Enterprise AI Needs Control
Most enterprise AI strategies, stripped of their slides, are procurement plans. Copilot, vector databases, agent plaforms, and more. The spend is real and the activity is genuine, but the architecture is accidental. Two years in, leadership is starting to notice that the sum of these investments doesn't behave like a capability. It behaves like a portfolio of pilots.
The instinct is to blame the models, the data, or the change management. The more useful diagnosis is structural. Enterprise AI has three planes, data, model, and agent, and almost no organisation has built the fourth: a control plane that makes the other three accountable to the business.
The data plane is the substrate. It's what AI sees, remembers, and grounds itself in: warehouses, lakes, vector stores, document corpora, telemetry, the long tail of operational systems. Most enterprises have spent a decade investing here and still describe the result as "messy." That's because the requirement has changed. Data infrastructure designed for analytics is typically clean, periodic, and broadly aggregated; however, it differs fundamentally from infrastructure built to serve models in real time, which requires fine-grained data provenance, dynamic permissioning, and strict freshness guarantees at the point of query.
The model plane is the intelligence. Frontier models, fine-tunes, embeddings, classifiers, the mix of hosted and self-hosted endpoints that any advanced organisation now runs in parallel. The interesting question on this plane is no longer "which model is best" but "which model, for which task, under which constraints, at which cost." Answering that requires more than a leaderboard. It requires routing, evaluation, and the discipline to change your mind as the frontier moves - which it does, weekly.
The agent plane is the action layer. This is where the system stops describing the world and starts changing it: writing to systems of record, orchestrating tools, executing multi-step workflows on behalf of a human or another agent. The agent plane is the newest of the three, the least understood, and the one most likely to be deployed without the safeguards the other two accumulated over decades.
Each plane is necessary. None is sufficient. And the gap between them (between data that can be trusted, models that can be chosen, and agents that can be trusted to act) is where most enterprise AI initiatives quietly fail.
Borrow the metaphor from networking and cloud, where it has done useful work for thirty years. The data plane carries traffic. The control plane decides how. One does the work; the other makes the work coherent.
An enterprise AI control plane is the layer that sits across data, model, and agent and answers the questions a business actually has. What is this system allowed to do? How do we know it's working? What did it cost? Who is accountable when it's wrong? How quickly can we change it? It is the surface where governance, observability, evaluation, policy, routing, and cost management live as first-class concerns rather than slideware.
This is not a product category yet, which is part of the problem. Most vendors sell into one plane and gesture at the others. The result is that the control plane gets assembled by accident, in tickets, by the team that happens to be on call. That works at pilot scale. It does not survive contact with a regulated production environment.
The organisations getting compounding value from AI are the ones that have started treating the control plane as the actual product. Their model choices are reversible. Their agents are observable. Their data flows are policy-aware. They can answer "is this working?" with evidence rather than anecdote, and they can change that answer next quarter without rebuilding from scratch. That optionality, the ability to swap a model, retire an agent, or tighten a policy without breaking the surrounding business process, is what separates an AI capability from an AI portfolio.
A well-built control plane doesn't just keep the lights on. It is the only place from which the whole system can actually be optimised, because it is the only layer that sees the whole system.
A model swap is meaningful only if you can measure the downstream effect on cost, latency, and quality across every workflow that uses it. An agent's permissions are meaningful only if you can trace what it touched and why. A data source's freshness is meaningful only if you can tie it to the decisions it informed. Each of these is a control-plane function. Without it, optimisation collapses into local heuristics: each team picks the model they like, each agent gets the permissions someone signed off on in a meeting, each data pipeline is tuned to the SLA of whoever shouted loudest. The system gets bigger but not better.
With a real control plane, optimisation becomes a business discipline rather than a technical one. The organisation can ask (and answer) questions like which workflows are worth running on a frontier model and which aren't, where are agents adding measurable value and where are they adding measurable risk, what is the marginal cost of one more nine of reliability on this process. Those are the questions that move P&L.
The shift is from sponsoring use cases to architecting a layer. Use cases will come and go; some will scale, most will not. The layer either compounds value across all of them or it doesn't. Leaders who understand this stop asking "what can AI do for function X" and start asking a different set of questions:
None of these are AI questions. They are operating-model questions, dressed in new clothes. The organisations that answer them well will spend the next five years compounding. The ones that don't will spend the next five years renewing pilots.
The model plane will keep moving. The data plane will keep being hard. The agent plane will keep getting more capable, faster than anyone is comfortable with. The control plane is the part you actually build, and it is the part that turns enterprise AI from a line item into a layer.
To find out more contact us
Uncover fresh perspectives with our handpicked blog posts on AI advancements.
