A&O Shearman (Allen & Overy), one of the world's most prestigious law firms, deployed Harvey to 4,000 lawyers across 43 jurisdictions. This wasn't a limited pilot—it was a firm-wide infrastructure overhaul.
Results:
- 30% reduction in contract review time
- 2-3 hours saved per week per lawyer on routine tasks
For a firm that bills by the hour (or increasingly by fixed fee), this is a massive productivity lever. Harvey effectively created thousands of new hours of capacity without increasing headcount.
The Horizontal Failure
Microsoft's "Copilot for Legal" has struggled to gain the same traction among elite firms. The reason is context depth. A horizontal tool cannot easily ingest millions of documents from a firm's Document Management System and index them with the semantic understanding of legal concepts required for high-stakes work. For the full competitive landscape and why legal represents a unique AI opportunity, see State of Legal AI.
Harvey's specialized embeddings (like voyage-law-2-harvey) provide retrieval accuracy that generic tools cannot match. The moat isn't intelligence—it's context. This pattern—where domain expertise creates defensible moats—is what we call The Legal AI Exception.
The Healthcare Vertical: Abridge
If the legal vertical is about automating text generation, healthcare is about automating attention. The primary crisis in modern medicine isn't a lack of knowledge—it's a lack of time.
Physicians spend nearly two hours on documentation ("pajama time") for every hour with patients. This administrative burden drives burnout and medical error. Ambient Clinical Intelligence—technology that listens to patient encounters and autonomously generates documentation—is solving this.
For the complete story of how Abridge reached $5.3B by solving the pajama time crisis with Epic integration and medical-specific AI, see our Abridge deep dive.
The Epic Moat
Abridge has emerged as the leader, securing a $5.3 billion valuation following a $300 million Series E. Unlike horizontal transcription tools, Abridge positions itself as a "Clinical Operating System" mediating interaction between physicians and Electronic Health Records.
The single most significant barrier to entry is integration with Epic Systems, which holds records of over 250 million US patients.
Abridge achieved "Deep Integration":
- AI parses conversations into structured fields (HPI, ROS, Physical Exam, Assessment & Plan)
- Inserts directly into Epic workflow
- Queues orders based on conversation ("Order CBC and metabolic panel")
A generic agent like ChatGPT Enterprise has no mechanism to securely write data into a hospital's Epic instance. The technical and regulatory friction of achieving this integration serves as a massive defensive moat.
The Long Tail of Medical Specificity
Why not use OpenAI's Whisper? It's cheap and accurate.
The answer lies in the long tail. Abridge published comparative studies showing the delta:
- 45% relative error reduction on newly FDA-approved medications vs standard ASR
- Contextual reasoning distinguishes a patient reporting a symptom from a physician prescribing it—a conflation that generic models make routinely
- Linked evidence: every claim hyperlinked to the specific audio timestamp where evidence exists
In medicine, confusing "Vincristine" with "Vinblastine" is fatal. Generic accuracy isn't good enough when the long tail includes medication names that sound similar but do radically different things.
Regulatory Moat
Healthcare is governed by HIPAA and GDPR. Vertical agents are architected for this reality: Business Associate Agreements, zero data retention for model training, strict PII redaction pipelines.
Horizontal agents, while improving, often default to data retention policies that are non-starters for large health systems. The "Compliance Shield" is core to the vertical value proposition.
The Coding Vertical: Cursor
Software engineering is arguably the most mature market for AI agents—and the most instructive battleground.
GitHub Copilot is the fastest-growing SaaS product in history. But in 2025, Cursor matched its milestone, growing from $1M to $100M ARR in under two years.
The difference is architectural.
Plugin vs Environment
Copilot (Horizontal): AI is a plugin. It lives in the sidebar or as "ghost text" overlay. It helps you write the next few lines of code.
Cursor (Vertical): AI is the environment. Cursor is a fork of VS Code—it is the editor. This gives the agent "root access" to the developer's workflow, file system, and terminal.
