When Efficiency Becomes an Expensive Illusion

When Efficiency Becomes an Expensive Illusion

One of the most attractive promises of artificial intelligence is that it can help businesses do more with fewer resources. Documents are generated faster, customer enquiries are processed automatically, large datasets are examined in seconds and administrative activities that once occupied entire afternoons can be completed before someone has finished their morning coffee. An extraordinary achievement, certainly. But before declaring another revolution in productivity, I would like to examine something considerably less glamorous. What exactly has become more efficient, and who benefits from the improvement? Because reducing the time required to perform a task is not necessarily the same as increasing the economic value created by that task. We may have made the operation faster without making it more useful, more accurate or more profitable. And if the activity itself contributes very little to the business, accelerating it may simply give us the pleasure of accomplishing something unnecessary at remarkable speed.

The economics of automation deserve more serious examination than the number of hours a software provider promises to save. Imagine a small company introducing an AI system that eliminates sixty hours of routine administrative work each month. At an internal labour cost of forty euros per hour, the apparent saving amounts to 2,400 euros. Excellent figures for a presentation about digital transformation. But suppose the company also pays 350 euros per month for the technology, spends twenty five hours checking automated outputs and another fifteen hours correcting errors and managing exceptions. At the same assumed labour cost, verification and correction account for another 1,600 euros. The original saving of 2,400 euros has now become a net operational benefit of 450 euros per month, before accounting for implementation, employee training, system maintenance or possible changes in customer satisfaction. These are illustrative figures, but the economic principle is anything but hypothetical. A business cannot measure the value of automation solely by the labour it appears to eliminate while ignoring the resources required to make the automated process reliable.

And even these calculations require interpretation. Time saved does not automatically become money earned. If an employee spends fewer hours on administration, the organisation may gain additional capacity without reducing its actual expenditure. That capacity has potential economic value, but the benefit depends on what happens next. Can those hours be redirected towards activities that generate revenue, improve service quality or reduce genuine operating costs? Or have we simply created more available time without changing the economic performance of the organisation? I would also want to know whether the automation has introduced new dependencies, additional software expenses or operational risks that were absent from the original process. A sensible investment decision requires an assessment of total costs, expected benefits and the uncertainty surrounding both. The subscription fee is merely the most visible part of the calculation.

I find it particularly revealing when companies celebrate the number of automated tasks without establishing whether the business actually needed those tasks in the first place. We have reached a point where an inefficient process can be automated, monitored through an attractive dashboard and proudly presented as evidence of technological progress. But if the process was poorly designed, why are we investing in making it run faster instead of questioning its existence? Perhaps the most intelligent decision would have been to remove the unnecessary activity altogether. Admittedly, eliminating pointless work does not provide quite the same marketing opportunities as announcing a sophisticated AI transformation.

Executing a Task Is Not the Same as Exercising Judgment

The distinction between execution and judgment becomes particularly important when AI systems move beyond repetitive operations and begin influencing management decisions. A computational system can analyse available data, identify statistical relationships, compare alternatives and generate recommendations according to its design and instructions. Some systems can also operate tools, coordinate sequences of actions and adapt their responses to changing information. These capabilities can be extremely useful. But none of them automatically establishes that the system is applying the right criteria to the right business problem. Before evaluating the quality of a recommendation, we need to understand how the problem was defined, which information was considered relevant and what the proposed decision is actually intended to achieve.

Consider a company using an automated pricing system to increase sales. The model identifies opportunities to adjust prices, customer response improves and transaction volumes begin to rise. Management is delighted. The reports show increased conversion, the sales department celebrates and the technology appears to be delivering exactly what was promised. Now imagine that the original objective was defined primarily in terms of transaction volume, without sufficient attention to contribution margins. The system may recommend discounts that successfully attract additional orders while reducing the economic contribution of each sale. Is this a failure of artificial intelligence? Not necessarily. The system may be optimising the objective it was given. The more uncomfortable possibility is that the organisation selected an inadequate objective and then invested in technology capable of pursuing it with extraordinary efficiency.

