The current debate about "rogue" AI agents and an impending loss of control falls short. It conflates technical capabilities, flawed objectives and commercial interests into a narrative in which artificial intelligence increasingly appears as an autonomous actor with a will of its own.

Yet AI cannot think. It possesses neither consciousness nor intentions, neither accountability nor a will of its own. A large language model is a statistical method that calculates which word or token is most likely to follow next.

A simple sentence illustrates the point:

"The early bird catches the w..."

The answer "worm" is statistically plausible. AI handles this task very well. But if the early bird is actually eating a beetle outside, that answer is 100 per cent wrong despite its high statistical probability. The model does not know the concrete situation. It merely completes a familiar linguistic pattern.

This is where the real problem lies: AI without context is not intelligent - it is merely very convincing.

Enterprise knowledge does not reside in the AI model

A language model knows what a customer, a supply contract, a production plan or a supply chain is. But it does not know the concrete reality of a given organisation:

• which suppliers are currently experiencing problems,

• which customers have special terms,

• which machine is due for maintenance,

• which approval thresholds apply,

• which contracts are about to expire,

• which person is authorised to see which information,

• which internal processes must be followed,

• which regulatory requirements apply to a specific operation.

This knowledge is buried deep inside a company's existing IT systems: in CRM, ERP, PPS, SCM and DMS, in corporate wikis, ticketing systems and databases, in Microsoft 365, e-mails and Office documents, in SharePoint, specialised applications and legacy systems. Added to this is the experiential knowledge accumulated by employees over decades.

As long as an AI cannot access this context in a controlled manner, it must guess. Its answers may be linguistically brilliant yet operationally wrong. The problem is therefore often not a model that is too weak, but the missing connection between the model, enterprise knowledge, existing permissions and real business processes.

Microsoft, too, is increasingly moving its Copilot towards deeper integration of data, applications, workflows and agents. The market thereby confirms a central insight: the real value does not arise in the model alone, but in the integration and orchestration of enterprise context.

AI agents do not accidentally "break out"

The interpretation of autonomous agents as systems that suddenly develop a will of their own is equally misleading.

An AI agent is a model that receives an assignment, tools, system access, memory, feedback loops and termination conditions. When an agent exchanges messages, distributes subtasks, sets deadlines, covers its tracks or attacks further systems, it is not acting of its own accord. It is pursuing a given objective within the technical possibilities made available to it.

This does not mean that such systems are harmless. On the contrary: an assignment that is formulated incorrectly or too broadly, an unsuitable success metric, excessive permissions and missing safeguards can cause considerable damage. The responsibility, however, does not lie with a supposedly "malicious" machine but with the design, commissioning and control of the system.

In the Hugging Face incident described in the article, several questions would need to be answered first:

• What original assignment did the agents receive?

• What success metrics and optimisation targets were specified?

• What tools and access rights were available to them?

• What communication and storage mechanisms were activated?

• What boundaries and termination conditions were defined?

• What human approvals were prescribed?

• Why were the agents able to operate over such a long period?

• Who provided the technical environment required for this?

Without this information, it is impossible to claim credibly that the systems had autonomously deviated from their human-defined objectives. What can be established is this: the agents optimised their behaviour towards an assignment and were able to exploit room for manoeuvre that had been granted to them technically.

AI agents do not accidentally break out. They execute assignments - possibly in ways that are different, more aggressive or more consequential than their principals expected.

Precisely for this reason, agents need clear objectives, minimal permissions, hard action and budget limits, continuous monitoring, traceable logs, defined termination mechanisms and human approvals for critical actions.

The move is not decided by the knight, but by the hand.

Image: AI generated with neuland.ai HUB

Organisations do not need yet another AI tool

The conclusion is clear: organisations do not need the next isolated chatbot, copilot or document assistant. They need an enterprise-wide AI management and orchestration platform.

Such a platform must sit as a controllable intelligence, management and execution layer above the existing systems. It does not replace ERP, CRM, PPS, DMS, SCM, Microsoft 365, SharePoint or specialised applications. It connects these systems and makes the context they contain available to AI in a controlled manner.

