Website icon Xpert.Digital

Even McKinsey has to do the math: When does artificial intelligence really pay off?

Even McKinsey has to do the math: When does artificial intelligence really pay off?

Even McKinsey has to do the math: When does artificial intelligence really pay off? – Creative image on the topic, featuring AI: Xpert.Digital

The AI ​​cost trap: Why companies need to control results, not computing power

How to correctly calculate AI costs: What really matters for companies

Is artificial intelligence worth it? The crucial calculation begins after the model price

Artificial intelligence can accelerate work, improve processes, and open up new business opportunities. However, increased usage does not automatically translate into increased profits: the more successful an AI application is, the more its costs can rise. In addition to model and cloud expenses, there are often costs for data, integration, monitoring, and post-processing. Therefore, simply looking at the price of individual requests is insufficient.

McKinsey illustrates just how challenging economic management can be, even for technologically experienced organizations. In the summer of 2026, CFO and Senior Partner Yuval Atsmon made it clear that the previously observed increase in internal AI costs could not simply continue until the following July. This is not evidence that the deployed systems are fundamentally unprofitable. However, it does highlight a key conflict of objectives: companies want to promote productive use – while simultaneously ensuring that expenditures do not increase faster than the actual benefits.

The crucial question, therefore, is not simply how many people have access to an AI assistant or what a single model request costs. Companies need to know what work is actually being done, how much a usable result costs overall, and whether it measurably improves the quality, speed, or efficiency of a process. Only when technical expenditures, human rework, and business outcomes are considered together can it be assessed when artificial intelligence truly pays off.

The more successful the implementation, the more dangerous the calculation – if no one controls the economic benefits

Artificial intelligence presents companies with a business paradox: a successful implementation can strain the budget more than a failed one. If new applications are hardly used, ongoing expenses remain manageable. However, if they prove helpful, demand, adoption, and computing requirements increase. A limited pilot project can quickly evolve into a company-wide infrastructure, with costs developing much more dynamically than originally planned.

McKinsey illustrates just how challenging this development is, even for organizations with significant technological and management expertise. Chief Financial Officer and Senior Partner Yuval Atsmon made it clear in the summer of 2026 that the previously observed increase in internal AI costs could not simply continue until the following July. This doesn't necessarily imply a general inefficiency of the systems in use. However, it does reveal a fundamental conflict of objectives: A company wants to promote productive use, but at the same time must prevent expenditures from growing faster than the resulting benefits.

For finance managers, this shifts the central question. The crucial factor is no longer solely how much a model call costs or how many employees have access to an assistant. What matters is what economically relevant work is actually being done, what the total costs are, and whether the result is reliable enough to be used within the company.

The real challenge, therefore, is not to buy AI as cheaply as possible. It lies in establishing a sound relationship between technical capabilities, organizational changes, and financial impact. Those who merely limit costs risk stifling productive applications. Conversely, those who interpret every use as progress risk expensively digitizing additional work without any corresponding added value.

From predictable workplaces to consumption-based knowledge work

A comparison with traditional enterprise software makes the difference clear. With a largely license-based model, immediate software expenditures can often be planned relatively well. A company procures a specific number of accesses, agrees on contract durations, and thus knows a significant portion of its ongoing obligations.

However, these costs were never entirely predictable. Implementation, integration, training, maintenance, and customization could incur significant additional expenses. Similarly, usage-based billing didn't begin with generative AI. Cloud services have already accustomed companies to paying based on actual usage.

AI, however, intensifies this logic because individual work steps can now directly trigger computing power. Research, document analysis, or processing a customer inquiry generates varying amounts of input and output data depending on the system. This may be compounded by searches, database queries, external tools, and additional verification steps. Therefore, the financial perspective must consider model access, cloud services, software packages, in-house computing capacity, and organizational operating costs together.

AI is not sold solely based on consumption. Flat-rate licenses, included usage quotas, and hybrid models remain relevant. Nevertheless, the economic impact of additional usage doesn't disappear. It can manifest itself in higher contract tiers, purchased additional capacity, performance limitations, or increased infrastructure requirements.

