Why the orchestration layer wins

Enterprise AI has three layers. Models supply capability. Systems of record, the CRM, the warehouse, the ticketing system, supply state. Between them sits the execution layer: identity, permissions, tool access, workflow orchestration, and the trace of what actually happened when the work ran.

Two of those layers are spoken for. The models are a fierce, well-funded race with several serious contenders. The systems of record are entrenched and not going anywhere; no one is ripping out the warehouse. The contested layer, the one that is up for grabs, is the one in the middle, where work actually runs across everything else.

It helps to be specific about the job, because it is unglamorous and hard. The execution layer has to authenticate agents through the customer's own identity provider and enforce the customer's existing permissions on every action. It has to reach heterogeneous systems, each with its own authentication and data model, under one coherent governance. It has to capture a complete, auditable trace of every action across all of them.

And it has to do all of that while staying neutral about which model handles which step. That last requirement looks like a detail. It is the load-bearing one, and it is where the whole argument turns.

It is worth dwelling on why the middle layer is hard, because the difficulty is exactly what makes it defensible. Every system the agent touches has its own identity model, its own permissions, its own idea of what a given user is allowed to see. Unifying all of that under one governance, so that an agent inherits the right person's access in every system and every action is traced the same way, is a large and unglamorous body of work.

It is also orthogonal to model quality. A better model does not help you reconcile a CRM's permission model with a warehouse's. That reconciliation has to be done by someone whose whole job is the layer, not the model, and it does not get easier when the models get smarter. It is a different kind of problem, and it rewards a different kind of company.

Enterprises have learned to fear lock-in, and model lock-in is the sharpest kind, because model capability and price move every quarter. A buyer who commits their execution layer to one vendor's models is betting that vendor stays ahead forever, and no one stays ahead forever.

So the buyer wants an execution layer that routes each task to whatever model is best and cheapest that week, and can change its mind the next week without re-platforming. Neutrality is not a nice-to-have feature on top of that layer. It is the thing the buyer is actually trying to purchase: the freedom to keep using the best model without rebuilding around it each time the best model changes.

Concretely, imagine the best model for a class of work is one provider's this quarter and a different provider's the next, which is roughly how the last two years have gone. A neutral layer reroutes and the customer never notices. A layer wired to one vendor either keeps sending the work to a now-second-best model, or the customer eats a migration. The neutral layer is the one that makes the quarterly churn somebody else's problem.

Now the structural point. A model vendor sells inference. Its incentive, at every level of the company, is to route work to its own models. That is not a criticism. It is the business.

An execution layer that does its job honestly will sometimes route away from the vendor that built it, to a competitor's model, because that is what is best for the task. So a model lab building a neutral execution layer is building a product whose correct behavior is to sometimes prefer a competitor. That is not something a company ships against its own core business, not for long and not wholeheartedly. The labs will build agent runtimes, and they will be good, and they will route to their own models, which is exactly the lock-in the enterprise was trying to avoid in the first place.

The strongest objection is that a lab could simply build or buy a neutral layer and run it at arm's length. They could build the technology, certainly. The problem is not the engineering. It is the trust.

A neutral layer's entire value is that the buyer believes its routing serves the task and not the vendor. A model vendor cannot credibly make that promise about a layer that also feeds its own inference revenue, however the org chart is drawn. The neutrality has to be believed, and belief is the one thing a conflicted owner cannot manufacture. The day the layer would have routed away from the house model is the day the conflict becomes visible, and the buyer knew it was coming.

This is why we do not think of ourselves as racing the labs. We adopt every model they ship, the day it is better, the same way we adopt every framework and method. A stronger frontier model makes the execution layer more capable, not less necessary.

The labs supply capability. We supply the neutral, permission-aware layer that turns capability into governed work, across all of a customer's systems and under their own rules. We get better every time the labs do. The relationship is not rivalry. It is supply, and a supplier improving is good news for the thing built on top of it.

So the real question for the enterprise deployment layer is not whether the labs can ship agent infrastructure. They can, and they will. It is whether they will build permission-aware trace capture across heterogeneous enterprise systems, under each customer's entitlements and governance, while staying neutral across competing model providers. That is a different company than a model lab, with a different business model and a different reason to exist.

The winner of enterprise AI is not the smartest model. It is the layer that lets an enterprise use the smartest model this quarter, and a different one next quarter, across all of its systems, under its own governance, without rebuilding anything. That layer has to be neutral, and neutrality is a position a model vendor cannot hold. Capability will keep moving between labs. The execution layer is where it gets put to work, and that is the position worth owning.