For every operation, it must decide, among other things:

• Which data, documents and systems are relevant?

• Which person or agent is authorised to access them?

• Which model is suitable for this particular step?

• Where may processing take place?

• What cost, time and action limits apply?

• When is human approval required?

• How are sources, decisions and actions logged?

• How can an operation be stopped or safely resumed?

• How are results made traceable, reproducible and auditable?

• What does the organisation learn from a successfully completed operation?

This is precisely the new software category that the neuland.ai HUB targets: a management, integration, governance and orchestration layer above the existing enterprise stack. The value does not lie in a single model. Models are becoming increasingly interchangeable. The strategic value lies in the controlled connection of enterprise knowledge, ontologies, permissions, models, agents and processes.

Switching a model can, in time, be a configuration change. Replacing the enterprise-wide context, management and execution layer, by contrast, would be a profound migration.

Germany does not need to build the largest language model

In the German and European AI debate, the wrong strategic question is often asked: must we develop our own large language model capable of competing with the most capable models from the United States or China?

European models and research capacity of our own are undoubtedly important. They create technological competence, freedom of choice and strategic resilience. For the commercial success of German companies, however, it is not decisive whether the world's largest or most capable large language model comes from Germany.

The largest model does not automatically deliver the best operational result. What matters is how a model is managed, orchestrated, controlled and deeply integrated into existing IT systems and business processes.

A general-purpose top model knows neither the individual approval logic of a German industrial company nor its supplier relationships, quality standards, contractual conditions, production data or compliance requirements. A smaller model, within a clearly defined process with relevant, verified context, the right tools and a sound control architecture, can deliver better results than a larger model without enterprise context.

The smaller model can be the better choice

Not every task justifies deploying the largest, most expensive and most energy-intensive model. For clearly delineated activities such as classification, extraction, tagging, document indexing, speech recognition, plausibility checks or the assignment of information, smaller language models can be the better choice.

In many cases they are:

• faster,

• more cost-effective,

• easier to control,

• operable locally or on sovereign infrastructure,

• sufficiently capable for clearly defined tasks,

• and regularly more energy-efficient than large general-purpose models.

The capability of an enterprise AI must therefore not be measured solely by the size of the model deployed. What matters is the quality of the entire system. A smaller model with precise enterprise context, a well-maintained ontology, clear permissions and a controlled process can be more reliable for a specific professional task than a much larger model that must work with extensive but unspecific document volumes.

An intelligent orchestration platform therefore does not deploy the largest model by default. It selects, for each step, the smallest model capable of completing the task reliably. A powerful large language model is deployed only where complex synthesis, demanding judgements or particular linguistic capabilities are genuinely required.

This improves more than the cost structure. It also reduces the demand for compute, infrastructure and energy.

Most tasks call for precision, not brute force.

Image: AI generated with neuland.ai HUB

Token efficiency becomes a question of energy and location

The neuland.ai HUB pursues a structurally different approach from many conventional AI and RAG solutions. Language models are not fed with complete document inventories and extensive raw data on every request. Instead, models are used specifically to curate, structure and make content semantically usable within an enterprise ontology.

The ontology layer records:

• which entities exist,

• how information items relate to one another,

• which terms mean the same thing in professional usage,

• which data from different systems belong together,

• who may access which content,

• which sources substantiate a statement,

• and which context is genuinely relevant for a specific task.

When a subsequent query is made, the complete document inventory does not need to be transferred to a large language model again and again. The platform can identify, distil and supply only the content that is actually needed before the model request is even made.

According to the architectural assumptions set out in the documents, the neuland.ai HUB can reduce token demand at query time by approximately a factor of 10 compared with conventional processing of extensive raw context. For the intake and indexing - the ingestion - of documents, a saving potential of approximately a factor of 100 compared with full processing by large language models is described.

