Banking AI: Why you should command the air-traffic control tower

Three years ago, I watched the board of a large US bank approve a nine-figure investment to build their own foundation model.

At the time, it felt inevitable. If AI was going to redefine banking, surely the winners would own the models that powered it. Today, that decision looks less like a competitive advantage and more like an expensive lesson, not because the investment was poorly executed but because the market changed faster than the strategy.

The assumption that dominated banking’s first wave of AI — that owning the model meant owning the moat — has collapsed. According to Stanford University’s 2025 AI Index Report, the inference cost of GPT-3.5-level performance fell more than 280-fold between November 2022 and October 2024, while the performance of open-weight models approached that of closed models, with the performance difference dropping from 8% to 1.7% on key benchmarks in a single year.

Building a proprietary foundation model for most banking use cases now resembles building your own power station to keep the office lights on: technically impressive but strategically unnecessary and economically difficult to justify.

Some institutions are gaining ground because they built a smarter system: the air-traffic control tower that coordinates models, workflows, decisions and regulatory constraints across the enterprise. A control tower doesn’t fly the aircraft. It decides which aircraft flies, on what route, at what altitude and with what clearance.

Routing: The battleground is the agent execution layer

Value has migrated from models to runtimes. Organizations are now building orchestration layers that coordinate autonomous workflows across core banking, legacy systems and customer channels without replatforming decades of infrastructure.

In September 2025, Gartner® predicted: “Forty percent of enterprise applications will be integrated with task-specific AI agents by the end of 2026, up from less than 5% today.”*

The aircraft are arriving whether the tower is staffed or not.

In the shift from conversation to consequence, banks are moving beyond chatbots and deploying agentic systems that execute multistep processes and inherit the access rights of the people they work alongside.

In practice, that means real-time fraud interdiction that freezes, verifies and resolves in a single journey; straight-through commercial loan routing across more than 140 discrete lending capabilities mapped in NTT DATA’s lending value-stream model, which is aligned to the Banking Industry Architecture Network (BIAN) industry standard; and proactive wealth rebalancing that executes, not merely recommends. The difference lies in whether the system can complete a transaction end to end without human intervention.

Radar: One picture, or no picture at all

A control tower with three incompatible radar screens is not a control tower. Financial services providers need to establish a unified semantic fabric: a governed, enterprise-wide view of customer, product and risk data that every intelligent workflow draws from. One source of truth, with many coordinated workflows.

Without that coherence, banks produce a sprawl of disconnected autonomous systems that is costlier and harder to govern than the point-solution fragmentation that preceded it.

Gartner predicts: “By 2030, 50% of AI agent deployment failures will be due to insufficient AI governance platform runtime enforcement for capabilities and multisystem interoperability.”**

If your intelligent workflows cannot see across value streams, you have built nothing but sophisticated silos.

Clearance: Hyperscaler dependence is now a board-level risk

Concentrating AI operations within a single public cloud now presents a systemic operational risk that regulators are already scrutinizing. The European Supervisory Authorities have designated major hyperscalers as “critical ICT third-party providers” under the European Union’s Digital Operational Resilience Act (DORA), while Germany’s BaFin has confirmed that GenAI must be governed within existing information and communications technology risk, testing and third-party frameworks. Exit strategies for high-concentration functions are now expected to be documented and tested.

Banks are building dynamic routing into their orchestration architectures from the outset, directing workloads to the models best suited for each task while avoiding dependence on any single provider.

When a provider experiences latency degradation or raises prices, a well-designed orchestration layer moves workloads to an alternative provider or to private, on-premises infrastructure without disrupting a single integration. For sovereign and regulated workloads, that private fallback isn’t optional. Resilience depends on portability.

Fuel: Token economics is the new FinOps

The era of seat-based software licensing is giving way to usage-based AI economics, exposing which programs were engineered to generate returns rather than headlines. Falling unit prices are not the same as falling bills.

Gartner predicts: “By 2030, performing inference on a large language model (LLM) with one trillion parameters will cost GenAI providers over 90% less than it did in 2025.”***

As AI token consumption rises faster than token costs decline, total inference spend is expected to increase. Cheap fuel does not make an airline profitable. Fuel discipline does.

The metric that increasingly matters to finance leadership is cost per transaction completed, not pilot satisfaction scores or demonstration velocity. Institutions now manage compute consumption with strict workflow budgets, aggressive caching, context compression and unit-cost optimization.

This is where banks can manage AI against measurable economic returns, priced per unit of value delivered rather than per seat occupied. If you cannot articulate the fully loaded cost of a single AI-completed credit decision, you have an expense account, not an AI strategy.

Choosing architecture over algorithms

The gap between institutions scaling AI with measurable operational impact and the majority still cycling through pilots is now well documented, and it’s not a technology gap. The Cambridge Centre for Alternative Finance’s 2026 Global AI in Financial Services Report found that 52% of financial institutions are already adopting agentic AI, yet only 23% have reached the scaling or transforming stages while 29% remain in piloting.

The same report found that although 81% of financial institutions have adopted AI at some level, only 14% see AI as transformational to their organizational strategy and competitive advantage. Most large institutions can access broadly comparable foundation models. Barely 1 in 7 — the 14% mentioned above — have turned that access into operating leverage. The differentiator is architectural discipline.

When you embed governance into the execution layer, governance controls can be applied at the point of execution. In banking, where a single governance failure can erase years of efficiency gains, that matters greatly.

Equally important is value-stream alignment. Deploying intelligent workflows against clearly mapped banking capabilities, rather than an ad hoc wish list of use cases, is what converts isolated experiments into durable operating leverage. Strategy drives the deployment roadmap, architecture makes it defensible and governance makes it sustainable.

The control tower is the competitive asset

Stop trying to own the factory. Foundation-model capability is rapidly commoditizing. Heavy investment in proprietary models does not necessarily translate into faster execution.

The durable competitive advantage stems from the orchestration system that coordinates models, governance and economics at enterprise scale. This system is built on hyperscaler-agnostic routing, workflow-level unit economics, a single semantic picture of the enterprise, and governance embedded throughout.

When you master this, your organization becomes far more adaptable: capable of onboarding new skills, entering new markets and responding to regulatory change faster than your competitors. That adaptability is what transforms technology investment into market position.

The model wars are over. The orchestration era has begun.

What to do next

NTT DATA helps financial institutions build the architecture, governance and operating model needed to scale AI confidently. Whether you’re moving beyond pilots or redefining your enterprise AI strategy, connect with us to discuss how we can build your AI control tower.