AI in SMEs: How companies can turn technology into real productivity gains – Europe's AI opportunity

Europe's AI competition will not be decided in the laboratory – but in the companies

Artificial intelligence is spreading throughout European companies – but simply using individual tools doesn't automatically translate into economic advantage. The crucial factor is whether companies can reliably integrate AI into their operations, improve processes, and measure the actual benefits. This is precisely where Applied AI Labs come in: In these application-oriented development units, technical experts and business leaders work together on concrete business problems. They examine where AI makes sense, what data and control mechanisms are needed, and whether the investment pays off in everyday use. This presents a major opportunity for Europe: Companies don't necessarily have to develop leading basic models themselves, but they do need to master the connection between available technology and their own business. The decisive competitive advantage arises where AI reliably leads to better processes and sustainable results.

Related to this:

• Introducing the Applied AI Lab: Closing the Enterprise AI Gap

Applied AI Labs: Europe's AI competition will be decided in the companies

Those who only buy models finance the progress of others – those who change processes create their own added value

European companies are facing a fundamental shift in artificial intelligence. For a long time, the discussion focused on the capabilities of large language models, data centers, investment sums, and which providers are technological leaders. These topics remain important. However, the economic success of most companies hinges on a different capability: they must transform widely available AI into reliable, affordable, and sustainable business applications.

It is precisely at this intersection that Applied AI Labs are gaining importance. These are application-oriented development units where AI specialists, software developers, data experts, and business process owners collaborate to address concrete business problems. The term does not refer to a uniformly defined organizational model. Such a lab can be part of a technology provider, a corporate group, a research institution, or a cross-industry network. What matters is not the label, but the ability to transform technical potential into tangible business value.

The core economic question is therefore no longer solely which model provides the best answers. It is which organization integrates AI into its processes in such a way that quality, speed, capacity, and profitability actually increase. Europe still needs research and infrastructure for this. However, it also urgently needs a professional implementation layer bridging the gap between technology and day-to-day business.

Europe's AI revolution is measurable, but far from widespread

The use of artificial intelligence in European companies is increasing significantly. By 2025, 20 percent of the surveyed companies in the European Union were using at least one AI technology. A year earlier, this figure was 13.5 percent. The increase of 6.5 percentage points represents a relative growth of almost half within a single year. However, behind this dynamic lies a sobering starting point: around four out of five companies in the surveyed group still did not use any of the AI ​​technologies surveyed. The statistics refer to companies with at least ten employees in selected economic sectors and therefore do not represent all micro-enterprises or the entire economy.

Equally important is the distinction between usage and economic penetration. Even a single application can be enough for a company to be statistically classified as an AI user. This doesn't reveal how many employees actually work with it, which business processes are affected, or what contribution the technology makes to the bottom line. A company with an AI-supported text tool and a company with end-to-end AI-supported production planning could belong to the same basic category in such a survey.

For an economic evaluation, the mere question of implementation is therefore insufficient. Crucial factors are the intensity of use, the depth of integration, and the impact on results. Are only individual tasks completed faster, or is an entire business process improved? Are demonstrably fewer errors occurring? Are delivery times shortened? Can additional customers be served?

This distinction explains why the next phase of the AI ​​race is likely to be characterized less by spectacular demonstrations and more by unspectacular integration work. Implementation is a start. Economic success only comes with its integration into businesses.

The real gap lies between different implementation capacities

The challenge becomes particularly clear when comparing company sizes. In 2025, around 17 percent of small, 30.36 percent of medium-sized, and 55.03 percent of large companies in the EU were using AI technologies. Significant differences also existed between sectors: In the information and communication sector, the figure was 62.52 percent, while in professional, scientific, and technical services it was 40.43 percent. In the construction sector, however, it was only 10.79 percent.

These differences cannot be explained solely by varying openness to new technologies. Large companies can spread development costs across more locations, employees, and business processes. They more often have specialized IT departments, structured data repositories, in-house legal departments, and the financial resources to test multiple approaches simultaneously.

In contrast, a medium-sized industrial supplier may have to make do with a small IT team already responsible for day-to-day operations, security, and existing digitalization projects. Even a promising AI application would then compete with urgent daily tasks.

