The Orchestrator Is More Than a Connector
The strategic subject here is the relationship between a vertical AI firm and the general-purpose agent that brings its capabilities to users. The specific question is why an open protocol can make a specialist easier to call while leaving the orchestrator with power over discovery, routing, substitution, and learning. Seeing those control points allows a decision-maker to distinguish technical interoperability from economic independence and to locate defensibility in the assets that make an output trustworthy.
A protocol answers a narrow question: how can one system request a capability from another? It does not, by itself, answer who decides which capability is shown, which one is called first, how a failure is judged, or where the customer’s relationship accumulates. Those decisions sit above the connection layer. They are where the orchestrator can become more than a neutral relay.
How an open protocol becomes a control surface
In the open-protocol pattern, the orchestrator aggregates users, tools, data sources, and third-party capabilities at the surface where work is requested, routed, and completed. That combination gives it a platform position. The vertical AI firm may own a specialized function, but the orchestrator owns the place in which the customer asks for that function and sees the result.
This position matters because the user’s request is not just an input. It is the beginning of a sequence of choices. The orchestrator can decide which tools are eligible, which provider is the default, how alternatives are described, and whether a result is accepted as sufficient. It may coordinate a capability from the vertical firm with data from one provider and actions from another. The specialist is present in the workflow, but it need not control the workflow’s order or the customer’s interpretation of the output.
Platform envelopment describes one way this power develops. An actor with a shared user base and complementary capabilities absorbs another platform’s functionality into its own offering. In an agent ecosystem, the orchestrator can build a native capability that substitutes for the vertical tool. It does not need to reproduce every detail of the specialist’s product. It only needs to make the customer’s task feel complete without the separate component. The specialist’s function is then pushed toward commoditization or irrelevance.
The orchestrator can exert influence without building a substitute. This is the gatekeeping route. It can change default tool selection, alter ranking, deprecate a category, or extend a protocol in a way that favors some complementors over others. A vertical firm can remain technically available while becoming harder for users to find or less likely to be invoked. The connection still works; the economic opportunity attached to that connection has changed.
There is also a slower learning route. Where data-use terms permit, the input and output traces of customer interactions may be used to train or fine-tune the orchestrator’s model. That can expose the specialist’s behavior to the platform while giving the platform information with which to approximate it. The mechanism is not limited to formal model training. Even where enterprise terms prohibit using customer data for training, the orchestrator can observe which tools are invoked, where workflows fail, which outputs users accept or revise, and which vertical functions attract repeated demand.
Those observations reveal what the customer values and where the specialist is weak. Over many interactions, the orchestrator can improve planning, routing, evaluation, and task completion across domains. The vertical firm’s specialized work, by contrast, may be valuable only within one domain. The platform can learn from many categories of use while the complementor bears the cost of improving one.
An open protocol reduces one kind of friction: the friction of making systems connect. It does not constrain the orchestrator’s residual power to absorb, rerank, substitute, or learn. The interface is portable; the economic position may not be. A vertical firm becomes especially exposed when it optimizes its service for one orchestrator’s ranking rules, evaluation habits, permission model, or protocol extensions. Technical adaptation can then become economic dependence.
This is a systems problem because the relevant state is distributed across different owners. The vertical firm controls the domain capability and may control the evidence behind its result. The orchestrator controls the user-facing surface and the routing logic. The customer experiences a single task, but the choices that determine the task’s path are divided among them. If the customer cannot tell which party selected a tool, evaluated its answer, or retained the invocation history, accountability and value capture can drift toward the surface owner.
The exposure is not hypothetical merely because the protocol is open. Openness lowers the cost of entry for other complementors as well. A specialist can be added more easily, but so can a substitute. The same standardized connection that expands distribution also reduces the effort required to compare or replace providers. For a component business, that may be acceptable: reach and usage are the objective. For an accountability-bound business, it can be dangerous if the evidence and responsibility remain outside the surface where the decision is made.
Where durable value actually sits
The control question becomes clearer through the idea of complementary assets. An AI capability is not automatically a complete product. It needs other resources to become useful, trusted, and commercially durable. These resources differ in how dependent they are on the capability and how easily each side can find a substitute.
Generic complementary assets are widely available. Compute is the clearest example in this setting: a firm can rent it, and competitors can do the same. A generic asset may be necessary, but it does not by itself give a vertical AI firm a durable advantage.
Specialized complementary assets have a one-way dependence. Proprietary domain data may be valuable because it improves a capability or gives its owner an advantage. Yet, in principle, the capability might work with an alternative source of data. The dependence is meaningful but not necessarily binding.