The Shadow Workspace
Cursor's dominance is built on a specific architectural innovation.
The problem: When AI writes code, it often makes syntax errors or breaks dependencies. In a plugin model, you paste the code, run it, see the error, prompt the AI to fix it. This loop is slow and manual.
The solution: Cursor spawns a hidden background instance of the project—a "Shadow Workspace." When the AI generates a solution, it applies changes in the shadow realm first. It runs the project's linter and compiler.
If the shadow linter throws an error, the AI sees it, self-corrects, and retries. This cycle repeats until the code is valid. Only then is the solution presented to the user.
This "Agentic Loop" creates a user experience that feels like magic—code that just works—and is technically impossible for a simple plugin to replicate due to lack of deep system access.
Model Agnosticism
Microsoft locked Copilot users into OpenAI models. When Claude 3.5 Sonnet emerged as superior for coding tasks, Copilot users were stuck.
Cursor, being model-agnostic, allowed instant switching. This captured the power user market that felt constrained by Copilot's GPT-4 dependence.
The lesson: vertical agents can be flexible about the model layer precisely because their moat is in the workflow layer.
For the complete story of how forking VS Code led to $1B ARR in 24 months, see our Cursor deep dive. For the other side of the coding agent debate—autonomous agents that work in the background rather than assisting in real-time—see our deep dive on Cognition AI's Devin.
The Finance Vertical: Policy as Code
In finance, vertical AI is transforming the office of the CFO by moving from expense tracking to autonomous spend management.
Ramp's Policy Agent
The problem: Corporate expense policies are PDF documents nobody reads. Auditing is manual, reactive, and universally hated.
The solution: Ramp allows finance teams to define policy as logic ("No travel expenses over $200 without pre-approval"). The Policy Agent sits between transaction and ledger.
When a receipt is uploaded, the agent extracts data via OCR and LLM (Merchant, Date, Line Items, Alcohol presence), then evaluates against policy logic.
Results:
- 85% of transaction reviews automated
- 3x more out-of-policy spend captured vs traditional rule-based systems
The improvement over rule-based systems comes from understanding nuance—distinguishing between a "Client Dinner" and a "Team Party" based on receipt details.
Klarna: Service Automation at Scale
Klarna provides a parallel case study. In 2025, Klarna's AI assistant handled 2.3 million conversations—equivalent to 700 full-time agents.
The agent wasn't just a chatbot. It was connected to Klarna's banking backend: processing refunds, changing billing dates, explaining specific line items.
Financial impact: $40 million profit improvement, resolution times dropping from 11 minutes to 2 minutes.
This is the Service-as-Software thesis in action: Klarna didn't buy software to help its agents. It built software to replace them.
Outcome-Based Pricing
Intercom's Fin illustrates the business model transformation. Priced at $0.99 per resolution, it represents ultimate incentive alignment.
If the AI hallucinates or fails to help, the customer pays nothing. This pricing model is only possible because Intercom has high confidence in the vertical capabilities of its agent.
Customer support becomes a variable cost utility rather than a fixed headcount expense. For a complete analysis of the $50B support automation wave, see Customer Support Agents.
The Technical Architecture
The superiority of vertical agents isn't magic—it's architecture. A distinct "Vertical Stack" has emerged.
| Layer | Horizontal | Vertical |
|---|---|---|
| Model | Off-the-shelf frontier LLMs | Fine-tuned or RAG with domain embeddings |
| Context | User prompt history | Vault (full case files), EHR history, entire codebase |
| Action | Text generation | System of Action (API calls, SQL, file edits) |
| Verification | None / user review | Shadow Workspace, linters, citations, policy logic |
| Compliance | Terms of Service | BAA, SOC2 Type II, zero-retention, audit trails |
The Shadow Loop
The defining technical innovation is the Shadow Loop (Agentic Loop):
- Think: Plan the action
- Act: Execute in sandbox/shadow environment
- Observe: Check for errors (linter, policy violation, hallucination)
- Correct: Fix errors
- Output: Only present when checks pass
This loop transforms the probabilistic nature of LLMs into the deterministic reliability required by enterprise workflows.