Furthermore, there is the issue of minimum size. Regardless of company size, a custom-developed solution requires certain prerequisites: data access, interfaces, testing, security measures, and responsible operation. With few processes, these fixed costs are spread across a small usage volume. The same technical effort might be profitable for a large corporation but initially uneconomical for a smaller company.

Applied AI labs can close this gap by making development components reusable, pooling industry-specific knowledge, and giving companies access to skills they couldn't otherwise afford to finance on their own. However, this must not lead to a permanent dependence on external specialists.

What an application-oriented AI lab must actually be able to do

An economically viable Applied AI Lab is neither a mere training room nor a showcase for new software. Its task is to systematically reduce uncertainty. It must determine which problems can be effectively addressed with AI, which technical solutions are sufficient, and under what conditions productive use is justifiable.

Therefore, the first step is not selecting a model, but describing the problem. For example, a company might be experiencing excessively long processing times in its technical service. The cause could lie in scattered information, unclear responsibilities, poor documentation, or unnecessary approval steps. A language model can help with information retrieval. However, it doesn't automatically eliminate organizational inconsistencies.

A high-performance laboratory therefore combines technical development with process analysis. It examines what data is available, what decisions need to be made, and where human responsibility remains indispensable. It then develops a solution that is tested under realistic conditions.

This also includes the willingness not to use AI. If a problem can be solved more cost-effectively and reliably through a clearer data structure, a simple search function, or a fixed business rule, this is often the better economic decision. An Applied AI Lab should therefore not maximize the most extensive use of AI, but rather its operational benefits.

The difference between a technology lab and an effective implementation unit ultimately lies in the responsibility for the results. A convincing prototype is an intermediate step. A reliable contribution to quality, productivity, or growth is the real goal.

Why more model performance doesn't automatically mean more profit

The quality of an AI model is important, but it alone does not determine business success. A high-performing model may generate little benefit in a poorly prepared environment. Conversely, a smaller and less expensive model can operate very efficiently in a clearly defined process.

The interplay of several components is crucial. The system requires relevant information, clear work instructions, reliable tools, and clearly defined scopes of action. It must be able to recognize when an answer is uncertain and when a process should be handed over to a human.

Especially in companies, the bottleneck often lies outside the model. Product information may be outdated. Permissions may prevent access to relevant documents. Customer data may exist in multiple and contradictory versions. A technically sound result is of little use if it cannot be written back to the primary business system.

This results in an important economic shift: With more easily accessible models, the quality of complementary business resources gains in importance. These include proprietary data, process knowledge, customer relationships, and the ability to actually implement changes.

This presents an opportunity for European companies. They don't necessarily need to develop a leading baseline model themselves to become competitive with AI. However, they do need to master the connection between available technology and their own business model. Those who completely relinquish control of this connection risk losing a growing share of the value chain to external platforms and integration partners.

Productivity data speaks in favor of AI, but against simple promises of salvation

Empirical findings show that AI can have economically significant productivity effects. A study published in early 2026, based on linked data from more than 12,000 non-financial companies in the EU and the US, estimates that the introduction of AI will increase labor productivity by approximately four percent. The study attributes this effect to increased capital resources and finds no negative employment effects at the company level in the short term. However, the productivity gains are concentrated in medium-sized and large companies. Complementary investments in software, data, and training play a crucial role.

These results are significant but must be carefully interpreted. An estimated level difference of four percent does not mean that every company will automatically become four percent more productive after implementing an AI tool. Nor is it a guaranteed annual growth rate. The study uses an econometric identification approach, the validity of which depends on certain assumptions.

Even the absence of negative employment effects in the short term does not allow for any reliable predictions about long-term changes. Companies can initially use additional capacity to reduce backlogs, offer more services, or acquire new customers. However, job profiles, staffing needs, and market structures can change over longer periods.

This presents a clear task for Applied AI Labs: they must organize the complementary investments that transform technical availability into productive application. The benefits do not arise from AI alone, but from a combination of technology, data, skills, and organizational transformation.

