AI is no longer the experiment. Now it has to pay off.

A management review of the kind that could currently be taking place in many companies: Several AI initiatives were launched a few months ago. Customer service is testing an assistant, marketing is producing content faster, sales is having conversations summarized automatically, and an internal knowledge solution is in pilot operation. Usage figures are developing well, employees report time savings, and the project status presentation gives the impression that the company has now set quite a lot in motion in artificial intelligence.

Then an uncomfortable question is asked: What of this has actually reached the business? Not in the number of users or prompts, but in productivity, revenue, costs, speed, customer retention, or a better offering. This question is becoming more important because AI is currently crossing a threshold in many companies. It is increasingly less about whether the technology should be addressed at all. What is becoming more interesting is which applications actually generate economic value and which primarily show that a company uses AI.

A recent study by KfW Research provides a noteworthy finding. Even after taking other key performance drivers into account, SMEs that use AI show stronger revenue growth, higher profitability, and higher productivity than companies that do not use AI. However, the caveat made by KfW itself is important: For productivity, the positive correlation can be identified for a broad group of AI users. For revenue growth and especially profitability, the group for which a positive correlation can be demonstrated is smaller and more concentrated among companies that are already growing or highly profitable.

This is interesting, but it is not proof that introducing AI automatically creates growth or profitability. For companies, a different question is therefore more relevant: Under what conditions does a technological possibility actually become a business case?

Usage is not yet impact

Consider an AI assistant in sales. After a few months, a large share of the team is using it, meeting notes are created automatically, and proposals can be prepared more quickly. Employees report saving several hours a week.

That may be a successful project. Whether it actually is one, however, depends on the problem it was originally meant to solve. If the goal was to relieve employees of tedious administration, the time saved may already be a relevant benefit. But if the aim was to increase sales productivity, it should eventually become visible what happens with the time gained. Are more qualified customer conversations taking place? Do proposals reach customers faster? Is the sales cycle getting shorter? Is conversion improving? Can the same team manage more business?

The distinction matters because companies often measure what is technically easy to measure. Usage figures, logins, generated content, or minutes of processing time saved become available quickly. Yet they do not necessarily say whether the business has improved. A technology can be used intensively and still generate only limited economic value. Conversely, an unspectacular use of AI in a critical core process can have a substantial effect even though hardly anyone in the company talks about it.

The business case should come before the AI use case

When a company considers new AI projects today, the first question should not be where else the technology could be used. That question almost inevitably produces a long list because generative and analytical AI can now, in principle, be applied to a great many activities.

For the business case, the reverse sequence is more useful: Where is the company currently losing business, time, or margin? Where is a bottleneck preventing growth? Which work becomes proportionally more expensive as revenue increases? Where are customers waiting unnecessarily long? Which decision is too slow because information is missing? And is there perhaps even a customer problem from which a new product could emerge using new technological possibilities? Only then should the company examine whether AI is a suitable lever.

This may sound like a minor difference in framing. In practice, it changes project selection considerably. Those who start with the technology find applications. Those who start with the economic problem may find investments.

This logic is not specific to artificial intelligence. In larger digitalization and business initiatives, the technical possibility was rarely the hardest question. The harder task was always to establish a clear link between the business idea, the technology, the changed process, and the economic effect. AI merely makes this problem more visible because impressive prototypes can now be built very quickly.

The time-saving trap

Business cases that rely almost exclusively on saved working time are particularly critical. The calculation is temptingly simple: A tool saves each employee two hours a week, the company employs 100 people, so 200 hours of additional capacity are created. Multiply that time by internal personnel costs, and a six-figure annual benefit appears within minutes.

The mathematics may be correct and the business case may still be too optimistic. If the same 100 employees continue to receive the same salaries after the tool has been introduced and the time gained is not used economically elsewhere, the company has initially created capacity but has not yet realized a corresponding cost saving. That can still be valuable, for example because overtime falls, quality rises, or employees have more time for demanding tasks. These effects simply should not be confused with realized savings.

