The Workflow Is the Vertical AI Moat

Models are becoming infrastructure

Access to advanced AI models was briefly a competitive advantage. That advantage is disappearing as developers gain access to increasingly capable models from OpenAI, Anthropic, Google and others.

A vertical AI moat comes from the workflows, context, integrations and feedback built around the model. Model selection still matters for reasoning, latency, tool use, accuracy and cost. Most application companies are unlikely to train frontier foundation models themselves. Their value moves toward understanding a problem better than anyone else.

Domain knowledge becomes software

Consider two legal AI products built on exactly the same underlying model.

Company A gives lawyers a chat interface. Company B understands how due diligence is performed, which documents need review, which clauses matter, how exceptions are classified, which precedents should be referenced, how findings should be structured, who approves the final output and what happens after the review. Company B can encode that expertise into a repeatable workflow.

Harvey provides a useful real-world example. In February 2026, Harvey reported that customers had created more than 25,000 custom workflows incorporating organisational templates, examples and guidelines. By March 2026, Harvey described more than 400,000 agentic queries processed daily and positioned Agent Builder as a way for organisations to build reusable agents around their own processes. These figures are company-reported, but they show how much harder a working domain system is to replicate than a prompt.

The vertical AI flywheel

A strong vertical AI business can develop a flywheel: domain expertise improves workflow design; a more useful product drives usage; usage produces corrections and outcome data; those records strengthen eval datasets and system behaviour; better behaviour deepens adoption and returns more domain expertise.

The system learns from data produced by operations. Every interaction can show what users ask, where cases become difficult, which recommendations get rejected, where humans intervene, which steps create friction, which exceptions recur and what outcomes follow different decisions.

The result is a dataset about how work actually happens, including the decisions and outcomes a static corpus misses.

Integrations and workflow state

Vertical workflows rarely exist in isolation. A healthcare product may need electronic medical records, appointment systems, patient communication channels, devices, laboratory results, and task management. A logistics product may need fleet data, routes, customer requests, driver assignments, invoicing, and communications. A legal product may need document management, precedent libraries, matter management, billing, and knowledge repositories.

Every serious integration increases the context available to the AI and the actions it can safely perform.

Most early AI applications focus on conversational memory. Business software also needs state: what has already happened, what is waiting, who owns the case, what decision was made, what information is missing, and what comes next. A useful system cannot rediscover those answers every time someone opens a chat.

Underneath the intelligence still sit identity, permissions, records, workflow engines, databases, integrations, audit trails, and business logic. The intelligence layer does not replace software architecture. It makes that architecture more capable.

Evaluations become proprietary infrastructure

As a vertical AI system matures, another asset appears: the eval set.

Imagine an insurance underwriting platform that has accumulated 10,000 carefully evaluated scenarios covering normal applications, edge cases, policy exceptions, difficult documents, fraudulent submissions, and known historical failures. Each scenario contains expected behaviour. Whenever the company changes its model or workflow, it can measure whether the system got better or worse.

A new competitor may have access to the same model. It does not have those 10,000 tests. Anthropic argues that evals allow teams to identify regressions and determine much faster whether new models genuinely improve their products. For vertical AI companies, domain-specific eval datasets become strategic assets.

Outcome data closes the loop

Stripe shows what this looks like in a mature system. Payments create observable outcomes: accepted, failed, fraudulent, charged back, authenticated. Those outcomes feed risk systems.

Stripe reported in 2025 that adaptive Radar rules combining machine-learning risk information with issuer signals increased payment success rates by 1.3 percentage points with minimal change in fraud rates. At Stripe Sessions, Stripe described its AI as evaluating potential transaction pathways across variables such as authorisation probability, fraud risk, and refund likelihood.

The intelligence, workflow, and outcome are connected.

Build around the work

There will always be another model release, a larger context window, better reasoning, better tool use, and falling costs. AI application companies should benefit from all of that. They should not mistake model improvement for their own product advantage.

The durable opportunity is to understand a vertical deeply enough that the AI becomes embedded in how the industry operates: capture context, encode workflows, integrate systems, build evaluations, learn from corrections, measure outcomes, and keep improving the system around the intelligence.

The foundation model may provide the reasoning engine. The workflow becomes the product.