Minutes saved do not yet translate into a business profit

A particularly common error in valuation is to immediately classify every minute of work saved as a cost reduction. In practice, the relationship is more complex.

If employees only need 45 minutes instead of 60 for a task, this initially frees up capacity. Whether this translates into a financial benefit depends on how this capacity can be used. In high-demand areas, additional orders can be processed. In a busy service department, waiting times can be reduced. In other cases, the time savings are spread across many small interruptions and barely affect the bottom line.

In addition, there is the effort required for review. An AI can create a design in seconds, but checking it afterwards takes several minutes. If the design contains errors that are difficult to detect, the net effect can even be negative. Therefore, the entire process chain must be considered: preparation, creation, review, correction, approval, and any necessary rework.

A sound cost-benefit analysis separates different categories of impact. Direct cost reductions must be evaluated differently than increased capacity, improved quality, or avoided risks. Additional revenue, in turn, must be adjusted for the associated variable costs.

Applied AI labs should therefore work closely with controlling. Without this connection, impressive usage figures can easily be generated without any corresponding improvement in results. The economically decisive metric is not the number of responses generated, but the value of reliably completed business transactions.

A transparent model calculation reveals the limits of optimism

The importance of this distinction is illustrated by a simplified, explicitly hypothetical calculation. A company employs 100 people in a document-intensive department. Through an AI application, each person saves an average of 20 minutes over 200 working days. This results in approximately 6,667 hours of additional annual capacity.

Assuming total costs of €50 per working hour, this equates to a calculated capacity value of approximately €333,000. However, this sum does not yet represent realized savings. If the company can only utilize half of the freed-up time economically, a usable value of approximately €167,000 remains.

Assuming ongoing costs for models, infrastructure, support, quality assurance, and internal maintenance amount to €90,000 annually, this leaves a net annual return of approximately €77,000 before taxes and financing effects. With one-time implementation costs of €150,000, the simple amortization period under these assumptions would be just under two years.

The calculation changes significantly if the actual time savings are less than expected or if more control is required. Conversely, the benefits can increase if faster processing generates additional revenue or prevents costly errors.

The key insight, therefore, does not lie in a specific result. It lies in the sensitivity of the assumptions. A professional laboratory must reveal which variables determine the business value and how strongly the result reacts to changes. Only then can a sound investment decision be made.

In medium-sized industrial companies, process knowledge counts more than demonstration

For industrial companies, there is considerable potential in combining technical knowledge with recurring workflows. Possible applications range from evaluating technical documentation and providing customer service support to preparing test reports and quotations.

For example, a machine manufacturer can provide service technicians with controlled access to maintenance manuals, fault histories, and approved troubleshooting information. This becomes economically advantageous when the search for relevant information is shortened, follow-up questions decrease, and systems are available again more quickly.

However, the requirements differ fundamentally from those of a general text application. A technically plausible but incorrect course of action can lead to damage or security risks. Document versions, machine types, and permissible work steps must therefore be clearly defined. The system should not conceal uncertainty behind fluent language.

Furthermore, a distinction must be made between voice-based assistance and industrial control. A speech model can explain information or prepare a workflow. This does not mean, however, that it should control a machine without additional technical safeguards.

Applied AI labs in such environments require close collaboration with production, service, quality assurance, and occupational safety. The added value arises from a precise understanding of the application context. Those who treat industrial AI solely as a software project underestimate the importance of physical processes and their economic consequences.

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

A managed AI platform is your all-inclusive, worry-free solution for 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:

• Managed AI Platform

In logistics, the combination of forecasting, optimization, and communication is gaining importance

In logistics and supply chain management, AI is not a single tool, but rather a bundle of different methods. Demand forecasting, inventory optimization, image recognition, document processing, and speech-based support each solve different problems.

A language model can, for example, summarize a delivery deviation in an understandable way or extract the relevant information from various messages. For concrete route planning, however, mathematical optimization methods are often crucial. Statistical and machine learning methods may be more suitable for predicting arrival times.

A well-designed, economically sound Applied AI Lab combines these methods instead of assigning every task to a single language model. For example, it can develop a solution that detects malfunctions, calculates possible courses of action, and provides the responsible dispatcher with a transparent basis for decision-making.