The additional capacity becomes economically interesting where a second change follows. A service team can serve more customers with the same workforce, sales can hold additional conversations, development can bring products to market faster, or an administrative function does not have to add staff proportionally as the company grows.

A robust AI business case should therefore distinguish between two levels:

• Operational effect: Which activity becomes faster, better, cheaper, or disappears entirely?
• Economic translation: How does that operational effect change revenue, costs, capacity, risk, margin, or growth?

The part between these two levels is often the part that an impressive demo has not yet answered.

The greatest leverage may not lie in the easiest use case

For the first steps with a new technology, it makes sense to begin with relatively easy applications. Employees gain experience, technical and legal questions become visible, and the company develops a better sense of what works and what does not. However, this learning phase should not be confused with a long-term AI strategy.

If all competitors can use the same generally available assistant, this may create a productivity gain but not yet a sustainable competitive advantage. The greater leverage may lie somewhere else entirely: in pricing, production planning, procurement, maintenance,
in a central sales process, in product development, or in a new data-based offering for customers.

Such applications are usually more difficult. They require better data, a deeper understanding of processes, integration into existing systems, and people who understand both the business and the technological possibilities. Precisely for that reason, they may be more economically relevant.

For the first phase of AI, the question “What can we try?” was entirely legitimate. In the next phase, companies should ask more often where this technology can change something that truly matters to the business.

An experiment needs different rules from an investment — but not forever

Demanding a fully robust return on investment from an early AI pilot after only a few weeks would be shortsighted. Innovation needs room for learning, and anyone who assesses every idea solely against short-term result metrics from day one will ultimately end up mainly with projects whose outcome was already largely known.

Nevertheless, every experiment eventually needs a decision point. A company should know in advance what it wants to learn and when it will decide whether a use case should be scaled, changed, or terminated. The option to terminate is particularly important. A use case can work technically, be popular with employees, and still fail to generate an economic effect that justifies further investment. Stopping it then does not mean that the experiment failed. If it leaves the company with a clearer understanding of what works and what does not, it has served its purpose.

It only becomes problematic when a pilot project turns into a permanent state because no one fundamentally asks why the company is still investing.

Not every effect has to be immediately measurable in euros

At the same time, it would be too simplistic to assess AI exclusively by short-term cost savings. Some applications improve quality, reduce risks, or create capabilities whose economic value only becomes visible later. A company may, for example, make knowledge more accessible, increase the speed of decisions, or build capabilities that are relevant to future products.

Such effects also belong in a business case. They simply need to be named clearly. If a company consciously decides that a project should initially build a strategic capability and therefore does not yet need to make a short-term contribution to earnings, that is a legitimate investment decision. Less
convincing is a project initially sold as an efficiency initiative that, after savings fail to materialize, is suddenly described only as a strategic learning project.

Good business cases are not good because every benefit can be calculated to two decimal places. They are good when it is clear before an investment which assumptions are being made and how management will later judge whether those assumptions still hold.

The decisive question therefore does not begin only after go-live. During a digitalization or AI project, it should already be clear how it will be possible to tell a few months after implementation that real value has been created. After all, a technically functioning system and a successfully completed project are not the same as a changed operational reality.

AI becomes a management responsibility when the demo is over

The first phase of the AI debate was strongly shaped by possibilities. What can the technology do, which activities will it change, and which applications should companies try? This phase was necessary because economic potential can hardly be realized without first understanding the technology.

The next phase is likely to be more sober. KfW Research currently finds a positive correlation between AI use and company performance, but with clear differences depending on the metric and group of companies. That is a good reason to look more closely, but no reason to automatically turn every AI initiative into a driver of growth.

For a management team, this leads to five questions for its most important ongoing AI initiatives:

• Which specific business problem or opportunity was the starting point?
• What operational change was AI meant to create?
• How can that change actually be observed?
• How does it translate into economic or strategic value?
• What decision follows: scale, change, or terminate?

Those who can answer these questions do not necessarily need the largest number of AI use cases. What is more valuable is a robust understanding of which technology investments genuinely move the business forward.