The instinct that drives enterprise AI investment toward new capabilities is understandable. New capabilities are visible. They generate executive attention, analyst commentary, and competitive narrative. A new AI-powered interface, a generative assistant for knowledge workers, an agentic workflow that automates a previously manual process, each of these produces a story that is easy to tell and easy to measure at launch.
The instinct is also misdirecting investment away from the highest-return AI opportunities in most large enterprises, which are not visible in the same way because they are embedded in operational systems that have been running quietly at scale for years.
The scale argument
The most important difference between decision optimisation within existing operational systems and new AI capabilities at the edge of operations is not sophistication or technical complexity. It is scale.
A large bank’s payment authorisation system makes several billion decisions per year. Each decision is a binary outcome: approve or decline. Each outcome has a financial consequence: approved revenue retained, legitimate transaction protected from fraud, or fraud loss incurred, or legitimate customer declined. The aggregate financial consequence of those billions of decisions, and the gap between their current quality and their achievable quality, is a very large number.
A new AI capability deployed to a workforce of ten thousand knowledge workers will influence some number of decisions made by those workers. The total volume of decisions influenced is bounded by the workforce size and the frequency with which individual workers encounter the relevant decision type. The impact surface is fundamentally different in scale from the impact surface of an improvement to the authorisation model.
The scale asymmetry does not mean new capabilities are not valuable. It means they are not comparable to decision optimisation in existing high-volume systems, and treating them as equivalent investment alternatives, as most AI portfolio planning processes implicitly do, leads to systematic underinvestment in the higher-return opportunity. The underlying mechanism is decision density: the product of decision volume, unit financial consequence, and decision frequency. Operational systems running on IBM Z and equivalent core infrastructure operate at decision densities that no workforce-facing AI capability can match. That density is where AI investment produces the highest leverage.
The investment profile difference
The investment required to generate AI value from an existing operational decision frequently produces faster and more measurable returns than the investment required to build a new AI capability, for reasons that are consistent across industries and decision types.
Decision optimisation in an existing system starts with a working production system, an established data pipeline, a known decision type, a measurable outcome, and an existing integration between the AI and the operational process. The incremental investment to improve the model is concentrated in training data refresh, model development, validation, and deployment to a system that is already configured to consume the output. The time from investment decision to production value is measured in weeks to months. The same logic applies whether the decision is a payment fraud score, a logistics network routing allocation, a manufacturing yield classification, or a telco network fault prioritisation. The decision path, outcome metric, and economic consequence already exist and are measurable.
New AI capability development starts from a lower base. The use case must be defined precisely enough to specify a model. The data pipeline must be built or adapted. The integration must be designed and implemented. The organisational change management required for adoption must be planned and executed. The time from investment decision to production value at meaningful scale is measured in months to years, and the realised adoption rate at the end of that period is frequently lower than the business case projected.
Existing operational systems are not necessarily simpler to improve technically. Many are deeply integrated, operationally sensitive, and carry significant feature debt and governance coupling. The advantage is not that the technical work is easier. It is that the economic case is clearer: the decision volume, the outcome metric, and the financial consequence of a quality improvement are all measurable before the investment begins, and progress is visible against a known baseline.
The production risk profile is also different. An improvement to an existing operational model carries the risk that the improved model performs differently than expected, which is detectable quickly in production against a known baseline. A new AI capability carries the risk that the use case was not correctly defined, that adoption does not reach the level required to generate the projected value, or that the integration with downstream processes creates friction that the business case did not anticipate. These are larger and harder-to-detect risks.
What the portfolio resequencing looks like
The practical implication of the scale and investment profile argument is a resequencing of the AI portfolio that most enterprises have not yet made.
The first tranche of AI investment should be directed at the highest-volume operational decisions already running in core systems, ordered by the product of decision volume, unit financial consequence, and current quality gap. For a bank, that typically means fraud scoring, credit adjudication, and collections prioritisation before it means knowledge worker productivity tools. For an insurer, it means claims routing, underwriting accuracy, and fraud referral before it means agent assist interfaces.
The second tranche should be directed at net-new capabilities where the use case is clearly defined, the adoption path is credible, and the investment case does not depend on achieving adoption rates that the organisation’s change management capacity cannot realistically deliver.
Most enterprise AI portfolios are sequenced in the opposite order, not because the high-volume operational opportunity is less valuable, but because it is less visible. It does not generate a product launch announcement. It does not produce a demo that captures executive attention in a strategy presentation. It produces a change in a model that influences billions of decisions per year and shows up as a two-point improvement in the fraud detection rate and a measurable reduction in false decline revenue loss.
That outcome is less visible and more valuable. The organisations that have learned to prioritise it are the ones generating AI ROI that compounds year over year rather than cycling through successive rounds of pilot investment in search of the capability that finally delivers.