This changes how digital knowledge work is calculated. Another generated report or additional analysis is not automatically virtually free. Marginal costs may be low, but they must be considered in conjunction with testing, integration, and use. This very connection between technical production and operational application is often missing from early cost-benefit analyses.

Why falling model prices don't mean the all-clear

At first glance, technological progress seems to alleviate the cost problem. Powerful models are becoming more efficient, and prices are falling significantly for certain tasks. The Stanford AI Index 2025 documents a decrease in inference costs of more than a factor of 280 between November 2022 and October 2024 for a system comparable to GPT-3.5 at a defined test level. In the underlying analysis, the price fell from $20 to $0.07 per million tokens. This refers to a specific performance level, not the overall cost development of every AI application.

This limitation is crucial. A lower price for comparable technical capability does not mean that a company's overall AI costs will decrease proportionally. For that to happen, factors such as the scope of tasks, frequency of use, quality requirements, and system architecture would have to remain unchanged.

A simplified example illustrates the relationship. If the average technical price per use decreases by 60 percent, while the number of uses increases fivefold, the corresponding expenses double. The original 100 cost units become 200. The technology has become significantly cheaper per use, yet the overall budget is still larger.

In addition, companies can tackle more demanding tasks. Instead of short text drafts, they analyze extensive contract portfolios, process complex programs, or prepare multi-stage decisions. Such applications may require more powerful models, longer contexts, and multiple processing passes.

Economically, this represents a potential feedback loop: Increased efficiency lowers the barrier to adoption and can thereby trigger additional demand. Whether total expenditure subsequently rises or falls depends on the strength of this demand effect. Lower unit prices are therefore an important step forward, but no substitute for quantity planning and cost-effectiveness analysis.

When an assistant becomes a digital process chain

The cost of a single chat response is only one aspect of the problem. With increasing automation, the unit that needs to be financially managed changes. It's no longer just the visible response that matters, but the entire processing chain.

A quotation generation system could first classify a customer inquiry, then search technical documents, compare product information, retrieve prices, and create a draft. A further processing step would then check whether mandatory information is missing. In case of conflicting data, part of the processing process would restart.

From the user's perspective, a service is offered. From the infrastructure's perspective, numerous model calls and tool accesses may have occurred. Therefore, repetitions, excessively long inputs, unnecessary tool calls, and an inappropriately large model are among the starting points for efficient process control.

A particularly important question is when such a process chain ends. A human employee often recognizes that further research no longer yields any significant added value. An automated system requires suitable rules for this. Otherwise, it may use additional computing power even though the quality of the results hardly improves.

This leads to a business management design challenge: Every automated process needs a defined goal, appropriate quality requirements, and limits on time, repetitions, and resource consumption. Technical autonomy must not be confused with unlimited spending authority.

Tokens explain consumption, but not business success

Tokens are processing units of language models. They can represent words, word parts, or other strings of characters. They are relevant for technical billing and consumption analysis. However, they are not sufficient as the sole key performance indicator for business management.

A CFO can initially only deduce from an increasing token volume that more corresponding model processing has taken place. Whether this has resulted in faster customer service, avoided errors, or generated additional revenue remains unclear.

Conversely, high token consumption is not automatically wasteful. Extensive technical testing can be economically sound if it identifies a costly problem early on. A very short, inexpensive text, on the other hand, can be useless if it misses the point.

The viable perspective, therefore, is that consumption data must be linked to business results. Accordingly, the FinOps Foundation recommends considering costs per successful outcome and linking these to the results for which the respective business unit is responsible.

This doesn't mean abolishing technical metrics. Without transparency in consumption, misconfigurations and unusual cost increases are difficult to detect. Tokens, however, are a diagnostic tool, not a standalone measure of success. They explain what infrastructure was used for. Whether this investment was worthwhile must be determined at a different level.

The complete calculation begins after the model calculation

A reliable AI cost analysis encompasses more than just the model's price. Beyond direct use, it often includes data preparation, interfaces, access rights, monitoring, quality assurance, training, and ongoing adjustments. The internal time commitment of business departments and IT must also be considered.

