Implementation costs – the hidden economics of AI adoption

There is a point in any system under pressure at which what appears to be working begins to separate from what can actually be trusted. It is not dramatic. There is no visible failure, no alarm, no obvious breakdown.

Outputs continue to flow, timelines are met, and confidence builds. Decisions are made more quickly than before, often with less friction, and the system appears to be doing exactly what it was designed to do. It is only later, usually when a decision is challenged, that the underlying structure is tested, and that is where the economics change.

In operational environments where outcomes carry consequence, cost has never been defined solely by what it takes to produce an answer. It is defined by whether that answer can be defended. This distinction is not theoretical.

In aviation, when an incident occurs, investigators do not ask whether the aircraft was efficient. They reconstruct the sequence of decisions that led to the outcome. The system is designed so that each action can be traced, each input accounted for, and each decision understood within the context in which it was made. In search and rescue operations, the same principle applies. Communications are logged, transcribed, and preserved, not because it improves speed, but because it protects the integrity of the decision-making process. When a plan is executed, it is anchored to something that can be examined later. The system assumes, from the outset, that every decision may need to be justified.

Artificial intelligence, as it is currently deployed across most industries, does not operate on this assumption. It produces outputs, often with impressive accuracy, but the pathway that leads to those outputs is not inherently preserved in a form that satisfies operational or legal scrutiny. The system can generate an answer, but it cannot always demonstrate how that answer was constructed in a way that can be independently verified. This introduces a cost that does not appear in implementation plans, and it accumulates quietly.

The prevailing economic narrative around AI adoption has been shaped by efficiency. Systems that can automate workflows, reduce manual effort, and increase throughput are positioned as cost-saving mechanisms. This is supported by widely cited projections. A 2023 report by McKinsey & Company estimates that generative AI could contribute between $2.6 trillion and $4.4 trillion annually to the global economy through productivity improvements across industries.[1]

These projections are not incorrect, but they are incomplete. They assume that the outputs generated by these systems can be relied upon without introducing disproportionate downstream costs, and they assume that correctness, measured statistically, is sufficient. In many domains, it is not.

The cost structure begins to shift when decisions are no longer evaluated solely on whether they are correct, but on whether they can be explained and defended. In healthcare, this distinction is already visible. Clinical decisions must be traceable. Regulatory frameworks require that outcomes can be linked to reasoning that is both transparent and reproducible. A 2023 industry survey by Black Book Research reported that 76% of healthcare organizations are unable to fully reconstruct AI-assisted decision pathways, raising concerns about audit readiness and compliance exposure.[2] The implication is not simply that systems are incomplete; it is that their economic value is conditional on capabilities they do not yet possess. A system that cannot explain its decisions introduces a secondary cost layer. It may operate efficiently in the moment, but it creates exposure that must be managed elsewhere.

A similar dynamic is emerging in financial services. AI systems are increasingly used in credit adjudication, fraud detection, and risk assessment. These applications benefit from speed and scale, but they also operate within regulatory environments that require fairness, transparency, and accountability. When a credit decision is challenged, the institution must demonstrate how that decision was made. If the system cannot provide a clear and defensible explanation, the cost is no longer technical; it becomes legal, regulatory, and reputational. This is not a hypothetical risk. Regulatory bodies globally are moving to address it. The European Union Artificial Intelligence Act, formally adopted in 2024, establishes requirements for high-risk AI systems, including obligations around transparency, human oversight, and the ability to provide meaningful explanations of automated decisions.[3] These requirements do not add new functionality to AI systems, they expose what is missing.

The economic model of AI adoption is often presented as a combination of infrastructure costs, integration costs, and operational savings. This model captures the visible components of deployment, but it does not account for the cost of maintaining trust in the system’s outputs. A more complete representation can be expressed as

C_total = C_infrastructure + C_integration + C_verification + C_liability, where

C_infrastructure includes compute, storage, and licensing,
C_integration includes engineering, workflow adaptation, and training,
C_verification includes the cost of ensuring outputs can be reconstructed and justified, and
C_liability includes the cost incurred when decisions cannot be defended.

The first two components are typically addressed during implementation, while the latter two are often deferred. They are not eliminated. They accumulate.

Verification cost is not simply a technical problem. It is an operational requirement. It includes the systems, processes, and controls needed to ensure that decisions can be traced back to their origin. This may involve maintaining detailed logs, preserving intermediate states, and establishing frameworks for interpreting model behaviour.

When these capabilities are not built into the system from the outset, they must be added later, often at significantly higher cost.

Liability cost is even more complex. It is not incurred during normal operation. It appears when something goes wrong, or when a decision is challenged. It includes the cost of investigation, remediation, legal defence, regulatory penalties, and reputational damage. These costs are highly asymmetric. A system may produce thousands of correct outputs with minimal issue, but a single incorrect decision in a high-consequence context can outweigh those gains. Traditional performance metrics do not capture this asymmetry.

AI systems are typically evaluated using aggregate measures such as accuracy, precision, and recall. These metrics provide a statistical view of performance, but they do not reflect the distribution of errors or their impact. In many cases, errors are not evenly distributed. They occur in edge cases, where the system encounters conditions that are underrepresented in training data or poorly defined in context. These are precisely the scenarios where decisions are most critical. The cost of an error in these situations is not proportional to its frequency; it is proportional to its consequence.

This creates a tension between efficiency and admissibility. Efficiency focuses on throughput and scale, valuing speed and volume. Admissibility focuses on whether a decision can be justified under scrutiny, valuing traceability, transparency, and constraint. These are not opposing goals, but they operate under different assumptions. Efficiency assumes that most outputs will be acceptable, while admissibility assumes that any output may be challenged. When systems are designed primarily for efficiency, admissibility becomes an afterthought, and the cost of addressing it later is often higher than the cost of building it in from the beginning.

Some organizations are beginning to adjust their approach. Rather than treating AI as a stand-alone capability, they are integrating it into governance frameworks that prioritize accountability. This includes defining clear boundaries for where AI can be used, establishing protocols for human oversight, and ensuring that decision pathways are recorded and auditable. This approach does not eliminate cost, but it redistributes it. Investment shifts from reactive correction to proactive control.

There is also a cultural dimension to this transition. AI systems often present outputs with a level of confidence that can influence how they are perceived. When a system produces an answer quickly and fluently, there is a tendency to accept it as authoritative. This can reduce the level of scrutiny applied to individual decisions, particularly in high-volume environments. Maintaining appropriate skepticism requires discipline. It requires recognizing that confidence is not the same as correctness, and that correctness is not the same as admissibility. Organizations that understand this distinction are better positioned to manage the hidden costs of AI adoption.

The question is not whether AI can deliver economic value. It can. The question is whether that value can be realized without introducing disproportionate risk. This depends on how systems are designed, how decisions are governed, and whether the underlying assumptions about cost are aligned with the realities of consequence. The visible costs of AI are easy to measure. The hidden costs are not. They only become visible when the system is asked to explain itself.

References

[1] McKinsey & Company, The Economic Potential of Generative AI: The Next Productivity Frontier (2023).

[2] Black Book Research, AI Adoption and Readiness in Healthcare (2023).

[3] European Union, Artificial Intelligence Act (2024).

(Mark Jennings-Bates – BIG Media Ltd., 2026)