Let us make the problem concrete. Suppose a business initially sells 1,000 units at 100 euros each, with variable costs of 75 euros per unit. Revenue amounts to 100,000 euros, while the total contribution towards fixed costs and profit is 25,000 euros. An automated pricing strategy then reduces the selling price to 90 euros and increases sales to 1,250 units, with the same variable cost per unit. Revenue rises to 112,500 euros, an increase of 12.5%. Transaction volume grows by 25%. On the dashboard, both indicators look rather impressive. But the contribution per unit has fallen from 25 euros to 15 euros, leaving a total contribution of only 18,750 euros. The company now generates more revenue and completes more transactions while producing 25% less contribution towards fixed costs and profit. The calculations assume unchanged unit variable costs and exclude additional effects such as returns or customer acquisition expenses. Even under these simplified conditions, the contradiction is difficult to ignore. What exactly did the company optimise? And who decided that increasing transactions was more important than maintaining economic contribution?

This is where managerial judgment becomes indispensable. The economic objective of a business is rarely reducible to a single indicator. Pricing decisions may affect profitability, customer retention, market positioning, operational capacity and competitive responses. These considerations do not always point towards the same solution. Increasing volume might be perfectly reasonable when a company has unused capacity and sufficient contribution margins. It might be disastrous when production costs are rising and operational resources are already stretched. The appropriate decision depends on context, priorities and constraints. Technology can help us analyse these relationships, but the responsibility for defining acceptable outcomes remains a management function.

I am especially sceptical of the assumption that a technically sophisticated recommendation deserves greater authority merely because it comes from a computational system. A mathematical calculation may be correct while the management decision built around it is economically unsound. A forecast may be internally consistent while relying on assumptions that do not reflect the market. A recommendation may be perfectly aligned with a specified objective, even when that objective contradicts the interests of the enterprise. If the problem has been defined incorrectly, computational precision does not rescue the strategy. It can make the consequences more systematic.

Business Context Cannot Be Reduced to a Convenient Dataset

There is another aspect of AI adoption that deserves considerably more attention than it usually receives. Business decisions are made within environments shaped by formal rules, contractual obligations, customer expectations, organisational relationships and information that may never have been systematically recorded. A database can contain sales histories, product prices, customer interactions and operational metrics. It may not contain the reason a particular client received exceptional payment terms, the commercial implications of a delicate negotiation or the fact that an apparently profitable product is being discontinued because its supplier can no longer guarantee delivery. These circumstances can sometimes be documented and made accessible to AI systems, but they should not be assumed to exist in the data simply because the system has access to a substantial amount of information.

I would be particularly careful with the distinction between information availability and contextual relevance. A system may have access to thousands of records and still lack the particular information needed to make a reliable recommendation. It may identify patterns that are statistically meaningful but commercially misleading. Historical customer behaviour, for example, can provide useful evidence about demand, but it does not necessarily reveal how customers will respond to a new competitor, a regulatory change or a disruption in purchasing power. Nor does a confident explanation guarantee that the underlying evidence is complete. The practical question is not how much information the system can process, but whether the information is suitable for the decision being considered.

This is not an argument that AI systems are incapable of contextual analysis. They can combine multiple information sources, recognise inconsistencies, identify missing details and help decision makers examine alternatives. Under appropriate conditions, these capabilities can improve the quality and efficiency of business analysis. But the relevant context must be available, sufficiently reliable and interpreted with appropriate attention to uncertainty. Where information is incomplete, the system should not be treated as an authority merely because it can produce a coherent response. A convincing explanation can make an unsupported assumption appear surprisingly respectable.

I also find the language surrounding business intelligence rather revealing. We talk about intelligent systems, autonomous agents and automated decision making as though these expressions settle the question of competence. They do not. A system's ability to generate a plausible recommendation is evidence of a capability, not proof that the recommendation is suitable for a particular commercial situation. The distinction is especially important when decisions affect contracts, financial commitments, customer rights or employees. In these circumstances, knowing what the system has produced is not enough. We need to know which information supports the result, which assumptions remain uncertain and whether the decision can be challenged before it creates consequences that are difficult to reverse.

If an AI system recommends a commercially attractive course of action that violates an existing contractual commitment, the problem cannot be solved by admiring the sophistication of the analysis. Someone must establish which obligations take priority, what risks are acceptable and whether the recommendation can be implemented at all. These are questions of governance, organisational knowledge and accountability. They cannot be answered simply by increasing the amount of text the system is allowed to generate.

