Generative artificial intelligence (AI) was undeniably a prominent topic last year, drawing significant public and corporate attention. Contributions from AI scientists were awarded Nobel Prizes in physics and chemistry, underlining the importance of automation.[i] Analysts and economists diligently assess AI’s potential impact, while practitioners cautiously explore its adoption, mindful of the associated risks. In this article, we examine recent trends and share our projections concerning the advent of agentic AI.
The AI boom
The AI industry has enjoyed sustained growth, beginning with deep learning advancements in the 2010s, followed by notable image and video processing achievements. Generative AI’s rise in the 2020s was driven by the development of transformer architectures in 2017, culminating in the release of generative pre-trained transformer (GPT) models and the creation of a dynamic AI development ecosystem.
Since 2010, improvements in cloud computing and computational power, combined with decreasing costs, have driven progress in deep learning and AI. Today’s efficiency is astonishing; by 2020, models from 2012 required 44 times less computing power, thanks to algorithm and hardware advances.[ii] [iii] The computational requirements for AI models, which used to double approximately every 21.3 months before 2010, have accelerated to doubling every 5.7 months during the deep learning era.[iv]
Since 2020, large language models (LLMs) have demonstrated significant and disruptive capabilities, transitioning from academic research to becoming a focal point for major technology firms. Over the past twelve months, leading tech companies and innovative startups have launched increasingly sophisticated models with remarkable frequency, naturally capturing public attention.
The release of GPT models significantly boosted interest in AI, and usage became increasingly popular with generic LLMs handling public domain information and conversational AI features. This allowed users to ask for general advice, such as creating a vacation plan. By 2024, many companies experimented with fine-tuning techniques and retrieval-augmented generation (RAG), enabling AI tools to provide specific policy references and step-by-step procedures. This allowed users to ask how to request vacation days within a particular team and receive detailed guidance tailored to company-specific policies.
With ongoing advancements in multimodal and visual explorations, and AI development techniques like RAG and fine-tuning, the initial public enthusiasm for LLMs has begun to wane. In response, a new area of exploration has emerged: agentic AI.
Looking ahead to 2025, with the right integrations and tools, AI agents may not only help build a vacation plan and provide tips on requesting time off but also check flight availability, book hotels and handle the entire vacation request approval process.
“Future agents will not only be more accurate, sophisticated, and accessible but may also have their own marketplaces, registries, and information exchange protocols.”
Liubomyr Bregman, Senior Manager – Artificial Intelligence & Data, and Nicolas Griedlich, Partner – Artificial Intelligence & Data, Global ESG FSI Tech Leader, at Deloitte
What is agentic AI?
Agentic is an AI system that uses complex reasoning and planning to solve problems and achieve goals without human supervision.[v] The idea of agents has existed almost as long as the idea of AI and early software products. However, recent advancements in algorithms, falling infrastructure costs, and efficiency gains in LLMs have made agentic AI feasible for routine business applications.
AI agents have gained traction, grounded in the “intelligent agent” concept. This involves external communication with a system until specific tasks are satisfactorily completed, originating from economic theories of rational agents, and further developed by Google’s research on chain-of-thought prompting.[vi] Combining these principles has yielded innovative solutions like multiple agentic AI frameworks, capable of automating not only minor tasks but also achieving a satisfactory level of task completion.
In the realm of business, AI agents mirror human tasking. Both must be meticulously selected, thoroughly trained, and equipped with the right tools to function optimally. Strategic deployment and steady management are vital for maintaining efficiency and enhancing value. By integrating AI agents endowed with human-like cognitive abilities, businesses can boost productivity, streamline operations, and achieve heightened success.[vii]
For example, in March 2025, during a private industry meeting in Deloitte AI Leaders Hub, we demonstrated a shift from simple agents to more complex multi-agent systems, where the same task may be executed by a few agents: a team lead agent, designer agent, analyst agent, and software engineer agent. This mirrors a full process where one agent builds a plan and distributes tasks, and multiple agents collaborate to complete them efficiently and faster than ever before.
In our experiment, within an hour from scratch, multi-agent systems could build a service capable of planning travel, booking a flight and hotel, integrating with a calendar to plan meetings, and collecting contacts.
Deloitte has used agents for data collection as well as document and report generation across industries like finance, banking, the public sector, and logistics.
Future expectations
Given agentic AI solutions’ rapid evolution from exploratory innovation to production deployment, further examination of security, scalability and performance risks are required in 2025. Nonetheless, we foresee substantial advancements in automation and innovation, driven by the development of larger and more sophisticated agents, as well as the formation of agent ecosystems collaborating across different computational bases. Some experts even speculate about creating a comprehensive new universe, referred to as Agentic Mesh.[viii]
Future agents will not only be more accurate, sophisticated, and accessible but may also have their own marketplaces, registries, and information exchange protocols. At Deloitte Luxembourg, in collaboration with key technology partners, we proactively connect business cases with clients, explore practical applications, and focus on developing impactful solutions. We remain dedicated to ensuring these advancements translate into tangible benefits for our clients, addressing specific business needs with precision and efficiency.
[i] Ben Li and Stephen Gilbert, “Artificial Intelligence awarded two Nobel Prizes for innovations that will shape the future of medicine,” Nature, 25 November 2024.
[ii] Vipra, J., & West, S. M. Computational Power and AI*. AI Now Institute. 27 Sep 2023
[iii] Danny Hernandez and Tom B. Brown, “Measuring the Algorithmic Efficiency of Neural Networks,” arXiv, 8 May 2020.
[iv] Jaime Sevilla et al., “Compute Trends across Three Eras of Machine Learning,” arXiv, 9 March 2022.
[vi] Jason Wei et al., “Chain-of-Thought Prompting Elicits Reasoning in Large Language Models,” arXiv, January 2022.
[vii] Ronan Vander Elst, Nicolas Griedlich, Anke Joubert and Liubomyr Bregman, “Agentic AI: The new frontier in AI evolution,” Deloitte Future of Advice, 1 April 2025.
[viii] Eric Broda, “Agentic Mesh: The Future of Generative AI-Enabled Autonomous Agent Ecosystems,” Towards Data Science, November 2024.