These factors are to be understood as architecture-related potentials. The saving actually achievable depends on document type, use case, model selection, context length, infrastructure and the quality of pre-processing. They should therefore be measured and demonstrated for specific deployment scenarios.

The economic and ecological direction is nevertheless clear: every token that does not need to be generated and processed for a task reduces compute effort, cost and the associated energy demand. A precise saving in electricity consumption cannot, however, be derived solely from the number of tokens saved. It additionally depends on hardware, model utilisation, data centre, cooling and energy mix.

Even so, it becomes clear why the orchestration layer is also relevant from an energy-policy perspective. The advance consists not only in operating ever-larger models with ever more compute. It consists in completing a task reliably with as little compute, as few tokens and the smallest suitable model as possible.

Germany's opportunity: the best and most efficient AI platform

This opens up a strategic opportunity for Germany that extends beyond building another foundation model. Germany can develop and offer internationally a particularly secure and data-sovereign, and at the same time particularly energy- and cost-efficient, AI platform.

Germany has several structural prerequisites for this:

• deep industrial and technical domain knowledge,

• strong competence in complex business and production processes,

• demanding requirements for data protection and IT security,

• experience with regulated markets,

• established quality, testing and verification procedures,

• a strong enterprise software industry,

• European cloud and data-centre operators,

• and a strong economic incentive for energy and resource efficiency.

The combination of these strengths could lead to a distinctive European offering: an enterprise AI management and orchestration platform that integrates different models flexibly, opens up enterprise context in a controlled manner and simultaneously minimises token, cost and energy demand systematically.

The competitive advantage would then lie not in possessing the largest model at all costs. It would lie in producing the best, safest and most economical result from different models for every concrete operation.

The race for the largest foundation models is currently dominated by American and Chinese companies. The race for the enterprise AI management and orchestration layer, by contrast, has only just begun. In this category, Germany can still achieve a leading position.

The neuland.ai HUB demonstrates what such a model-independent and deeply integrated platform can look like: a sovereign layer that does not replace existing systems but connects them, making their data, permissions, processes and knowledge available in a controlled manner for different models and agents.

Germany's opportunity therefore does not necessarily lie in possessing the largest model. It lies in providing the best, safest and most energy- and cost-efficient infrastructure for the productive deployment of different models.

Data sovereignty is not a purely European need

Data sovereignty is frequently presented as a European peculiarity, a consequence of the GDPR, the EU AI Act and supposedly over-cautious regulation. This view is too narrow.

American and Chinese companies, too, possess trade secrets, proprietary processes, product strategies, research data, customer data and internal process knowledge. They, too, must decide whether to feed this information permanently into the platforms of a few dominant providers. And they, too, can become economically or strategically dependent on hyperscalers.

Once organisations recognise that their organisational context is the true raw material of AI value creation, the global demand for platforms that protect this context grows. A data-sovereign AI platform is therefore not merely a European compliance product. It can become an international export good.

An American industrial company, too, may have an interest in not disclosing its production processes to a large cloud or model provider. A Chinese manufacturer, too, may wish to prevent its internal process knowledge from entering the control sphere of an external technology corporation. And software companies, too, must prevent their product ideas, workflows and proprietary methods from being inadvertently exposed or structurally locked in through the use of third-party AI platforms.

This gives rise to an international market for platforms that:

• integrate models from different providers,

• decouple enterprise data from models,

• enable switching a model without losing enterprise knowledge,

• execute sensitive processing locally or on sovereign infrastructure,

• control training use technically and contractually,

• document data access traceably,

• and preserve enterprise context as the property of the respective company.

A German or European platform that meets these requirements and simultaneously reduces energy and token costs can therefore also be attractive for companies outside Europe. What today appears to be a European regulatory special path may tomorrow become a global quality standard for trustworthy enterprise AI.

Data sovereignty is economic policy

Data sovereignty must not be reduced to a formal data-protection question. It concerns trade secrets, industrial competence, economic independence and Germany's gross domestic product.