Optimization should not focus solely on a single key performance indicator. Lower inventory levels can reduce tied-up capital, but at the same time impair delivery capability and resilience. Higher vehicle utilization can, under certain conditions, lead to longer waiting times or reduced flexibility.

The task, therefore, is to make conflicting objectives transparent. AI can help prepare better decisions. However, the specific combination of costs, service levels, and risk that a company aims for remains a business policy decision.

This is precisely where the value of application-oriented laboratories becomes apparent: they combine technical processes with an economically sensible objective function.

Data quality is becoming an investment issue, not an IT side issue

Many AI projects encounter difficulties that already existed before the technology was introduced. Inconsistent master data, unclear document status, and poorly maintained knowledge bases are not automatically resolved by AI. In some cases, their consequences even become more apparent.

A system that receives conflicting information cannot reliably resolve these contradictions through well-formulated explanations. Similarly, integrating a knowledge database does not fundamentally prevent incorrect or incomplete answers. It merely improves the information base, provided that selection, timeliness, and access function properly.

From an economic perspective, high-quality data is a complementary capital asset. While its preparation incurs costs, it often enables multiple applications simultaneously. A well-structured product data model can benefit sales, service, purchasing, and technical documentation.

Therefore, companies should not attribute data investments solely to a single AI project. Part of the effort creates a reusable foundation for further improvements. At the same time, this platform approach must not become an excuse for endless preliminary work.

A step-by-step approach is advisable: First, the data is improved for an economically relevant use case. Then, reusable structures are systematically expanded. This creates a link between short-term benefits and long-term data quality.

An Applied AI Lab must also clarify ownership and responsibility issues. Who updates information? Who decides on approvals? Who is internally liable for faulty data? Without such responsibilities, even a technically sound solution will lose reliability over time.

The transition from assistant to acting system increases the need for control

A key development is that AI is no longer just generating information, but is increasingly expected to execute work processes. For example, a system can classify a process, extract data from documents, create a draft in the business system, and prepare it for approval.

With each additional step, the potential benefit increases, but so does the consequence of an error. An inaccurate summary is different from an incorrectly triggered order or an unauthorized change to customer data.

The economically sensible answer is not blanket full automation. Instead, tiered authorization is needed. Some tasks can be automated, others only within narrow limits, and still others only after human approval.

Furthermore, errors must be limited and correctable. This includes traceable logs, controlled access, rollback procedures, and clear rules for exceptions. A production system should not force operation when uncertain, but rather stop at an appropriate point.

Applied AI labs must therefore not only ask whether a task is technically feasible. They must also examine the consequences of incorrect execution and the resulting control costs. The optimal level of automation lies where the additional benefits and risks are in reasonable proportion.

The underestimated ongoing task begins after the successful pilot project

A pilot project can demonstrate that an application works in principle. It does not yet prove that it will function reliably over months under varying conditions.

During operation, data, products, processes, and user requirements change. Even replacing the underlying model can alter an application's behavior. Therefore, a solution that performs well in testing requires ongoing monitoring.

Operating costs include not only model calls and computing power, but also testing, security audits, error analysis, support, documentation maintenance, and adaptation to new business requirements. For more complex applications, this organizational effort can be more economically significant than the model's price alone.

An Applied AI Lab should therefore clarify early on who will take over the application after its development. If it remains permanently with the lab, a bottleneck may arise. If it is handed over too early to an unprepared department, quality losses are likely.

A viable model combines central technical standards with decentralized professional responsibility. The platform provides secure basic functions. The respective business unit is responsible for the specific process and its benefits. The lab supports further development but does not permanently replace operational responsibility.

This handover is not a secondary detail. It determines whether a successful experiment will result in a robust business capability.

External implementation partners can accelerate the process but also create dependencies

The expansion of application-oriented units by major technology providers underscores the growing importance of customer proximity. OpenAI has announced an additional 200 positions in Dublin related to an Applied AI Lab. Combined with a previous announcement of 250 positions, this brings the planned total to 450. These are announced hires, not a fully staffed workforce. The intended tasks include developing customer applications, as well as data protection and integration work.

