Tech Navigator: Agentic AI Systems

Insights

• AI agents, unlike traditional automation, employ adaptive, autonomous decision-making to perform tasks with minimal human intervention, bridging the gap between rigid automation and human-like reasoning.
• AI has evolved to surpass human cognition in various fields, yet efforts continue to make the technology more intuitive and adaptable — much like humans.
• By 2025, a significant portion of companies using generative AI will begin exploring agentic AI pilots, with widespread adoption expected by 2027.
• The emergence of large language models has revolutionized the agentic world, serving as increasingly sophisticated intelligence engines powering autonomous agents.
• There are currently promising uses for agentic AI, however, as an emerging technology, not all have been tested at enterprise scale yet.

Artificial intelligence (AI) has steadily surpassed human cognition in fields once thought to be uniquely ours — from image recognition and speech processing to algorithm design. Yet, despite its astonishing power, computer scientists are still trying to make AI behave more like humans: intuitive, adaptable, and independent.

People are naturally skilled at recognizing patterns and making sense of chaos, even when information is disorganized or incomplete. However, comprehension doesn’t always strike immediately: We rely on books, online searches, and the wisdom of others to make more informed decisions that lead to better results. Generative AI follows a similar trajectory by retrieving information, generating insights, and sometimes taking actions — whether that’s analyzing a customer’s purchase history to recommend tailored products or automating essential tasks, such as sending emails and processing transactions.

The expectation that AI will deliver highly personalized experiences stems from the level of tailored care humans have come to expect across the services they consume. Businesses have refined their hyperpersonalization efforts, leveraging streamlined communication and interaction history to enhance customer engagement. Basic and routine inquiries were efficiently managed by chatbots that respond quickly and reduce human intervention.

However, while early chatbot implementations served their purpose, their lack of context awareness and empathy soon became clear, exposing a gap in user experience. The shift from static acknowledgments to dynamic, adaptive responses — capable of inheriting context from the industry domain and acting on insights — represents a significant opportunity for AI advancement. Developing systems that can understand, adapt, and respond intelligently in real time is the next frontier in AI-driven personalization.

When AI gains the ability to reason, coordinate tasks, and act with purpose, it transcends being a mere tool and becomes an agent. And as companies seek more adaptable, self-sufficient solutions, these emerging AI agents will evolve from technology tools into indispensable business partners.

The evolution of AI agents

Agentic AI’s arrival is accelerating — reminiscent of the rapid rise of generative AI just over two years ago. Gartner considers agentic AI to be the top technology trend for 2025. And Deloitte forecasts that by 2025, one-quarter of companies that use generative AI will initiate agentic AI pilots or proof-of-concept projects, with adoption increasing to half by 2027. The consulting firm also projects that in certain industries and use cases, agentic AI applications could begin integrating into existing workflows in late 2025.

At their core, agents are autonomous software entities that use a simple yet potent operational loop: They observe their environment via sensors, process this input, and use either mechanical or digital actuators to change their environments and achieve specific objectives. This traditional sense-plan-act cycle remains as relevant in the AI era as when it was initially conceptualized for robotics.

As agents evolve through each stage of enterprise expansion and adoption, their capabilities progressively advance. Traditional robotic process automation relies on rules-based or instruction-driven configurations, operating strictly within predefined parameters. These systems respond only to programmed rules, lacking adaptability or contextual reasoning.

In contrast, agentic AI systems introduce adaptive, autonomous decision-making, bridging the gap between rigid automation and human-like reasoning. With minimal to no human intervention, AI agents perform specific tasks using capabilities that are integrated across multiple layers in the overall agentic system.

Now, the rise of large language models (LLMs) has transformed the agentic world by acting as increasingly sophisticated brains behind the agents. Agents and LLMs function as partners with agents orchestrating LLMs as needed. However, the agent retains control over when to call and execute an LLM, ensuring structured, task-driven interactions that align with the system’s objectives and security requirements. From simple automation to the potential for artificial general intelligence (AGI), AI has steadily progressed from a world of rules-based actions to greater independence (Figure 1).

Figure 1. The progress of AI agents

Source: Infosys

Tier 0 – Simple automation: Localized automation applies to a specific segment of a broader process. It lacks agentic behavior and operates through rules-based, deterministic automation. A common example is component test automation, where scripts execute predefined tasks.