These costs are often underestimated in pilot projects. A small team can manually bridge missing interfaces, personally check results, and resolve special cases informally. What works for a few applications quickly becomes a bottleneck with high usage.

The economic analysis should therefore differentiate between implementation costs and ongoing operating expenses. One-off expenditures should not disappear simply because they are booked in a different budget. Conversely, they should not permanently burden each individual use as if the entire implementation had to be constantly refinanced.

A clear example: If an application incurs €120,000 in direct technology costs annually, plus €80,000 for operation and integration, and €100,000 for technical review, the total ongoing costs amount to €300,000. If this is used to complete 60,000 actually usable processes, each such process costs an average of five euros.

The direct technology alone would only generate two euros per transaction. Both figures are mathematically correct, but they answer different questions. The five euros are crucial for determining whether the overall process is economically viable. For targeted optimization of model usage, the two euros remain relevant.

The cheapest model can produce the most expensive process

Model prices are relatively easy to compare side-by-side. Evaluating the subsequent work involved is more difficult. It is precisely there that a supposed price advantage can turn into its opposite.

A purely hypothetical comparison illustrates the scale of the costs. A simple model costs €0.02 per process but requires an average of four minutes of manual post-processing. A more efficient model costs €0.12 and requires only one minute of post-processing. Assuming an internal time cost rate of €60 per hour, this results in €4.02 per process in the first case and €1.12 per process in the second.

The model with six times the immediate usage cost would therefore be significantly cheaper in the overall process. The difference arises not from a particularly low technology price, but from less human effort.

However, the opposite conclusion would also be wrong: the most powerful model is not necessarily the most economical choice. If a simple classification task can be reliably performed with a smaller system, additional model performance may be unnecessary.

The appropriate architecture therefore assigns tasks to different methods based on difficulty and risk. Simple cases can be automated cost-effectively. Difficult cases receive more powerful support or are handed over to humans. The crucial factor is not maximum intelligence at every processing step, but rather the right performance at the right place.

Why only usable results belong in the denominator

One of the most important key performance indicators (KPIs) is the effort required per successfully completed business transaction. The concept of "success" must be precisely defined in a professional context.

When responding to a customer inquiry, it's not enough that a reply has been generated. What matters is whether the issue was resolved correctly and no avoidable follow-up questions arise. Similarly, when processing an invoice, it's not enough that fields are filled in. The data must be accurate, complete, and usable for the subsequent process.

An application generating 100,000 results can be less cost-effective than one generating 70,000 results if only half of the former are usable. Dividing the total cost by all system expenses then results in a misleadingly low unit price.

It is also useful to highlight differences between case groups. An average transaction price can obscure the fact that standard cases are profitably automated, while rare special cases incur exceptionally high costs.

The distribution of processing times is also relevant. A good average is of little use if individual urgent tasks take an uncontrolled amount of time. Therefore, companies should consider not only averages but also typical ranges, outliers, and the frequency of manual intervention. Economic efficiency lies not only in favorable, typical scenarios but also in the ability to manage challenging situations.

 

🤖🚀 Applied AI Lab: Faster, safer & smarter to AI solutions with UNFRAME.AI

Applied AI Lab – Creative image on the topic, using AI: Xpert.Digital

Here you will learn how your company can implement customized AI solutions quickly, securely and without high entry barriers.

The Applied AI Lab is your all-inclusive package for applied artificial intelligence. Instead of dealing with complex technology, expensive infrastructure, and lengthy development processes, you receive a ready-made solution tailored to your needs from a specialized partner – often within just a few days.

The key advantages at a glance:

⚡ Rapid implementation: From idea to ready-to-use application in days, not months. We deliver practical solutions that create immediate added value.

🔒 Maximum data security: Your sensitive data stays with you. We guarantee secure and compliant processing without sharing data with third parties.

💸 No financial risk: You only pay for results. High upfront investments in hardware, software, or personnel are completely eliminated.

🎯 Focus on your core business: Concentrate on what you do best. We take care of the entire technical implementation, operation, and maintenance of your AI solution.