The machine may have access to the data. That does not mean the data contain everything the business needs to know.

What If AI Simply Allows an Existing Mistake to Spread Faster?

Perhaps the most underestimated risk of automation is not that technology will suddenly introduce an entirely new category of business failure. It is that existing weaknesses in management, data quality and organisational processes can be reproduced at a speed and scale that manual operations would never have allowed. A poorly designed procedure may once have affected a few customer transactions before an employee noticed something was wrong. Automate the same procedure, connect it to several business systems and allow it to operate continuously, and the original weakness can become an organisation wide problem before anyone has examined the first set of consequences.

Imagine a company with an unclear refund policy. Employees have traditionally interpreted exceptional situations individually, sometimes inconsistently, but the problems remain relatively limited because every case requires human attention. Management decides to automate customer requests using an AI system connected to the order database and communication platform. The system is instructed to follow the company's existing policy and resolve enquiries efficiently. Unfortunately, the policy contains ambiguous conditions that nobody thought important enough to clarify before implementation. The automated workflow begins interpreting those conditions in ways that differ from the company's intended commercial practice. Customers receive inconsistent decisions, some cases are escalated unnecessarily and others are resolved under terms the business did not mean to offer. The technology has not invented the ambiguity. It has simply made the consequences of that ambiguity considerably more visible and potentially more expensive.

Now consider what happens when the same logic is applied to larger operational systems. Incorrect product classifications can distort automated recommendations. Outdated customer information can influence marketing decisions. Poorly defined approval rules can create financial commitments without adequate review. Faulty assumptions about demand can affect inventory planning and purchasing. In every case, the central concern is not merely whether the software performs its assigned operation. It is whether the underlying process, the information it depends on and the controls surrounding it are sufficiently reliable for automated execution.

Scale changes the economics of error. If a manual process produces one incorrect decision out of every hundred cases, the organisation may have time to recognise and correct the problem before it expands. An automated process operating across thousands of cases can generate a much larger number of problematic outcomes before the pattern becomes apparent. The frequency, severity and recoverability of errors all matter. Some mistakes are inexpensive and easily reversed. Others can damage customer relationships, trigger contractual disputes or create obligations that cannot simply be cancelled with a software update. The cost of a failure therefore depends not only on whether an error occurs, but also on how widely it propagates and how quickly the organisation can intervene.

This is why I would examine the quality of the business process before celebrating its automation. What assumptions does the process rely on? Where are decisions made? Which exceptions require special handling? What information must be verified? Which outcomes could create unacceptable financial or legal consequences? And perhaps the most revealing question of all: if we cannot explain how the process works today, what makes us so confident that automating it tomorrow will improve the business?

I have no difficulty imagining a company proudly announcing that it has automated 90% of its operations while quietly discovering that nobody fully understands the remaining 10%, where the most consequential decisions happen to reside. Impressive technological transformation, certainly. But I would be considerably more interested in whether the company has become easier to manage, more reliable and economically stronger. Otherwise, we may have achieved nothing more than an exceptionally efficient distribution system for mistakes that were already there.

Human Oversight Is Not a Signature at the End of the Process

The phrase human oversight appears frequently in discussions of responsible AI adoption. It sounds reassuring, and in principle it addresses a genuine need. Yet I would be reluctant to accept its presence in a company policy as evidence that meaningful supervision actually exists. There is a substantial difference between having a human somewhere in the workflow and designing a system in which that person can understand, challenge and, when necessary, prevent an automated action. A manager who receives hundreds of recommendations and approves them without examining their basis is not necessarily exercising judgment. The organisation has simply introduced an additional click between the algorithm and its consequences.

Meaningful oversight requires operational authority. The person responsible for reviewing an automated decision must have access to relevant information, understand the purpose of the system and be able to recognise circumstances in which its output should not be trusted. There must be a practical mechanism for escalating uncertain cases, stopping inappropriate actions and correcting mistakes. The system should maintain records that allow significant decisions to be examined after the event. These controls become particularly important when AI is connected to tools capable of changing prices, issuing communications, modifying customer records or initiating transactions. At that point, the distinction between generating a recommendation and executing an action is no longer theoretical. It determines whether a questionable output remains a suggestion or becomes an operational consequence.