Neuro-Symbolic Guardrails
Vertical agents increasingly use a neuro-symbolic approach:
- Neural (LLM): Understanding language and intent
- Symbolic (Code): Enforcing rules
In Ramp, the LLM reads the receipt, but a hard-coded Python script checks if the amount exceeds $50. This combination ensures the agent cannot be "talked into" breaking rules via prompt injection.
The LLM handles the unstructured interpretation; the code handles the non-negotiable constraints.
The Three-Way War
The rise of vertical agents has triggered a three-way battle for the enterprise interface. For a comprehensive view of the vendor landscape across all tiers, see the Top 100 AI Agent Companies.
Incumbent Response: Salesforce Agentforce
Salesforce recognized the threat early. If Harvey or Abridge becomes the interface, Salesforce becomes a "dumb database" backend.
Their thesis: "Data is the Moat." An agent is only as good as the data it can access. Since Salesforce holds the customer record, its agents have immediate, grounded context for sales and service tasks.
Salesforce is rolling out industry-specific agents (Agentforce for Healthcare, Manufacturing) to compete directly with startups.
Advantage: Distribution—they're already on the contract. Disadvantage: Depth—they can't match a pure-play's specialization.
Foundation Model Threat: OpenAI Operator
The bear case for vertical AI is that foundation models become capable enough to handle vertical tasks without specialized middleware.
OpenAI's Operator and similar "Computer-Using Agents" represent this threat. They can look at a screen and click buttons, navigating software visually. This potentially bypasses the need for API integrations. If Operator can "use" Epic or "use" Salesforce like a human, does the vertical agent lose its moat?
The consensus in 2025: highly regulated verticals (Law, Medicine) require liability protection and workflow crafting that OpenAI is unlikely to provide. OpenAI prefers to be the platform, not the liability holder. The risk of a hallucinated legal argument or medical recommendation is existential for OpenAI's brand—better to let Harvey and Abridge absorb that liability.
Google's Hybrid Approach
Google is taking a different path: "Industry AI" accelerators. They provide models (Gemini, MedLM) and infrastructure (Vertex AI) but rely on partners to build the final vertical mile.
This appeals to massive enterprises (Mayo Clinic, Bayer) who want to build their own vertical agents rather than buying off-the-shelf. Google becomes the platform layer without absorbing vertical-specific liability.
The Vertical Thesis
The evidence is conclusive: vertical agents are winning because context is king—and more importantly, because verification is domain-specific.
Enterprises don't buy "intelligence." They buy solutions. General intelligence is a commodity; applied intelligence is a product. The question of who aggregates the AI market isn't "who has the most context?"—it's "who can verify quality?" Legal has citations. Code has tests. Healthcare has clinical standards. General knowledge work has... human judgment. That's why horizontal platforms can't aggregate the way Google aggregated search.
Horizontal agents are tourists in the enterprise. They can visit any department, speak any language, attempt any task. But they don't belong anywhere. They lack the tribal knowledge, the workflow integration, the compliance frameworks that make AI usable in high-stakes environments.
Vertical agents are natives. They were born in the workflow. They speak the domain language. They understand what "consideration" means in M&A, why "Vincristine" and "Vinblastine" must never be confused, why code must compile before it's shown to the user.
The winners of the AI era will be those who constrain the infinite possibilities of Large Language Models into specific, reliable, and compliant workflows. By solving the "Last Mile" problem through:
- Deep integration with domain systems (Epic, DMS, IDE)
- Proprietary data vaults that accumulate tribal knowledge
- Agentic shadow loops that transform probabilistic into deterministic
- Compliance frameworks that absorb liability
These companies are building the new operating systems of the economy. They're transitioning software from a $600 billion tool market to a multi-trillion dollar labor market.
The era of SaaS is ending. The era of Service-as-Software has begun.