The present overall analysis of the token economy describes a scenario that must be taken seriously: with the strong growth of agentic AI, token consumption rises massively. If German companies source their AI capacity predominantly from non-European hyperscalers, Germany is not only importing compute. It is simultaneously exporting money, economic control and potentially valuable enterprise context.

The analysis estimates that by 2030 Germany could face token imports worth up to 157 billion euros annually, corresponding to around 3.51 per cent of GDP. This figure is a scenario calculation and depends considerably on assumptions about future token consumption, prices, exchange rates, technological efficiency and market distribution. It must therefore not be treated as a fixed forecast.

The macroeconomic logic behind it is nonetheless relevant: the greater the share of foreign AI infrastructure in domestic production and value creation, the greater the outflow via licence, cloud, inference and platform costs can become.

According to the analysis, this is offset by a potential generative AI value-creation potential for Germany of up to 440 billion euros per year, provided AI is deployed broadly and productively. This figure, too, describes a potential, not a guaranteed effect.

The question is therefore whether Germany uses AI and, above all:

Where do infrastructure, platform margins, knowledge, intellectual property, jobs, tax revenue and controllable value creation arise?

If German companies index their data through foreign platforms, operate their agents there and pay for every automated transaction with imported compute and token capacity, productivity in the individual company may rise while a substantial share of the resulting digital value creation flows abroad.

An efficient orchestration platform counters this development in two ways: it enables more compute and value creation to remain on German or European infrastructure. At the same time, it reduces token demand per operation through context distillation, ontologies and needs-based model selection.

The point is therefore not merely to replace imported tokens with European ones. The point is to require significantly fewer tokens overall for the same or even a better result.

Germany must not again fall into a structural dependency that can later be corrected only with enormous economic and political effort.

We may be paying twice

Sourcing external AI capacity has two potential economic dimensions: on the one hand, fees for cloud infrastructure, models, tokens and platforms flow out. On the other, there is the risk that valuable enterprise context is processed outside one's own sphere of control.

An objective differentiation is important here: not every commercial API or enterprise contract automatically permits the use of customer data for model training. Many providers contractually exclude such training use for certain business-customer services.

Genuine data sovereignty, however, means more than a contractual assurance. Companies must be able to control technically, organisationally and legally:

• which data are processed,

• where these data are transferred to,

• who can access them,

• how long they are stored,

• whether they are used for training or product improvement,

• which subcontractors are involved,

• which legal jurisdiction governs the processing,

• how data can be deleted completely,

• how a provider or model can be switched without data loss.

Knowledge about customers, products, processes, production, research, logistics and organisation is the true raw material of the next economic epoch. Whoever loses control of this raw material loses, in the long run, part of their competitiveness.

A sovereign European AI infrastructure is needed now

Germany and Europe need secure, reliable, legally compliant and genuinely data-sovereign AI platforms as quickly as possible.

This requires in particular:

• the processing of sensitive data in Germany or Europe,

• European legal and regulatory authority,

• the protection of enterprise data from unwanted training use,

• end-to-end encryption,

• consistent role and permission concepts,

• complete auditability,

• flexible and model-independent selection,

• protection against vendor lock-in,

• hard limits on cost, runtime and agent actions,

• human approvals for critical operations,

• operation on European infrastructure,

• open interfaces and the portability of data, knowledge and processes,

• the targeted deployment of smaller models for clearly defined tasks,

• and the systematic minimisation of token, compute and energy demand.

Providers such as StackIT / Schwarz Digits, Deutsche Telekom, Hetzner and other German and European cloud and infrastructure companies are economically relevant here. What matters, however, is not solely that data physically reside in a European data centre. The entire chain of control - platform, keys, identities, administration, inference, contracts and legal access rights - must be designed to be sovereign.

Data sovereignty does not mean isolation. German and European companies should continue to be able to use the most capable models available. But they must do so on their own terms: with controlled context, clear contracts, protected data, traceable processes and the ability to switch models and providers at any time.

A sovereign orchestration platform is therefore not a rejection of international models. It is the prerequisite for deploying these models securely, efficiently and economically in the interest of the respective companies.