📈 Future-proof & scalable: Your AI grows with you. We ensure continuous optimization and scalability, and flexibly adapt the models to new requirements.

More information here:

 

Cost control and its impact on productivity

Minutes saved are not the same as money saved

Many AI business models begin with a simple calculation: the number of employees multiplied by the number of minutes saved and an hourly rate. The result can be impressive, but it's not sufficient proof of actual savings.

If an employee spends 20 minutes less per day on drafts, their salary won't decrease. Initially, this creates additional capacity. This can be valuable, but it must be used productively.

An economic benefit becomes apparent, for example, when overtime decreases, external services are no longer needed, vacancies don't have to be filled, or additional customer orders can be handled by the existing team. Similarly, more time can be available for quality improvements or customer service.

These effects differ financially. Avoided external costs are often immediately apparent in the profit and loss statement. Additional capacity, on the other hand, is initially a potential benefit. Improved quality can be of considerable value, but requires a different measurement than pure time savings.

The distribution of time savings is also important. Many small reductions in workload cannot automatically be combined to create a completely vacant position. This would require corresponding changes to tasks, responsibilities, and workflows.

A sound cost-benefit analysis therefore separates theoretically saved time, actually available capacity, and realized financial impact. Mixing these categories easily creates a false sense of precision that worsens rather than improves the investment decision.

The underestimated value of faster decisions

Not every economically viable AI application has to reduce personnel costs. In many processes, the decisive benefit lies in the fact that information becomes available earlier and decisions can be made in a timely manner.

An industrial company might be able to react more quickly to an impending supply shortage. A sales department could submit offers sooner. A finance department could investigate anomalies earlier. The value then arises from a changed course of action, not from the sheer volume of text produced.

Here too, caution is needed in the evaluation. Faster decisions are not automatically better decisions. If a system merely disseminates unreliable information more quickly, the economic damage can increase.

Therefore, the analysis should combine speed and quality. For example, in the case of offers, processing time, closing rate, and achieved contribution margin would all be relevant. In logistics, response time, delivery capacity, and the costs of special measures could be considered together.

Assessing avoided losses is particularly difficult. A production outage that did not occur cannot simply be attributed entirely to AI. Plausible comparison groups, historical experience, and documented decision-making processes can help. Nevertheless, a degree of uncertainty often remains, which must be explicitly included in the economic analysis.

Four tests replace the enthusiasm for mere use

Four related economic checks can be formulated for each AI-supported workflow.

First, it must be clear whether a relevant task has been completed. An additional report is not progress if no one needs it. An automated summary only has value if it supports or replaces a meaningful activity.

The full costs must then be recorded. This includes not only technology and infrastructure, but also human oversight, post-processing, operation, and handling of exceptions. The calculation must not end at the interface between IT and the business department.

The third test concerns usability. The result must meet the technical requirements of the specific application. An acceptable error rate for an internal design may be completely unsuitable for a safety-critical test.

Finally, the connection to the company's impact must be evident. Has a process been accelerated? Has quality improved? Were costs actually avoided or additional contribution margins achieved?

These checks are not independent checkboxes. They are part of a chain of effects. Without a relevant task, quality is meaningless. Without a usable result, time savings are questionable. Without a complete cost analysis, the economic benefit cannot be assessed.

Why the current procedure is the correct comparison

The alternative to an AI application isn't always to do nothing. Often, manual processes, traditional software, rule-based automation, or external services already exist. The new solution must compete with these existing options.

For clearly structured tasks, a conventional rule can be more cost-effective, stable, and easier to understand than a language model. Conversely, AI can offer advantages where unstructured documents, linguistic diversity, or frequently changing information need to be processed.

A sound decision therefore requires a baseline measurement. How long does the current process take? What errors occur? What are the rework efforts and costs? What quality is actually achieved?

Without these initial values, it remains unclear later whether performance has improved. A seemingly more modern process may be economically weaker than an unspectacular existing method.

The comparison must also remain fair. If the AI ​​solution is compared to a particularly poorly organized legacy process, its benefits may appear greater than they would be compared to a simple process improvement. The crucial question, therefore, is not only whether AI helps, but whether it offers the best balance of effort, impact, and risk among the available alternatives.