I would also question the belief that human involvement automatically makes an automated process safer. A poorly trained reviewer may overlook errors. A person under excessive time pressure may approve recommendations routinely. An employee may assume that a technically sophisticated system is more reliable than their own assessment, even when the circumstances clearly require further investigation. Human supervision can reduce risks, but only when the organisation has designed it to be effective. Simply attaching a human name to the approval record does not establish that independent judgment has taken place.

The level of supervision should also reflect the potential impact of the task. I see little value in requiring senior management to approve every automatically formatted internal document. That would undermine the very efficiency the technology was introduced to provide. But an automated system that can influence financial commitments, employment decisions, contractual obligations or access to essential services deserves considerably stronger safeguards. The objective is not to make automation unnecessarily complicated. It is to place meaningful controls where mistakes could have serious consequences.

And those controls have economic costs of their own. Verification takes time, monitoring requires resources and audit procedures must be maintained. A responsible automation strategy therefore needs to account for governance as part of the investment, not treat it as an unpleasant administrative detail to be added once the technology has already been deployed. If a business case looks profitable only because it excludes the cost of supervising the system, I would question the business case before approving the implementation.

After all, what is the purpose of reducing operating costs if the savings depend on abandoning the controls that make the operation reliable?

Who Defined Success, and Who Will Explain Failure?

The question I would ultimately bring into every AI implementation discussion concerns the definition of success. Before choosing a model, purchasing a platform or integrating an automated workflow, I want to know what the company is trying to improve. Revenue? Profitability? Customer response times? Error rates? Employee capacity? Service quality? These objectives are related, but they are not interchangeable. Improving one can sometimes weaken another. And if the management team has not decided how competing objectives should be balanced, introducing a sophisticated optimisation system will not resolve the disagreement. It may simply conceal it beneath a set of performance metrics.

A company can reduce customer support costs by automating more interactions, but what happens if customers find it increasingly difficult to resolve complex problems? It can generate more marketing content, but does that content improve qualified demand or merely increase the volume of material competing for attention? It can accelerate the preparation of financial reports, but are the underlying classifications, calculations and assumptions reliable? I am not opposed to any of these applications. I would simply refuse to treat an improvement in operational speed as sufficient evidence of business value.

This also raises an uncomfortable issue of responsibility within organisations. Who approved the objectives? Who selected the system? Who established the acceptable level of risk? Who is authorised to override its recommendations? And when an automated process produces a commercially damaging result, who has both the information and the authority to investigate what went wrong? These responsibilities need to be assigned before problems occur, not discovered through a series of increasingly urgent meetings after the damage is done.

The distinction between a developer, a technology provider and the organisation deploying a system is particularly important. Different parties may be responsible for different aspects of design, performance, integration and use. Legal obligations can also vary according to the application and jurisdiction. But from the perspective of business governance, the organisation cannot reasonably assume that adopting an external technology eliminates its duty to make informed decisions about how that technology is used. A system provider may explain technical limitations or correct a software defect. That does not replace the management team's obligation to assess whether the chosen application was suitable for its business purpose.

And here lies a rather uncomfortable contradiction in the enthusiasm surrounding artificial intelligence. Companies increasingly want systems capable of operating with greater autonomy, while simultaneously expecting accountability to remain simple, inexpensive and somehow external to the organisation. They want automated decisions, but not the burden of examining their consequences. They want more sophisticated technology, but not necessarily more sophisticated governance. They want to reduce human involvement, yet expect someone to appear immediately when the system makes a costly mistake.

I find that expectation economically and organisationally difficult to defend.

The value of artificial intelligence in business should be assessed through the outcomes it helps create, the resources it consumes and the risks it introduces or reduces. It should not depend on the number of automated workflows, the sophistication of the interface or the confidence with which the system presents its recommendations.

The most important investment may not always be in another AI tool. It may be in better data, clearer processes, more precise objectives and management practices capable of distinguishing a useful recommendation from an expensive mistake.

Because technology can improve the execution of a decision. It cannot make an inadequate business objective economically sound merely by pursuing it more efficiently.