Ingestion, permission-aware retrieval, structured output, review workflow, audit trail — every vertical AI product needs all five before a customer sees a single feature that's actually theirs. Most teams build all of it from scratch, for every vertical, every time.
The same capability, two ways: nine months of custom infrastructure milestones on the left, versus a declarative blueprint — topology, taxonomy, tasks, scope bindings — deployed to a first client in fourteen days on the right.
Ingestion pipeline, vector index, permission-aware retrieval, structured output validation — all of it has to exist before a customer sees a single feature that's specific to your vertical.
Legal, support, procurement, compliance — the underlying pieces are nearly identical, and nearly every team building one rebuilds them independently because there was no shared engine underneath.
The actual expertise you're selling — the document taxonomy, the review workflow, the domain judgment — sits on a roadmap behind months of retrieval and permissions engineering that looks the same for every competitor.
It gets you retrieval, not a vertical product. Permissions, structured output contracts, review workflows, and audit trails are still yours to build on top, and that's most of the actual engineering effort.
The fork diverges immediately, and now there are two codebases to maintain instead of one configurable engine — every fix to one has to be manually ported to the other.
More engineers on the same undifferentiated plumbing is still months of work before the differentiated part starts — headcount doesn't change what has to be built first.
Retrofitting permission-aware retrieval onto a live index means re-tagging and reindexing everything you already ingested — usually right when your first enterprise customer's security review demands it.
Knowledge Base, Document, Blueprint, Workspace, Run — the same five primitives handle legal, support, procurement, and compliance. A new vertical is a new blueprint, not a new codebase.
Document taxonomy, task output schemas, and scope bindings are declared in a blueprint release. Retrieval, permissions, review, and audit come from the engine underneath, already built.
The domain taxonomy, the task output contract, the instruction modules — the parts that are actually your expertise — are what you spend the first weeks on, not the ingestion pipeline.
Ship a master blueprint to 100 clients via copy-on-distribution — never a shared reference that would leak one publisher's policy text into every client's index.
Retrieval, permission scoping, review workflow, and audit trail are already built — the blueprint declaration is what's left, and that's days of work, not months.
The run viewer, workspace drop-zone, and task output renderer ship as embeddable components — your product's UI wraps them instead of reimplementing them.
Permission-aware retrieval and the audit trail exist on day one of a new vertical, not as a scramble the first time an enterprise prospect's security team asks about them.