Rigid annual budgets are being replaced by controllable usage scenarios

With rapidly growing AI usage, a single planned figure often provides a greater sense of security than actually exists. Multiple scenarios that reflect different usage intensities and cost developments are more meaningful.

For example, a company might have 10,000 employees. If 4,000 of them use an application ten times a day for 20 working days, that results in 800,000 uses per month. Assuming direct technology costs of an average of €0.03, that would amount to €24,000.

If the number of active users rises to 8,000 and daily usage doubles simultaneously, the monthly volume increases to 3.2 million transactions. At an unchanged price, this results in costs of €96,000. If the price were to halve in the meantime, the costs would still amount to €48,000, which is twice as high as in the initial scenario.

This example calculation is not a market forecast. It shows which factors need to be planned together: active users, usage per person, complexity per process, and average costs.

Additionally, assumptions about acceptance, quality levels, and economic benefits are needed. A higher budget can be justified if the additional contribution margin increases significantly. Conversely, a lower budget is not a success if it merely results from underutilization of an already expensively built infrastructure.

Cost control must not penalize productive applications

Across-the-board cuts seem simple, but can create perverse economic incentives. If a team demonstrably performs profitable work with additional AI usage, a rigid consumption limit could potentially be counterproductive.

Conversely, a generous budget without accountability for results is equally problematic. Employees might use elaborate models for simple tasks or create additional content whose usefulness is not verified.

A differentiated approach to management is advisable. Experimental applications require limited learning budgets and clearly defined testing deadlines. Established processes need reliable resources and transparent unit costs. Growth applications can afford to spend more if their additional results justify the increased expenditure.

Costs and benefits must be aligned organizationally. If IT pays for all expenses while business units expand usage without taking responsibility, a perverse incentive is created. Conversely, if individual teams have to justify every technical step, meaningful collaborative solutions can be prevented.

The challenge, therefore, is to combine transparency with the ability to act. Budgets should enable sound economic decisions, not merely limit technical activity.

Finance and technology must speak the same language

The extent to which AI has become a key component of corporate financial management is demonstrated by the 2026 State of FinOps survey. In it, 98 percent of respondents reported managing AI spending; two years prior, this figure was 31 percent. These figures reflect the surveyed FinOps community and should not be interpreted as representative of all companies. Nevertheless, they illustrate the growing importance of AI within this functional area.

The organizational consequence is closer collaboration between finance, IT, and specialist departments. None of these groups possesses all the necessary information on its own.

The finance function understands capital requirements, performance targets, and budget constraints. Technical teams understand model behavior, architecture, and consumption. Business units can assess whether results are practically applicable and can improve workflows.

Joint management therefore requires clearly defined responsibilities. For each important application, it should be clear who is responsible for its technical benefits, who manages its technical operations, and who monitors its economic performance.

Equally important is a shared set of terms. If IT talks about successful requests, but the business department counts unresolved customer issues, they are measuring different things. Only a jointly defined unit of performance makes cost reports meaningful for decision-making.

Data quality and control are not additional costs incurred later

The quality of an AI application depends not only on the model. The underlying information must also be suitable, up-to-date, and accessible. Unclear responsibilities, conflicting documents, and outdated data can lead to additional research, queries, and manual corrections.

This raises an important investment question. Sometimes it is more economical to improve the data foundation first, rather than implementing an even more powerful model. A system cannot compensate for organizational disorder at any cost.

Human oversight must also be designed to account for the risks involved. A complete audit of every expenditure can largely negate any productivity gains. Conversely, a complete abandonment of auditing can lead to unacceptable consequences of errors.

The appropriate solution depends on the application. An internal draft text requires different safeguards than legally binding customer information or a process with significant financial consequences. Phased approvals, random sampling, and the targeted transfer of difficult cases are all conceivable options.

It is economically crucial to plan this control architecture from the outset. If it is only added after a problematic incident, costs and processing times may change so drastically that the original cost-benefit analysis no longer holds true.