Tier 1 – Robotic process automation: An advancement over basic automation, this approach extends beyond point solutions to cover entire process segments using predetermined rules-based logic. While AI or agentic behavior remains minimal to nonexistent, it enhances efficiency in tasks such as screen scraping or automated form completion.

Tier 2 – Elementary AI augmentation: This stage introduces the first meaningful presence of agentic AI, offering an opportunity to replace tier 1 automation, particularly in areas requiring human oversight. By leveraging a language model, agents enable intelligent interactions while maintaining a limited, yet impactful, role in automation. Examples include sentiment analysis or ticket data labelling, where agents classify information into appropriate categories.

Tier 3 – AI agent twins: These systems function as digital twins to users, interpreting intent and autonomously taking action to achieve specific outcomes. Some of the best-known instances of AI agent twins include GitHub Copilot and Microsoft 365 Copilot, which assist users by generating code, automating tasks, and enhancing productivity through intelligent decision-making.

Tier 4 – Refined AI agents with reflection: These systems represent an advanced class of AI agents that many organizations are eager to implement. They can decompose tasks from a given objective, formulate plans to achieve the intended outcome, and analyze results to adapt their approach in response to failures or unexpected events through complex reasoning sequences. This tier can be seen in credit decision-making systems, where agents process loan applications, extract and analyze documents, and match them against stored information to ensure accuracy and compliance.

Tier 5 – Complete autonomy, AGI or artificial superintelligence: When this tier arrives in the future, these agents will possess the capability to conduct entirely original research, independently reason through complex problems, and develop innovative solutions beyond their initial training data. With advanced logical reasoning and adaptive learning, they will continuously acquire new skills, refine their methodologies, and tackle previously unsolved challenges, pushing the boundaries of AI-driven discovery and problem-solving.

Although tier 5 is still somewhere in the future — with optimistic predictions ranging from 2026 to 2029 — AI has already shifted possibilities and expectations. This technology has evolved from rules-based automation to intelligent, self-improving systems capable of operating independently, dynamically responding to their environments, and optimizing decision-making in real time. Organizations can now deploy these agentic solutions across domains — such as software development, IT operations, and customer care — to drive unprecedented efficiency and adaptability.

The blueprint for agentic AI

Business leaders recognize the value AI has already delivered and see its potential to drive further transformation. However, fully capitalizing on AI-driven opportunities requires a strong foundation in both architecture and operational layers of advanced AI systems. The blueprint for agentic AI defines the core components and processes that drive their functionality. By analyzing this framework, it becomes clear how user inputs, orchestration engines, planning modules, reasoning layers, utilities, memory, integration interfaces, human oversight, and pretrained language models work together to enable intelligent and autonomous agent behavior.

The following are the core layers of an agentic system and an examination of how it operates (Figure 2).

• User input or event trigger: A predefined instruction set or event combination that activates an agent, prompting it to initiate actions or processes based on specified conditions.
• Goal-based orchestration engine: Serve as the foundation for decision-making and prioritization, orchestrating various actions to align with a system's goals.
• Planning module:
  • Reasoning layer: Algorithms designed to analyze inputs, develop strategies, and break tasks into subtasks to ensure efficient execution.
  • Tools and skills: Reusable software modules activated to achieve specific goals, such as document digitization, optical character recognition, and PDF generation.
  • Memory: Module maintains a record of current and past interactions, tool usage, and learned behavior, enabling context-aware planning.
  • Integration module: Module interfaces with the outside world, often through APIs.
  • Human in the loop: Humans approve certain critical decisions recommended by agents.
• Language model: Pretrained or fine-tuned models deployed responsibly on public or private cloud infrastructure.
• Security and compliance: Ensure strict compliance with data security, privacy, and ethical standards, aligning with responsible AI principles and regulatory requirements.
• Evaluation layer: Measures effectiveness and efficiency of the agent and provides feedback to the planning module for continuous improvement.

Figure 2. AI agent framework

Source: Infosys

Planning module

The planning module, often considered to be the agent's “brain,” is the central orchestrator of decision-making, task prioritization, and adaptive execution in an agentic system. It plays a pivotal role in transforming reactive automation into proactive, intelligent decision-making.

Functioning alongside the orchestration layer, the planning module interprets input triggers and breaks down objectives into achievable goals. The orchestration layer refines these objectives into a structured sequence of subtasks, oversees execution, engages reflection modules for validation, and generates user responses.

