The Value in Enterprise AI Is the Implementation Layer
The model was never the bottleneck. The workflow design, data access, authority, evals, and audit around it are where the value — and the capital — is flowing.
Watch where the capital is going and the story tells itself. Frontier labs are standing up deployment companies backed by billions in private-equity money. Consultancies are building agentic delivery practices. Systems of record are exposing agent interfaces. All of it is converging on one thing — and it isn't the model.
The bottleneck is how agents are built and operated
Even the labs now say it out loud: the constraint on enterprise AI isn't model capability, it's the implementation layer inside the company. That layer is specific, and it's mostly not model work:
- Workflow design — which decisions the model makes, which stay human, where the handoffs are, what counts as done.
- Data access — which sources are authoritative, which are stale, what permissions apply at the row and field level.
- Authority — what the agent may do, against which systems, with what limits. Reading, writing, and spending are three different risk profiles.
- Evals, audit, and recovery — how you know the output is correct before it ships, what gets logged, and who reverses an action that went wrong.
Why the money is moving now
Private equity spent years buying SaaS companies that all "tasted like chicken" — interchangeable, easy to analyze. That model broke when agents arrived, so PE is racing toward the completed workflow, where the disproportionate value now sits. The catch is that this value is custom. It can't be cloned in a weekend, and it can't be bought off a shelf.
The buyer's test
When a vendor says "our model is great, just trust your data to us," they're selling last year's market. Sit closer to the business object and probe the implementation layer: can they show you the workflow definition, the eval, the authority model, the audit trail? The ones who can't will reveal their cards. The value belongs to whoever assembles model, harness, and data into an actionable workflow — and that is usually built, not bought.