What a year for AI. In 2024, the Generative AI hype has turned into a full-fledged gold rush in which everyone and their dog are trying to capitalize on the new technology and its business applications, while users everywhere are being offered new, exciting ways to automate tasks and streamline workflows. RAGs, agents, multimodal LLMs, you name it.
We are moving from the hectic phases of initial experimentation to the specialization of technological maturity.
Time will tell, but we are probably witnessing the beginning of a new industrial revolution, propelled by foundation models that will redefine the way we work, while impacting everyday life in unfathomable ways.
Yet, “traditional” AI is not gone. Companies are still relying on Machine learning, Computer Vision and all technologies that can be subsumed under the labels of analytical or predictive AI. Moreover, we are today witnessing the rise of hybrid applications that combine the flexibility and analytical capabilities of ad-hoc ML models with generative techniques and natural language interfaces, opening up AI applications to non-technical users.
Accountability is also at stake: sophisticated observability and explainability capabilities are making AI systems increasingly responsible, transparent, and compliant with the growing regulations. At the same time, companies are more and more invested in the environmental and financial sustainability of their AI-powered use cases. The growing hunger for computational resources, in both cloud and on-prem systems, has prompted a new consciousness about the impact of AI and the need for scalable, optimized solutions.
The AI situation is excellent – yet there’s great chaos under heaven. New fads are born and buried in the matter of weeks. Tools and models, both open source and enterprise, are released on a daily basis. Moguls come up with new buzzwords and business applications every other day. Here at Radicalbit, we leveraged our expertise and hands-on experience to analyze the top AI trends and single out the 5 most relevant that will shape the landscape in 2025.
1. Agentic AI
You’ve seen LLMs and RAGs, now get ready for AI Agents. Hailed by Gartner as n°1 Top Strategic Technology Trend, agents are poised to revolutionize the operational effectiveness and autonomy of AI applications. We are talking about software tools that leverage foundation models – such as LLMs – to perform complex tasks independently, iterate based on external feedback, and adapt to protean scenarios.
Agentivity and multiplicity are the most exciting distinctive features of this technology. Agentic AI applications go beyond the content generation capabilities of LLMs and RAGs to perform actions on their surrounding environments. At the same time, they can handle non-linear workflows and complex scenarios, devising solutions that are simply unobtainable for rule-based automation systems. All this means that AI Agents can manage different inputs and outputs, draw up different plans, search the web, cooperate with other agents and applications, ask for user feedback and finally take action.
Establishing a productive relationship with the environment, AI agents can be used for a wide range of use cases. For example, they may revolutionize customer support, with intelligent chatbots that really understand users’ requests and act accordingly: “Can you retrieve and email me the last 5 monthly bank statements?”. Agentic AI may also be employed in media planning for marketing campaigns, performing multivariate analysis of historical data and market trends to come up with proposals in line with the strategic goals. And these are just two examples of the realms in which agents can be employed: cybersecurity, coding, business intelligence, you name it. As the newest sensation in AI, agents are definitely here to stay.
2. Vertical AI
The first wave of the generative AI revolution was characterized by horizontal LxMs and applications. ChatGPT, Gemini, Midjourney hit the market as out-of-the-box, general-purpose tools that address a wide range of situations and enable various content generation use cases. Yet, their major limitation lies in the datasets employed to train the foundation models. Since LxMs (mostly) rely on generic, publicly available information, they lack the in-depth domain knowledge that is necessary for highly specialized business applications. In fact, when employed in industry-specific use cases, general-purpose models run the risk of producing subpar predictions or altogether hallucinating, thus reducing the reliability of the entire AI system.
This is why companies are shifting towards vertical AI applications. These tools leverage high-quality training data and dedicated models to ensure relevant responses. They improve the efficiency by combining vertical knowledge and expertise, rarely found in horizontal applications, with domain-specific features and business logics. We are now witnessing the rise of AI applications for law firms, healthcare, marketing, finance, insurance – and the market is reacting positively. According to AIM Research, the vertical AI market was valued at $5,1 billion in 2024 and is projected to reach $47,1 billion by 2030.
