AI is rapidly becoming an invisible infrastructure that influences decision-making, cost structures, and the distribution of power. In 2023, many organizations were still asking what AI is capable of. Now, the more relevant question is who decides how it operates – and on whose behalf. 

At Netprofile, we took a moment to discuss what’s currently shaping the AI landscape, from the slowing performance race and evolving pricing models to broader questions of power, responsibility and dependency. 

Trust is built on ownership and transparency 

In a geopolitical climate that is, to say the least, rather peculiar, trust and preparedness for risks are becoming increasingly valuable. At the same time, AI is finding its way into decision-making, often without its role being clearly articulated. 

The question is no longer whether we trust the answer produced by an LLM, but who controls it, trains it, and uses it. 

The EU AI Act seeks to limit this power in use cases where the impact on people’s rights is significant. These include, for example, recruitment systems that screen applicants, as well as credit and financial systems that assess customers’ solvency and risk profiles. However, regulation does not resolve the fundamental question of who holds real power when AI is used to support or form part of decision-making processes. 

Different forms of dependency go hand in hand with trust. Many companies are considering where they want to store their data and whose tools they should rely on in the future. From a scientific perspective, it is also notable that nearly all significant language models today originate from commercial actors rather than research institutions or governments. 

AI’s continued tendency to hallucinate challenges trust in the accuracy of information. The Reuters Institute has identified Google’s AI Overview search results as a major structural threat to media, as answers provided directly in the search interface reduce the need for users to click through to original articles – despite estimates that the feature generates hundreds of millions of incorrect or misleading responses each day. 

The performance gap has disappeared 

By 2026, language model performance is no longer a sufficient differentiator. New players, models, and tools are emerging rapidly. Companies are increasingly selecting the model that best fits the use case, regardless of who developed it. 

The differences between frontier models have narrowed significantly, and GPT-4-level performance is now sufficient for many basic tasks at a fraction of the previous cost. Open source challengers are already operating at a level that meets the needs of many users.

As a result, companies are focusing on providers that can deliver reliable AI cost-effectively and models that perform well within their specific industry. At the same time, orchestration, meaning information retrieval, context management, and continuity across sessions, has become more critical than the raw performance of any single model. 

Agentic AI is everywhere – along with the risks 

AI is now rapidly being developed into agentic forms capable of modifying files, executing tasks, and making complex decisions. 

Estimates suggest that a vast majority of enterprise applications will include agentic AI during the current year. The potential productivity gains from agents connected to business-critical systems and data are compelling, but so are the risks. 

More concerning than hallucinated outputs are agents that perform the wrong measures. Unauthorized actions and breaches of operational boundaries can lead to irreversible consequences in real-world contexts and business environments. For example, Meta avoided serious impact from a security breach caused by an unauthorized agent largely by chance. 

This raises a key question: who bears responsibility for the damage caused by an agent? AI governance is now attracting attention even from those who previously saw it as slow-moving bureaucracy. 

In development, hype meets reality 

Almost all software developers now use AI as part of their workflows. Code is produced faster than ever, but reviewing it, fixing errors, and maintaining overall control takes up increasing amounts of time. 

At the same time, humans become the bottleneck in what is known as the productivity paradox: while individual developers may become more efficient thanks to AI, delivery capacity does not scale at the same rate. Benefits remain limited unless ways of working, responsibilities, and quality criteria are reorganised. 

Simultaneously, low-code and no-code solutions, along with “vibe coding”, are making software development accessible to a broader audience. Prototyping is moving earlier in the process, and developers are no longer required at every stage. This creates significant opportunities but also increases the need for architectural thinking. 

Pricing reveals how AI is actually used 

Agentic AI has disrupted traditional pricing models. While chat-based use consumed hundreds of tokens, an agent may consume millions within a single task. Fixed-price subscriptions simply cannot sustain this – especially when leading AI providers are still operating with significant losses. 

Pricing models are shifting increasingly towards usage-based billing, which creates challenges when cost predictability is required. This development pushes organisations towards more deliberate use of AI, as model selection, orchestration, and use cases are directly reflected in budgets. Heavy token usage needs to be justified in financial terms. 

In 2026, organisations are no longer pursuing AI benefits blindly, but are moving towards more controlled, cost-efficient, and responsible adoption. 

Would you like to hear more about the latest developments in technology and communications? Download our report Insight Track 2026!