This development is economically understandable. Suppliers need not only high-performance models, but also successful customers. The more deeply a technology is integrated into business processes, the greater its potential usage, customer loyalty, and additional revenue streams.

For companies, such collaboration can facilitate access to scarce skills. They receive support in architecture, development, and deployment. At the same time, switching costs can increase if applications are heavily reliant on proprietary tools and platform features.

The crucial question, therefore, is not whether external help is generally useful. It is which skills should remain within the company. Process knowledge, evaluation criteria, data structures, and decision-making authority should not reside entirely with the service provider.

A sustainable partnership accelerates implementation and simultaneously strengthens the company's own capacity to act. A problematic partnership generates good results in the short term, but makes the company dependent again with every expansion and adaptation.

Digital sovereignty means, above all, robust choices

The European discussion on digital sovereignty is often reduced to the origin of a model. For business practice, the term needs to be understood more broadly.

A company can use a European model and still be heavily dependent on individual infrastructure or integration providers. Conversely, the use of an international model can be combined with an architecture that strengthens data control, interchangeability, and operational continuity.

Economically relevant sovereignty is therefore demonstrated through concrete capabilities. Can the company export its data? Can models be exchanged? Are central audit procedures usable regardless of the respective provider? Is there a realistic plan for outages, price changes, or altered contract terms?

Complete interchangeability, however, comes at a price. Those who want to keep every application open to all platforms risk increasing development effort and complexity. Companies must therefore weigh flexibility against efficiency.

Applied AI labs should make this trade-off transparent. Particularly critical processes often justify higher requirements for independence and reliability. For less critical applications, a more standardized platform may be more economical.

The strategic perspective is therefore not technological isolation, but controlled dependence. A company must know where it is committing itself, what benefits this commitment brings, and what alternatives are available in a crisis.

Employees are not just users, but also bearers of crucial knowledge

The introduction of AI doesn't change work simply through new software. It changes who processes information, prepares decisions, and monitors results.

A large portion of the relevant knowledge resides with the employees themselves. They are familiar with typical special cases, informal agreements, and the points where documented processes deviate from reality. If this knowledge is not incorporated, applications are created that appear convincing on paper but are circumvented in everyday practice.

Therefore, training must go beyond simply operating a chat window. Employees should learn to professionally assess results, recognize uncertainty, and understand the appropriate limits of their use. Managers also need the ability to leverage productivity gains organizationally.

The incentive structure is also crucial. If employees get the impression that every disclosed improvement will be used directly against their own position, their willingness to share knowledge decreases. A company can then invest in technology while simultaneously weakening the social foundation for its implementation.

A realistic approach combines performance improvement with comprehensible development prospects. These can include more challenging tasks, better working conditions, reduced workload, or new qualifications. Which solution is appropriate depends on the company. However, without trust and participation, a significant portion of the potential benefits remains untapped.

Small and medium-sized enterprises (SMEs) need shared skills, not isolated prestige projects

For many small and medium-sized enterprises, a dedicated, fully staffed AI lab would not be economically viable. However, this does not mean they have to forgo application-oriented development.

Shared structures can distribute certain fixed costs. Industry-related competence centers, specialized service providers, or business networks can, for example, provide testing procedures, basic technical components, and training programs.

The greatest benefit often arises from recurring problem categories. Several companies can have similar requirements for technical documentation, order processing, or service support without having to pool their confidential trade secrets.

Such collaboration requires clear boundaries. Joint development is not the same as unrestricted data exchange. Competitively relevant information, customer relationships, and company-specific process knowledge must remain protected.

Furthermore, support should be geared towards transferability. A customized, individual project can be helpful for one company, but it won't have a broad impact. Reusable components and easily understandable implementation methods, on the other hand, reduce costs for subsequent users.

This raises an important economic policy perspective: The spread of productive AI does not depend solely on individual flagship companies. It depends on whether companies without large development departments also gain access to reliable implementation expertise.