The key to creating a vertical AI application or agent lies in the foundation model. There are two basic approaches: first, it is possible to train a vertical LxM from scratch with relevant datasets. This option offers personalization and flexibility, enabling in-depth domain knowledge. It also ensures regulatory compliance, avoiding the utilization of sensible or copyrighted data. However, training a vertical model might be too daunting a task for companies that are first venturing into the vertical AI world. Fine-tuning an existing model on industry-specific data may be a preferable approach to reduce development costs and time.
3. AI Observability
As AI-based applications become more numerous, sophisticated and central to companies’ core business, it becomes fundamental to collect real-time insight about their performance. There are two main reasons: first of all, Observability ensures that the AI models are generating value for the users and thus for the business. Online banks cannot afford downtime or false positives when the money of their clients is at stake.
Secondly, Observability is instrumental in assuring a responsible use of AI. Companies are increasingly asked to meet stakeholders’ and regulators’ expectations of fair and ethical AI, and ensure the reliability of AI-driven decisions. This is only possible with traceability, quality assurance and in-depth knowledge of model behavior.
Open Source tools such as the Radicalbit AI Monitoring platform are a free and flexible solution to gain situational awareness. It helps data teams guarantee model accountability and effectiveness in production, with comprehensive insights into Data Quality, Model Quality and Model Drift. Radicalbit AI Monitoring offers out-of-the-box support for the most common model tasks such as binary classification, multiclass and regression. To learn more and take part in the project, visit our GitHub page!
4. Energy Efficiency
As Generative AI develops and becomes mainstream, energy consumption and environmental impact are skyrocketing. Efficiency comes with a price: as the World Economic Forum website points out, Gen AI systems are likely to use 33 times more energy to perform a task than non-AI specific software. The sheer computational power needed to train and operate the most sophisticated models is the main culprit, prompting the need for new data centers that produce emissions and strain the electrical grid to the limit. Suffice it to say that Microsoft’s CO2 emissions rose nearly 30% in 2023 due to infrastructural expansion.
This is why power costs and environmental impact have become a central issue for AI. We cannot longer see AI development as simple technological advancement, but we must take into account the other side of the equation and incorporate sustainability principles into AI initiatives. This could mean preferring smaller, vertical LLMs that require less energy to train and operate, rather than power-hungry all-purpose models; creating and employing scalable AI applications that reduce resource consumption based on workload requirement; investing in energy-efficient hardware with better performance per watt and effective cooling systems.
5. Shadow AI
Shadow AI occurs when employees use AI-based applications without the oversight of the company’s IT department. The adoption of unsupervised AI solutions potentially jeopardizes security and compliance, especially in data protection. According to recent figures, in 2024 27% of corporate data entered into AI tools by employees was sensitive. The issue is particularly difficult to tackle because it often goes undetected – 74% of ChatGPT accounts used in the workplace are personal, thus falling outside the scope of corporate control systems.
To tackle these security and compliance vulnerabilities, companies have to take a two pronged approach. On the one hand, a cultural change is needed to make employees aware about the risks of unsanctioned and unsupervised AI. It is important to lay down clear guidelines on the use of AI-based applications, and to explicitly address these use cases within the data protection regulation.
On the other hand, companies must adopt software solutions that help ensure the compliance of AI-based software. Guardrails and LLM Evaluation are fundamental ways to guide the behavior of generative AI applications. Guardrail techniques include prompt engineering and output filtering that prevent the generation of harmful content or data privacy violations. LLM Evaluation assesses the quality and correctness of the responses generated by LLM-based applications, also employing other LLMs as “judges”.
As we step into 2025, the AI landscape continues to evolve at lightning speed, bringing both opportunities and challenges for businesses worldwide. From the rise of autonomous AI agents to the growing focus on vertical solutions, companies must stay ahead of these trends while ensuring responsible and sustainable AI deployment.
Radicalbit stands at the forefront of this evolution with its comprehensive LLMOps and AI Observability platform. Our solution empowers data teams to tackle critical challenges in AI deployment, observability and explainability, enabling businesses to harness the full potential of AI while maintaining the highest standards of reliability and responsibility.