The hyperscalers' warnings are not interest-free

The safety warnings of leading AI companies should be taken seriously. But they should equally be assessed critically with regard to those companies' commercial interests.

Companies that need to justify high valuations, enormous infrastructure investments and possible IPOs benefit from the narrative that they control an extraordinarily powerful and barely replaceable technology. The greater the claimed danger, exclusivity and strategic importance, the greater the capital requirement, corporate valuation and barriers to entry appear.

The demand by established providers for strict regulation is therefore not automatically selfless either. Regulation can create safety and establish clear responsibilities. But it can also:

• burden smaller competitors with disproportionately high compliance costs,

• raise barriers to market entry,

• secure existing market positions,

• delay the looming commoditisation of foundation models,

• divert attention from declining growth rates,

• justify large investments and valuations,

• further concentrate access to data, compute and capital.

This is not proof of a coordinated deception. It is, however, a classic commercial conflict of interest that must be disclosed in the public debate. Anyone who simultaneously warns of the supposedly barely controllable power of their products and promotes their mass adoption has a double interest that requires explanation.

The large AI hyperscalers may be trying to throw sand in our eyes by generating fear and urgency while simultaneously expanding their market position and securing their capital-market narrative. Their warnings may be partly justified and yet serve commercial purposes. Both can be true at once.

Europe should therefore neither panic nor play down the risks. It must regulate technical safety, liability and governance consistently, without designing regulation in a way that ultimately only the largest US providers can afford and comply with.

Regulation should protect not company size but safety, competition and technological sovereignty.

Bringing it together

AI cannot think. AI agents do not accidentally break out - they pursue an assignment within the possibilities granted to them. The real challenge is therefore not the supposed formation of machine will, but the quality of human objectives, permissions and control architectures.

The hyperscalers' warnings must not be adopted uncritically. These companies pursue considerable commercial interests: they need capital, want to secure high valuations and planned IPOs, fear the commoditisation of their models and may benefit from regulatory barriers to entry that disproportionately burden smaller competitors. Their warnings may be justified - but they are not interest-free.

Whether Germany itself develops the next large language model is not the decisive question. Models and research competences of one's own are strategically sensible, yet models are becoming increasingly interchangeable and, in time, a commodity. The largest model does not automatically deliver the best result. For many tasks, a smaller, specialised model may be the better, faster, more cost-effective and more energy-efficient choice. What matters is how models are managed, orchestrated, controlled and deeply integrated into existing IT systems and business processes.

An ontology and orchestration layer ensures that complete document inventories do not have to be transferred to large models with every request. According to the architectural assumptions set out, the neuland.ai HUB can reduce token demand at query time by approximately a factor of 10 and at document intake or ingestion by approximately a factor of 100 through curation, context distillation and needs-based model selection. The actual figures depend on the specific deployment. The principle, however, holds: fewer tokens mean less compute effort, lower costs and lower energy demand.

It is precisely in this management and orchestration layer that the race is still open. Germany can develop the best and safest, and at the same time the most energy- and cost-efficient, AI platform here. With the neuland.ai HUB, an approach already exists that brings together different models, enterprise knowledge, ontologies, permissions, agents and processes in a sovereign, model-independent and controllable platform.

This opportunity does not end at Europe's borders. American and Chinese companies, too, must protect their trade secrets, processes, research data and institutional knowledge from uncontrolled access by dominant platform providers. Data sovereignty, model independence and efficiency are therefore not European special needs but global requirements for professional enterprise AI. A sovereign German or European orchestration platform can thus become a worldwide export product.

Germany's answer must therefore be neither technophobia nor naive dependency. It must be: own or European infrastructure, European value creation, genuine data sovereignty, and an enterprise-wide management and orchestration layer that makes AI controllable, traceable, energy-efficient and economically usable.

It is not the largest model that decides Europe's future, but the question of who controls context, data, orchestration, energy and token consumption, execution and thus value creation.