Imagine a world where AI can easily tackle your trickiest tasks. In today’s economy, businesses need solutions that maximise efficiency and ROI. While Large Language Models (LLMs) show promise, they have limitations. Agentic AI offers a solution.

By dividing AI applications into specialised agents, Agentic AI overcomes these hurdles and unlocks new possibilities. But how can you prepare for an agentic economy?

Understanding Agentic AI: Autonomous Action

Agentic AI is an architecture paradigm where a single AI application is divided into specialised AI agents. Each agent has its own role, a specific set of tools, and potentially even its own large language model (LLM) and knowledge base. This method allows complex problems to be broken down into smaller, more manageable tasks, similar to a “divide and conquer” strategy.

Agents are often called autonomous because they can act and take initiatives to fulfil their tasks without user support. In some cases, they can even learn and enrich their long-term memory.

In short, Agentic AI focuses on creating AI systems that are capable of independent action and goal-oriented behaviour. These agents can automate tasks, manage systems, and provide highly personalised experiences.

How Agentic AI Works

Instead of a single, monolithic AI, Agentic AI breaks down complex tasks into smaller, more manageable pieces, with each agent having a specific role and toolset, which allows the agent to perform real actions rather than just a conversation. Each agent works on a limited task, making it easier to design, implement, and test. Using the right set of tools and the right sized LLM for a specific challenge is also more efficient and sustainable than always using a powerful LLM.

The agents can be orchestrated in many possible ways:
Workflows, when the activities to execute to complete a given task can be modelled as a business process. This facilitates the control at each step but is more rigid.
Free flow with supervision by a manager agent, where each agent tries to accomplish its task by interacting with the other agents, and the manager agent controls and validates the results, and may ask for improvements.
Free-flowing, unsupervised teams in which all agents have their personas and freely interact according to their own activities and personas.
Any mix of the above. It is also possible to involve a human when validation is required or additional information is necessary.

Looking forward, we can anticipate different kinds of Agents working together in a typical information system to support human users:

AI Agents in augmented Software (Gartner calls them “embedded AI”), referring to those agents present in commercial SaaS such as Salesforce or Sap ;
Personal AI Agents, such as device-based AI agents like Apple intelligence or Copilot+ PC (close to what Gartner calls “Bring Your Own AI”)
AI Agents used to develop custom Business functions (Gartner calls them “Built AI”) that could potentially “take over” all the backend business rules of traditional applications.

We also anticipate that AI Agents will strongly influence the User Interface of business applications, with more context-rich interaction models, potentially including voice and image or video for easier user request formulation. Personal Agents could serve as a universal proxy UI to activate the relevant agents based on user request.

Since these different types of Agents come from different vendors, integrating them might be a challenge. Natural language can be a minimal platform for integration, but critical processes will require stronger formalism for more trust. Anthropic has released a candidate standard, “Model Context Protocol,” to facilitate the integration of AI Agents with their environment (API, tools, data) or other Agents. Since this is an area for competition, there is no guarantee that a global standard will emerge soon.

Another aspect to consider in the context of agentic AI: we sometimes talk about Collective AI to describe how the interaction between AI agents and human agents can lead to a level of intelligence greater than the sum of its parts. It’s about finding the optimal balance between humans and AI, considering what each is best at to complete a task or process. This approach also focuses on setting up safeguards to ensure that humans remain in control of important decisions, preventing a potential “takeover” by AI.

Applications of Agentic AI

Don’t just take our word for it. Agentic AI is already being deployed in real-world scenarios with impressive results. Let’s explore some compelling use cases:

- Revolutionising Customer Service: Imagine a customer service chatbot that can seamlessly handle complex inquiries. With Agentic AI, this is a reality. A “User Proxy Agent” acts as the initial point of contact, analysing customer requests and intelligently routing them to specialised agents like “Product Experts,” “Billing Specialists,” or even “Upsell Agents.” This ensures customers receive fast, personalised support while “Quality Control” and “Compliance Agents” work behind the scenes to maintain service standards and adhere to regulations.
- Supercharging Sales: By automating tasks like lead qualification, proposal generation, and contract negotiation, Agentic AI empowers sales teams to focus on building relationships and closing deals.
- Streamlining Operations: From supply chain management to human resources, Agentic AI can optimise critical business processes. This translates to reduced costs, faster turnaround times, and improved productivity.
- Intelligent Document Processing, for example sales assistant for IT services company:
  - User proxy Agent: collects the requirements document from client and additional details such as location and start date, collects existing project context from knowledge management application, drives the workflow to other agents and ensures overall success of requests
  - Skills Identification Agent: identifies skills required from customer document
  - Staffing Agent: identifies best available consultants with the skills
  - Proposal Agent: prepares a templatised proposal to send to customer, optimises content based on the knowledge base of the IT services company and based on the publicly available information about client strategy
  - Reviewer Agent: challenges the proposal and suggests improvements
  - Compliance Agent: ensures that regulation is enforced

Devoteam has already delivered agentic AI applications in production, such as the employee chatbot of Strawberry based on LangGraph and VertexAI and experimented with advanced unsupervised agents teams with CrewAI and Autogen. This shows that these concepts are not just theoretical but are being applied to deliver real solutions.

Other technologies for Agentic AI include:

Langchain LangGraph
Autogen
AWS Bedrock
Microsoft Copilot Agents
Google Cloud Vertex AI
Service Now AI Agents
(Llama Index, Swarm AI, …)

Environmental and Technical Challenges

While the potential of Agentic AI is undeniable, it’s crucial to acknowledge the challenges that come with this emerging technology.

The Black Box Problem: As AI agents interact increasingly, understanding how they arrive at specific outcomes can be challenging. This “black box” effect raises concerns about accountability and bias. To address this, researchers are actively developing techniques to improve the transparency and interpretability of Agentic AI systems.
The Sustainability Puzzle: Agentic AI’s reliance on extensive communication between agents can lead to a significant carbon footprint. However, solutions are emerging.
The ROI Equation: Implementing Agentic AI requires a significant investment. Businesses must carefully assess the potential return on investment, considering increased efficiency, reduced costs, and improved customer satisfaction.
Ensuring Reliability: Unpredictable behaviour and silent failures can undermine trust in Agentic AI systems. Robust quality assurance frameworks, including rigorous testing and monitoring, are essential to mitigate these risks.

Challenges in Integration into IT Infrastructure

Integrating these technologies into existing IT infrastructure and workflows can be challenging. Key challenges include cybersecurity and GPU infrastructure optimisation.

The Future of Agentic AI

Evolving Roles of Human Consultants

Agentic AI doesn’t replace human consultants, it empowers them. While AI excels in automation and data processing, humans remain essential for tasks requiring:

Contextual Understanding: Humans can interpret nuances and complexities that AI may miss.
Building Trust: Human connection is crucial for establishing and maintaining strong client relationships.
Driving Sales and Customer Satisfaction: Humans will continue to lead activities that require empathy, persuasion, and emotional intelligence.

Agentic AI allows human consultants to focus on these high-value activities by automating routine tasks and providing data-driven insights.

Essential Skills

IT employees will need new skills to effectively collaborate with and manage AI agents. These will include understanding AI capabilities, developing and monitoring AI systems, and addressing ethical considerations.

Promising Future Trends

Future trends include the development of agent marketplaces, where different agents can compete and provide services.

Personal AI agents may become the main entry points for any interaction, connecting with other dedicated agents for different tasks.

If the interoperability question is solved, this could lead to an agentic economy with microservices being paid between agents.

Conclusion

Agentic AI is no longer a futuristic fantasy; it’s a tangible reality reshaping the technological landscape. From enhancing customer experiences to optimising business processes, this paradigm shift is showing a new era of intelligent, adaptable, and efficient systems.

But the journey is not without its challenges. As we embrace the potential of Agentic AI, we must also address concerns surrounding explainability, sustainability, and ethical considerations. By fostering collaboration between researchers, developers, and businesses, we can unlock the full potential of Agentic AI while ensuring its responsible and beneficial integration into our world.

Put forward by Olivier Mallet, Head of AI Agency at Devoteam

Managed Platforms & Sustainable IT