You Don’t Need a Better Model, You Need AI Control Plane

Large language models are rapidly becoming commoditized. Which makes it odd that so much of the conversation around AI in capital markets is still about the AI itself- the model selection, the prompt engineering, and the capabilities. As though the bottleneck is intelligence rather than infrastructure.

It isn’t.

For most firms, the real constraint is structural, and it sits well below the model layer. As firms move from pilots to production, competitive advantage will depend less on choosing the best model and more on building the right control plane.

There are two distinct problems that organizations often conflate.

• The data plane: the fragmented network of trading, clearing, settlement, finance, and operational systems where enterprise data resides.
• The control plane: the architectural layer that governs how AI agents access data, execute actions, and remain transparent, auditable, and secure.

Most firms have invested in the first. Very few have built the second. That gap is where the real work begins.

Why AI pilots fail in production

The 2023 ION Trading ransomware attack exposed more than an operational resilience issue. It revealed how dependent many financial institutions had become on data flows, reconciliation processes, and operational dependencies they hadn’t fully mapped. What appeared to be a resilience failure was, in many respects, an architectural one.

The same pattern appears in AI initiatives.

A pilot is, by design, a controlled environment. Clean data, defined scope, sympathetic conditions. The complexity doesn’t surface until you point an agent at production: twenty fragmented legacy systems that were never designed to talk to each other, inconsistent naming conventions, duplicated fields, undocumented relationships, schemas last updated when someone left in 2014. A text-to-SQL approach looks promising for about five minutes. Then you see what it returns.

AI works with whatever you give it, with equal confidence, regardless of quality. It won’t flag that the data is messy. It’ll pattern-match on the mess and present the results as though everything were pristine: the model doesn’t know the difference between clean data and a decade of accumulated inconsistency. Pilots don’t surface this problem because pilots don’t encounter it. Live deployment does.

The four pillars of an AI control plane

A mature AI control plane should provide four foundational capabilities. Without them, organizations may improve operational efficiency but struggle to demonstrate governance, accountability, and regulatory defensibility.

Data quality

Not as a one-time migration exercise, but as an ongoing capability. Continuous ingestion, normalization, validation, and reconciliation transform fragmented enterprise data into information that AI systems can safely use.

Visible logic

Every recommendation and action should be traceable. Organizations need to understand which data informed a decision, how conclusions were reached, and what actions an AI agent ultimately performed.

Authorization

AI agents should operate within the same permission boundaries as human employees. Access without governance is simply another form of operational risk.

Auditability

Every interaction should create a durable execution record detailing what occurred, when it occurred, under whose authority it happened, and which systems were affected.

None of these capabilities are particularly novel. They represent disciplined enterprise architecture applied to an area that has historically received less strategic attention: operational data, reconciliation, validation, and process orchestration.

Why AI infrastructure matters more than model selection

The reward for building these strong foundations is disproportionate. Once the control plane exists, the model becomes a plug-in.

You swap it, upgrade it, run three in parallel… It doesn’t matter because the control surface is consistent. The strategic decision for CIOs isn’t which AI model to adopt. It’s whether the infrastructure underneath your operations is mature enough to let any model, current or future, operate safely.

The capability ceiling isn’t the AI. It’s what’s underneath it. A mature data platform benefits every agent you deploy, because the hard problems – quality, governance, auditability – have been solved at the platform level rather than hacked around at the application level.

Getting it right

The path forward requires being clear about sequencing. Map the gap between your data plane and anything resembling a control surface. Identify one operational workflow – reconciliation, exception management, cash position reporting – where the data quality and permissioning problems are well understood.

Build the control plane for that workflow first. Get the four elements in place: quality, visibility, authorization, and audit. Then deploy a robust, governed agent into that environment and see what it can do when the infrastructure underneath it is actually ready.

The model isn’t the decision. The decision is whether to start building the infrastructure that will make every model you ever use perform better. The best time to start was last year, when the conversation began. The second best time is now.

In brief

The industry has spent the past two years debating models. The next decade will be defined by infrastructure.

The organizations that realize the greatest value from agentic AI will not necessarily be those adopting the newest models first. They will be those investing in the architectural foundations that allow every future model to operate securely, transparently, and at scale.

Models will continue to evolve. Architectures endure.

For technology leaders, the strategic decision is no longer which model to deploy. It is whether they are building the control plane that will enable every model that follows.