Europe's competitive advantage can arise from the combination of industry and application

There are significant differences in AI adoption between European countries. In 2025, Denmark (42%), Finland (37.8%), and Sweden (35%) were well above the EU average. While these figures indicate varying levels of adoption, they do not fully explain the underlying reasons. Industry structure, company size, and existing digital infrastructure must be considered when interpreting these results.

Therefore, for European competitiveness, it is not only crucial to further develop individual leading locations. Equally important is the dissemination of skills across industries and regions.

Europe possesses industrial value chains, specialized companies, and extensive technical expertise. However, this will only translate into an AI advantage if this knowledge is translated into usable applications.

An industrial dataset is not yet a business model. A high-performing research institution is not yet a scalable product. A funded demonstrator is not yet a permanently deployed solution.

Applied AI labs can support these transitions if they work closely with real-world business problems and evaluate their results not only technically but also economically. This also shifts the benchmark for successful innovation policy: What matters is not just the number of funded projects, but the number of sustainably productive applications and their demonstrable benefits.

Research, infrastructure and application must not be pitted against each other

A stronger focus on implementation does not mean that research or computing infrastructure are losing importance. Companies can only use what is technically available, accessible, and economically viable.

However, there is a risk of confusing different levels. New computing capacity initially creates a prerequisite. It guarantees neither suitable company data nor altered work processes. Conversely, good application skills cannot arbitrarily compensate for a lack of infrastructure when it comes to demanding tasks.

A European investment report describes both Europe's strong research base and industrial capacity, as well as the difficulties in financing and scaling innovative companies. At the same time, it points to a lag between European companies and the US in the use of big data analytics and AI. The six percentage point difference mentioned in the report refers to this shared indicator and should not be equated with the overall AI adoption rate.

A consistent strategy must therefore combine several levels: scientific progress, high-performance infrastructure, access to capital, and operational implementation.

Applied AI labs occupy the application level. Their success depends on whether they actually translate existing technological possibilities into productive processes. If they are merely established as another institutional showcase, their economic impact will remain limited.

Good management begins with measurable results and clear termination rules

Companies should not approach AI as a collection of as many individual experiments as possible. A more sensible approach is a portfolio that differentiates between short-term improvements, reusable foundations, and more strategically ambitious projects.

Not every project requires the same evaluation criteria. An internal search application can be evaluated based on search duration, response quality, and usage rate. A system for supporting proposals must additionally consider processing time, error rate, and potentially the economic success of the proposals.

A reliable baseline measurement is crucial. Without knowledge of the previous process, it's difficult to later assess whether the AI ​​has actually led to an improvement. Where possible, control groups or clearly defined timeframes are helpful. Different case complexities must be taken into account.

Equally important are rules for termination. A project should be terminated or fundamentally modified if data is unavailable, the costs of errors remain too high, or the benefits do not justify the operating costs. Stopping early is not a failure of innovation, but rather responsible use of capital.

The laboratory itself should also be measured by results. Meaningful metrics include the percentage of applications used continuously, the time to productive use, realized process improvements, and the reuse of developed components. The number of workshops and demonstrations can indicate activity, but it does not replace proof of economic performance.

Sustainable competitive advantage arises from faster corporate learning

A single successful AI project can reduce costs or improve quality. However, a lasting benefit only arises when the company learns more quickly and systematically from each implementation.

This includes reusable test data, documented error patterns, proven integration designs, and a better understanding of which tasks are suitable. With each successful project, the next implementation can be more cost-effective, secure, and faster.

This learning curve is strategically significant. Two companies can have access to the same models and still achieve very different results. One starts from scratch with every application. The other already has data access, evaluation procedures, security standards, and experienced teams.

An effective Applied AI Lab therefore not only builds applications, but also organizational capital. It improves the company's ability to identify technological opportunities, evaluate them economically, and implement them reliably.

This is precisely where the sound perspective for European companies lies: AI success is determined neither solely by the origin of a model nor by the size of an investment program. What matters is whether technology, expertise, data, and responsibility come together to form a productive unit. Applied AI labs can be a powerful tool for this. However, they only prove their value where they lead to improved processes, reliable performance, and economically viable results.

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