Cospecialized assets have two-way dependence. The asset depends on the AI-enabled offering, and the offering depends on that asset, with no easy substitute on either side. This is where control becomes strategically consequential. In vertical AI, professional signoff, regulated workflows, evidence trails, and trusted systems of record can be cospecialized assets.
Consider why. An AI system can make professional review faster, but the review gives the result authority. A regulated workflow tells the organization what the output means and what must happen before it becomes an action. An evidence trail makes the conclusion defensible to a regulator, court, auditor, or internal reviewer. A trusted system of record supplies the authoritative state against which the output is reconciled and into which the decision is written back.
The AI capability and these assets reinforce each other. Without the AI, review or reconciliation may be slower and less scalable. Without the evidence, authority, and system of record, the AI output may be difficult to defend or use operationally. This mutual dependence is different from a generic function that can be copied, rented, or replaced with little effect on the surrounding process.
The distinction changes what “defensible technology” means. A model that produces a plausible answer can be technically impressive without owning the assets that determine whether the answer is acceptable. Conversely, a vertical firm may possess a less visibly novel model but control the evidence chain, review process, or regulated workflow that turns its result into something a customer can rely on.
Appropriability determines how much pressure exists to control those complementary assets. Appropriability is the difficulty of imitating the innovation itself. Patents, secrecy, tacit knowledge, proprietary fine-tuning data, tightly coupled evaluation sets, or accumulated deployment knowledge can make a capability difficult to reproduce. When that protection is strong, a firm may capture value even without owning the complete accountability stack because competitors cannot quickly provide the same capability.
When appropriability is weak, the balance changes. If a frontier model can approximate generic document reasoning, structured extraction, or first-pass summarization through ordinary prompting, the underlying AI function may offer limited protection. In that case, the complementary assets become the more durable source of value. A firm that gives an orchestrator control over the evidence, evaluation, or trusted workflow may be giving away the part that competitors cannot easily reproduce.
This is not a universal ranking of tasks. A capability that appears easy to imitate today may gain stronger protection through proprietary data, an evaluation system tightly coupled to deployment, or tacit knowledge accumulated through difficult cases. A capability that seems specialized may become easier to substitute as models improve. The point is to examine both sides: how copyable is the AI function, and how replaceable are the assets that make it legitimate?
The answer also depends on who has an incentive to invest. An orchestrator’s investments in planning quality, tool routing, and general reasoning can be reused across accounting, law, healthcare, and many other domains. The same improvement can serve a large user base. A vertical firm’s investments in evidence systems, professional review, escalation paths, and regulatory relationships are less transferable. An audit evidence infrastructure does not simply become a clinical governance system; a clinical escalation path does not automatically serve procurement.
This creates a substitutability gradient. Where the combined customer output depends mainly on general planning, the orchestrator is the natural investor and potential owner. Where it depends mainly on domain-specific accountability infrastructure, the vertical firm or another accountable domain owner has the stronger reason to invest and maintain the asset. If control is left with a party that does not capture the returns, no party may invest at the level the domain requires.
The three boundaries can now be distinguished more precisely. The interface boundary marks where the customer meets the system and can often move. The value-capture boundary marks where control over copyability and complementary assets permits durable returns to accumulate. The accountability boundary marks where responsibility, evidence, and governance must reside. They may overlap, but they do not have to. Moving the first boundary can change the second, while moving the third requires a real change in who can defend the result.
That is why an API launch can have consequences that are invisible in the architecture diagram. The vertical firm may still own the model call, but the orchestrator may own the user relationship, the evaluation surface, and the evidence of which tools were used. If those become the assets through which trust and demand accumulate, value capture can move before the vertical firm recognizes that it has changed its economic boundary.
The strategic response is not to treat every orchestrator as an adversary or every protocol as a trap. It is to decide which dependency is acceptable. A component firm may deliberately exchange a standalone interface for broad reach, provided it can remain reliable and defensible as a component. An accountability-bound firm may expose tools and bounded functions while retaining permissions, evidence, review, lineage, and signoff. In both cases, the firm should know which assets it is asking the orchestrator to control and which it cannot afford to lose.
The next test is empirical. If agents are becoming the new integrators, current adoption data should reveal more than a rising headline rate. It should show whether AI is spreading across functions, appearing in worker tasks, receiving organizational investment, and producing structural changes. Those layers will help distinguish a technology that is merely easy to invoke from one that is changing how firms organize responsibility.