AI trends 2025: Always on the move
inpeek Blog
Fabian Hüni
The world of artificial intelligence is developing rapidly – anyone who wants to stay on the ball should keep an eye on the most important developments. New tools, models and protocols appear almost daily. This article shows in a compact and practical way which AI trends will be particularly relevant in 2025 according to our assessment and observations. And how you can prepare for them and benefit from them.
Agentic AI – AI that really acts
From reactive chatbots to active decision-makers: agentic AI is changing the rules of the game and is undoubtedly currently the biggest trend in the world of AI. For the first time, it will be possible not only to analyze or support business processes, but also to control them autonomously – going beyond purely deterministic methods. To this end, the LLMs are equipped with tools that perform individual actions such as making a calculation, calling up the current weather or triggering the sending of an invoice in the ERP system. The LLMs interpret data and decide independently which tools to call up in order to obtain the necessary information and carry out actions.
What are agents? Agents are AI systems that:
- make independent decisions,
- actively interact with systems and
- intelligently combine several actions to achieve a goal.
Practical example: An intelligent agent in customer support understands a ticket, triggers an action in the ERP system (e.g. sending an invoice) and informs the customer – completely autonomously.
Important Developments:
- Defines a standard for interaction with peripheral systems (e.g. ERP, weather API) through LLMs
- Agent2Agent Protocol: Defines the collaboration between two AI agents
- Assistants: e.g. Microsoft Copilot, SAP Joule
Forecast: Agentic AI will deeply automate processes and raise human supervision to a new level. The first use cases are currently being evaluated and will go into production in 2025. For companies, this means great potential to save time and therefore costs. Major providers such as Microsoft and SAP are launching agents (e.g. SAP Joule Agents) that are integrated into their tools and are intended to change or accelerate business processes in the long term. For in-house developments, the question will increasingly arise as to whether and how AI agents can interact with the software. The MCP (Model Context Protocol) will open up new possibilities for this. And if custom agents are built, they can be integrated into existing copilots such as SAP Joule or Microsoft Copilot using the Agent2Agent protocol so that users can continue to use the established chatbots.
Inference Time Compute – More time, better answers
Artificial intelligence is becoming more powerful, not only through larger models, but also through more intelligent “thinking”. Traditionally, models have been optimized by providing them with more data, more parameters or more training resources (GPU/TPU). Now a new approach is coming to the fore: more computing time during answer generation (“inference time”).
What does this mean?
Models are deliberately given more “thinking time”. Several thinking and revision steps improve the quality of the answers, especially in more complex tasks where logical thinking is required, such as in mathematics. This improves performance, but also increases costs and increases the time until an answer is available (latency). An example: The deep research agents of the major providers are not so strong because of other training, but rather because they are given several minutes or even hours to evaluate different sources and then use the best ones for a final answer.
Recommendation: Companies benefit from the improved response quality due to the inference time compute. However, this also increases costs, and the use of more “thinking time” is not always necessary. Performance and costs should always be weighed up carefully. Deep research agents make it possible to create reports such as a market analysis or a strategy paper in a much shorter time. Of course, such analyses should not be trusted blindly, but the results provide an initial sound basis for further elaboration.
Very large and very small models – two extremes are very popular
The future does not only belong to the giants: Small, specialized models offer new opportunities. While extremely large models continue to be developed, so-called smolLLMs (for text) and smolVLMs (for vision) are becoming increasingly established. These models deliberately focus on fewer parameters and can therefore be operated directly on smartphones, laptops or in the browser on-device, depending on their size.
Advantages of small models:
- Sufficient performance for many tasks (e.g. object recognition in an image, summarizing a text)
- Can be used directly on laptops, smartphones or in the browser
- No inference servers required
- Data protection-friendly: data remains local
New trends resulting from this:
- WebML: AI applications in the browser
- On-device AI: Mobile AI apps that can also be used offline
Insight: Small AI models enable flexible, secure and data protection-friendly applications and are therefore ideal for new generations of apps. This development brings AI directly into the browser or into a mobile app without the data ever leaving the device or requiring an internet connection. We consider the possibilities of such applications to be enormous, but assume that it will be several months before the first implementations in this area take place – partly because the focus is currently more on AI agents. Nevertheless, we see great potential here!
Human-software interaction – interfaces are becoming (almost) superfluous
The way we interact with software will change fundamentally. Large language models (LLMs) are opening up a new era of software operation: in future, natural language will increasingly replace classic text-based user interfaces (UIs).
What does this mean for companies?
- Applications should be developed at an early stage in such a way that they can also be used by LLMs.
- At the same time, systems are being developed that autonomously operate desktop or web applications via the existing UI. This is done on the basis of vision LLMs that interpret screenshots and then predict and execute the next action (e.g. mouse click).
Practical example: In our real estate portal Kibanda, users could in future simply reserve rooms via chat or voice command – instead of navigating through menus with clicks. This does not involve creating new functionalities, but rather making the existing APIs “AI-ready” so that LLMs can interact with the software.
Recommendation: Rethink interfaces and build your applications with AI integration and voice interaction in mind. In the same way that search engine optimization (SEO) has been used for websites up to now, we assume that applications and websites will also be optimized for AI systems in the future. We expect this to take place on the basis of the MCP and Agent2Agent protocol.
Data protection, regulations and ethical AI – trust becomes a competitive advantage
Technological sovereignty, data protection and ethics are moving to the center of the AI debate. In view of the dominance of US providers, data protection, compliance and transparency are becoming increasingly important for European companies.
Significant developments:
- Higher requirements for data processing (location, protection, control)
- Proprietary models vs. open source solutions
- Stricter regulations in the EU and Switzerland
- Ethical AI becomes mandatory: avoiding discrimination and bias
Insight: Trust is no longer a “nice-to-have”, but a key differentiator when using AI. Companies will have to ask themselves questions such as: “Where is the data located and which AI API provider do we trust? Do we rely on proprietary models (e.g. from OpenAI/ChatGPT) or do we focus on open source solutions?” This should be based on clear strategies.
Our conclusion: 2025 will be the year of AI agents
The trends presented here provide an initial insight into which topics we should all continue to keep a close eye on in 2025. Even if “AI Agents” lends itself well to buzzword bingo, this trend is likely to take hold and keep companies very busy this year. Many of the other trends are also indirectly or directly linked to AI agents.
Our assessment:
- Agentic AI will fundamentally change business processes.
- Repetitive tasks will increasingly be taken over by AI agents.
- Human-software interaction will become more natural and intuitive.
- In software development, the question “How does AI interact with our software?” will arise anew.
- Governance and ethics will be decisive success factors.
LEARN MORE
At inpeek, we deal with the latest technologies on a daily basis and support companies in the integration of AI solutions. Would you like to evaluate initial use cases? Prepare your applications for the future of interaction? Or adapt your data strategy to new requirements? Get in touch with us!