Many of the predictions have materialised in the form of specific products, platforms and pilot projects. Other topics, however, were characterised more by hype, buzzwords or regulatory debates. In this review, we analyse our predictions, compare them with actual developments over the course of the year and place the findings in a broader context: what does this mean for businesses, for technology teams and for the next steps in 2026?
Agent systems & multi-agent systems
In our annual outlook, we had highlighted modern agent systems and multi-agent systems as one of the key strategic trends. In our view, AI agents would begin to take on real tasks over the course of the year: independently gathering information, making decisions and carrying out process steps. Looking back, this forecast was essentially correct – albeit with some significant nuances.
What was actually achieved in 2025
In 2025, agentic AI made the leap from theoretical concepts and demo videos into corporate practice – albeit mostly in the form of clearly defined, well-structured use cases. According to recent market studies and analyst reports, many companies are now relying on so-called task agents or workflow agents, which automatically carry out defined steps. More complex agent-based architectures that coordinate multiple specialised agents are rare, but there are initial productive pilot projects, particularly in customer service and back-office automation.
Hype vs. Reality
The grand vision – fully autonomous agents that plan flexibly, learn and act within an open corporate context – will not yet be a reality by 2025. Some of the initial expectations were put into perspective over the course of the year; numerous experts now openly acknowledge that ‘agent-based AI’ was also used as a marketing term. At the same time, however, it is becoming clear that agent-based systems can create genuine added value when the technical and organisational prerequisites are in place.
Classification using the Dextralabs Maturity Model
The Agentic AI Maturity Model published by Dextralabs in November 2025 is particularly helpful for classification. It distinguishes between four levels – from simple task agents (L1) through to coordinated multi-agent systems (L2) and autonomous orchestration layers (L3), right up to self-learning agent ecosystems (L4). The majority of companies currently fall between Levels 1 and 2. A handful of pioneers are working on Level 3 architectures, whilst Level 4 – despite much discussion – clearly remains a prospect for the future.
Conclusion
Our forecast has thus largely been confirmed: by 2025, agent systems have become one of the most visible developments in the field of AI, albeit with a significantly more pragmatic level of maturity than the early visions would have suggested. Agentic AI is a reality – but is still rarely truly autonomous.
Multimodal conversational interfaces
In our annual outlook, we had anticipated that multimodal AI interaction would gain in importance by 2025 and that the traditional text or voice interface would increasingly be supplemented by richer forms of dialogue. The vision: systems that simultaneously understand and combine speech, text, images or even gestures to interact with users in a more natural and intuitive way.
What was actually achieved in 2025
Multimodality has clearly gained momentum this year. Large AI models such as Google Gemini or OpenAI systems now routinely process text, images and audio within a single architecture. Progress has also been made in research and product development: from multimodal agents and visually supported AI assistants to initial experiments with non-verbal signals such as facial expressions and gestures. This has essentially confirmed the trend – multimodal interaction is no longer a niche topic, but an integral part of many AI roadmaps.
However, what is not yet a widespread reality
Despite technological maturity at its core, genuine, fully integrated multimodal conversational interfaces remain the exception in practice. So far, companies have primarily deployed multimodality where it delivers clear, immediate benefits – for example, in the form of voice control with accompanying visual output, or through the combination of chat and image analysis in customer support. Holistic interaction across multiple modalities, as is often seen in demonstrations, remains largely at the pilot stage and is rare outside of major platform providers.
Assessment
Our forecast has thus essentially come true: interaction between humans and AI will become more varied, natural and versatile by 2025. At the same time, it is becoming apparent that true multimodality will not become the norm overnight. The industry is moving forward step by step – and multimodality is emerging more as a targeted feature than as a completely new generation of interfaces.
Emotionally intelligent AI
In our annual outlook, we had expected that by 2025, AI systems would respond better to moods and emotional signals, making interactions seem more natural and human. Looking back, it is clear that the emotional component of AI has continued to develop – albeit in a far more measured and cautious manner than some early visions had suggested.
What was actually achieved in 2025
Several AI models were stylistically adapted over the course of the year to sound friendlier, more supportive and ‘more human-like’ in conversational situations. This was particularly evident in models such as GPT-5.1, which received an update in November featuring a more conversational and warmer tone. Progress was also made in research and pilot projects regarding the analysis of sentiment indicators such as tone of voice or text meaning, and these can certainly be helpful in stable contexts.
What remains unresolved
The complex task of reliably recognising genuine emotions – across voice, facial expressions or body language – remains an open research problem. Many methods only work in controlled environments or produce unreliable results, particularly across cultural or linguistic boundaries. What is perceived as ‘empathetic’ in everyday life is often more a matter of stylistic optimisation than deeper affective intelligence. AI systems are still a long way from genuine emotional sensitivity in the sense of human perception.
Regulatory constraints
Furthermore, the EU AI Act imposes strict guidelines on this issue. The use of emotion recognition in the workplace, in schools or in public spaces is prohibited, and even in less sensitive areas, high standards of transparency and non-discrimination apply. These regulatory frameworks have put the brakes on many ambitious implementation plans and ensured that companies focus primarily on risk-free, clearly defined scenarios.
Assessment
The overall picture is therefore mixed: emotional intelligence in AI interactions has developed noticeably by 2025 – though primarily at the level of expression and conversation. The more far-reaching technological vision of an AI that precisely recognises human emotions and responds appropriately remains challenging and is heavily constrained by regulation. The trend is real, but its practical implementation has so far remained cautious, selective and clearly context-dependent.
Click here to continue to the second part.