The choice of supplier goes far beyond price

A low initial price is attractive, but says little about the long-term total costs. As integration grows, switching costs can arise: interfaces need to be adapted, processes retested, and employees trained.

Switching providers may also necessitate new quality checks. Two models that perform similarly well in general comparative tests may not deliver the same results in a specific business process.

Therefore, technological interchangeability is an economic asset. Modular interfaces, documented testing procedures, and organized data management can facilitate future migrations. At the same time, this flexibility also creates additional costs. Not every application requires multiple vendors operating in parallel.

The comparison between external services and in-house operation is similarly nuanced. In-house infrastructure offers more direct control over certain costs and operating conditions, but also entails investments, staffing needs, utilization risks, and the need for technical upgrades.

External utilization can be advantageous when demand fluctuates. In-house operation can become attractive with suitable, sufficiently high, and stable utilization. There is no universally winning solution. The decisive factors are the actual load profile, the requirements, and the total costs over the period under consideration.

Why consulting firms are feeling the change particularly clearly

For knowledge-intensive service providers, AI affects not only the cost structure but also the revenue logic. As research, analysis, and presentation preparation become faster, the ratio between time spent and billable output changes.

With time-based pay, higher productivity can initially mean that fewer hours are billed for the same task. An economic advantage then arises, for example, through additional orders, adjusted pricing, or higher-quality services.

With fixed prices, faster processing can improve the margin. This is conditional on maintaining quality and customer benefits, and on ensuring that additional technology and control costs do not negate the advantage.

At the same time, competition can intensify. As standardized analyses become more readily available, their mere creation loses its differentiating power. Reliable classification, access to relevant information, implementation expertise, and the assumption of responsibility become more valuable.

The McKinsey case therefore represents more than just a rising software market position. It highlights the need for companies to simultaneously examine their production methods and their value creation strategies. Those who generate labor faster but continue to sell it according to an unsuitable revenue logic may only partially realize the productivity gains.

More content does not automatically mean more value

Generative AI lowers the barrier to producing texts, analyses, presentations, and variations. However, this can create a new form of operational waste: additional content that consumes attention without improving decisions.

The immediate costs of creating the model may be low. However, the subsequent review, coordination, and approval processes can become expensive. If multiple managers need to review additional documentation, their time expenditure may exceed the total technical effort required for its creation.

Therefore, the demand side also needs to be examined. Does a process really need more reports, or rather a clearer decision-making template? Are additional variants being used? Does the new analysis replace existing work, or is it simply added to it?

A productive implementation does not indiscriminately automate every existing task. It distinguishes between necessary work, work that can be improved, and work that can be eliminated.

This presents a difficult management challenge. The simplest approach is to accelerate existing routines with AI. However, the greater economic leverage may lie in eliminating redundant routines. Not every task that can be automated deserves to continue.

The correct measure is additional economic surplus

Ultimately, AI must be measured against a clear economic question: What additional benefit does it offer compared to the best realistic alternative, and what additional effort is required to achieve it?

Companies should differentiate at least between realized cost reductions, available capacity, additional contribution margins, and plausibly assessed quality or risk effects. These categories can together present a compelling picture, but must not capture the same benefit multiple times.

The time perspective is equally important. An application may seem unattractive in the first year due to high implementation costs, but become profitable later. Conversely, an inexpensive pilot project may lose its initial appeal due to increasing complexity and ongoing maintenance requirements.

A well-founded decision therefore requires not artificial certainty, but transparent assumptions. What level of use is required? What quality standard is needed? How much manual rework is acceptable? At what cost increase will the architecture have to be changed?

The strongest financial leadership will not celebrate either the lowest AI bill or the highest usage rate as a success in itself. They will recognize when additional spending creates more added value and when it merely generates more digital activity.

This is precisely where the real test lies: AI will not become economically viable simply by producing impressive results. It will become economically viable when these results are reliably used to create real-world applications

 

Consulting - Planning - Implementation

Konrad Wolfenstein

I would be happy to serve as your personal advisor.

You can contact me at wolfenstein∂xpert.digital or

Just call me on +49 7348 4088 965 .

LinkedIn
 

 

Leave the mobile version