01 / definition
A vertical agent is a workflow system, not a model wrapper.
The easiest mistake is to define a vertical AI agent by the model it uses. The model matters, but it is only one layer. The agent becomes vertical when it understands the narrow world it operates in: the customer, the account, the patient, the case, the booking, the property, the claim, the ticket, or the codebase.
A chatbot waits for a prompt. A horizontal assistant helps with generic knowledge work. A workflow automation follows a brittle rule. A vertical AI agent sits closer to the operating surface of the business: it reads context, chooses the next step, uses tools, follows policy, and knows when to stop.
Reason over language, images, code, data, and instructions.
Understand the customer, account, case, property, patient, issue, or workflow state.
Read and write in the systems where work actually happens.
Apply policies, permissions, approvals, logs, and escalation rules.
Move the business process forward instead of producing only a chat response.
02 / why now
AI adoption is broad. Production impact is still scarce.
The market has moved past the question of whether companies will use AI. McKinsey's 2025 survey says 88 percent of respondents report regular AI use in at least one business function, and 62 percent say their organizations are at least experimenting with AI agents. Yet the same survey shows the gap: only 39 percent report enterprise-level EBIT impact from AI.
That gap is the vertical-agent opportunity. The winners are not merely buying tools. They are redesigning work. McKinsey reports that AI high performers are nearly three times as likely as others to fundamentally redesign workflows. Stanford's AI Index also shows AI adoption accelerating across business, while MIT's AI Agent Index points to a fast-moving agent ecosystem where autonomy is rising faster than transparency norms.
McKinsey's 2025 survey says nearly nine in ten respondents report regular AI use in at least one business function.
The same survey says 23% are scaling agentic systems and 39% are experimenting with them.
AI is common, but enterprise-level financial impact is still limited, which points to deployment quality.
McKinsey reports AI high performers are nearly three times as likely to fundamentally redesign workflows.
03 / examples
The strongest examples are already vertical.
The most useful agent companies do not sell “AI” in the abstract. They sell an operating improvement inside a domain: legal work, clinical documentation, personal injury claims, software engineering, customer service, or go-to-market execution.
Sierra builds AI agents for customer service across channels, where the agent has to resolve customer work rather than just answer FAQs.
02Legal and professional servicesHarveyHarvey frames its product around legal and professional-service work: research, document analysis, contract intelligence, and purpose-built agents that execute legal work end to end.
03HealthcareAbridgeAbridge turns clinical conversations into documentation, decision support, orders, summaries, and downstream care workflow across health systems.
04Personal injury lawEvenUpEvenUp focuses on the personal-injury case lifecycle, including intake, treatment, demands, negotiation, discovery, trial work, and client communication agents.
05Software engineeringDevin by CognitionCognition describes Devin as an autonomous software engineer that plans, writes, tests, and ships production code inside the tools engineers already use.
06Customer lifecycleDecagonDecagon positions AI agents as customer-concierge infrastructure: build, test, observe, iterate, and scale workflows across chat, voice, and email.
07Go-to-market11x11x packages digital workers such as Alice the SDR and Julian the phone agent around specific revenue workflows instead of generic productivity assistance.
08Software developmentFactoryFactory's Droid is another software-development example: an agent-native interface for moving engineering work through missions, code, and enterprise workflows.
04 / value
CompanyOS is the argument for vertical AI agents.
A company does not need AI everywhere. It needs an operating layer where agents can safely execute useful work. That is the CompanyOS argument: the model becomes a component inside a managed system of workflow truth, domain memory, tool access, policy, approvals, feedback, and production ownership.
This is why the model itself is not the moat. Models will keep improving, getting cheaper, and converging. The durable advantage is what the company builds around them: the specific workflows the agent can run, the context it can trust, the tools it can touch, the decisions it must escalate, and the feedback loop that makes the next version better.
workflow truth
The agent starts with how operators actually work: exceptions, handoffs, approvals, customer language, and the systems that hold the source of truth.
domain memory
The operating layer accumulates rules, cases, customer context, preferred responses, feedback, and examples that make the agent more specific over time.
tool access
The agent is useful only when it can take safe action in the tools of record: CRM, inbox, ERP, database, ticketing system, calendar, or internal app.
policy control
CompanyOS defines what the agent may decide, what requires review, what must be logged, and what should be escalated to a human.
feedback loop
Every deployment should create a tighter loop between real work, operator correction, measurement, and the next version of the agent.
AI does not deploy itself.
The hard part is not opening a model endpoint. It is mapping the process, redesigning the work, connecting systems, setting permissions, training people, and deciding what the agent is allowed to own.
The model is not the moat.
Frontier models keep improving and converging. The durable advantage is the operating layer around them: proprietary context, domain memory, tool paths, approval rules, and feedback from real work.
The task is automated; the job is not.
An agent can draft, classify, route, summarize, or click. Someone still has to decide which button, in which system, under which customer context, with which risk boundary.
Cheaper work creates more worthwhile work.
When analysis, content, research, and code get cheaper, more projects become economically rational. The bottleneck shifts to deciding what to build and making it operational.
Waiting is the expensive move.
The companies that compound learning now will know which workflows matter, which data boundaries matter, and which agent controls matter before the late adopters finish procurement.
05 / deployment
Forward deployed engineering is how CompanyOS starts.
The fantasy version of AI transformation is simple: buy a model, turn it on, and remove a large part of the staff. The real version is harder and more valuable: pick one workflow, map the operating truth, redesign the handoffs, connect the tools, define the risk boundaries, launch the agent, and keep improving it where work is actually happening.
Most companies do not have senior engineers, operators, and AI builders sitting idle for three to six months to do that work. That is where deployed.engineering fits. We sit between business units, data, tools, and model APIs. We do not just recommend; we design, build, integrate, deploy, and iterate until the agent is a working part of the operating system.
research notes