Together, these layers form a continuous cycle that guides agent behavior through the plan, act, and reflect stages. This process involves gathering data from internal and external sources, analyzing it using predefined logic, reasoning frameworks, or learned patterns, and determining the optimal next step based on the current state and objectives. By integrating these components, agentic AI systems achieve a high degree of autonomy, adaptability, and operational intelligence.

The layers that enable planning at scale include the following:

Reasoning layer
This capability shapes interactions with the language model by refining prompts and evaluating or ranking responses. It plays a crucial role in identifying specific milestones within subtasks, processing environmental inputs, and capturing sensory data. Additionally, it leverages knowledge bases and reference frameworks, such as knowledge graphs, to narrow down results and enhance accuracy.

Memory module
The ability to retain context from both current and past interactions, while continuously learning from ongoing and long-term activities, provides the reasoning layer with the necessary background and feedback for effective operations.

Tool, skills, and integration layer
This layer compensates for the inherent limitations of language models in directly interacting with the real world. It achieves this by leveraging reusable software components, such as PDF generation and document digitization, to accomplish specific goals. Additionally, it enables integration with specialized systems by accessing external web APIs, including retrieval-augmented generation (RAG) frameworks.

Language model

The language model serves as the central decision-maker, consisting of one or more models that employ reasoning frameworks such as ReAct or chain-of-thought. These models can be general, multimodal, or fine-tuned for specific business objectives. For optimal performance, the chosen model should align with target requirements and be trained on data relevant to the integrated tools. While the model is typically not trained on the agent’s specific configuration, its decision-making accuracy can be enhanced by providing contextual examples that highlight the agent’s capabilities. This approach ensures more precise and context-aware outputs.

The planning and reasoning cycle continues iteratively until the goal is achieved, or a stopping condition is met. The complexity of orchestration depends on the agent and task, varying from simple calculations to advanced logic, such as chained reasoning or machine learning algorithms.

Why do we need AI agents?

The rise of agentic AI is fueled by the growing complexity of modern businesses and the demand for intelligent automation. Agentic AI overcomes the limitations of traditional automation by integrating adaptability, contextual awareness, and autonomous decision-making.

Unlike conventional systems, which struggle with unstructured challenges and evolving conditions, agentic AI processes ambiguous data through continuous learning and real-time analysis to optimize workflows. This shift from rules-based automation to context-driven intelligence allows organizations to service increasingly complex demands. The following features make agentic AI particularly relevant for the modern enterprise.

• Operational efficiency: Integrates cross-functional tools, skills, and APIs to detect inefficiencies and autonomously implement improvements, ensuring process agility.
• Ability to scale rapidly: Unlike resource intensive traditional systems, it enhances capabilities without requiring proportional staffing increases by taking advantage of self-optimizing architectures.
• Dynamic problem-solving: The ability to manage unstructured tasks through contextual reasoning allows agents to generate intelligent responses to complex requests without relying on predefined rules.

When is agentic AI the best solution?

Frameworks play a crucial role in the effective deployment and use of agentic AI, ensuring that agents operate efficiently, ethically, and in alignment with business objectives. And with the limited number of real-world case studies, these frameworks serve as blueprints that can guide organizations that want to move quickly and decisively — rather than waiting for competitors to pave the way forward.

At Infosys, we use the SCOPE framework to assess the suitability of agentic AI for addressing business challenges. Each parameter is scored based on its impact.

Strategic alignment

• Does the opportunity align with long-term organizational goals?
• Are the workflows today largely manual?

Complexity of task

• Is the task planning intensive and multistep?
• Is the planning largely deterministic, and if so, are the rules expected to increase exponentially to comprehensively cover the possibilities?

Operational environment

• Is there a need for real-time action or decision-making, or a need to adapt to changing circumstances?
• Are there multiple data streams to integrate with and will the need grow to integrate with more systems?

Performance requirements

• Are there latency constraints or impacts that affect the cost?
• What are the ethical and security constraints
• Can the system’s actions be adequately governed, controlled, and monitored?
• Are there ethical risks with autonomous decision-making or recommendations?

In addition to insights provided by a robust framework, enterprises should closely examine the growing number of agentic AI use cases that are applicable to complex, real-world scenarios. Below are the most promising uses for agentic AI, although not all have been tested at enterprise scale yet. Remember that agentic AI is an emerging technology that is still building a track record.

Figure 3. Agentic AI use cases by industry

Source: Infosys

While AI adoption is accelerating, its enterprisewide deployment remains limited, with many organizations still in the experimental phase. Relatively few have successfully implemented and scaled agentic AI across specific operations. Below are examples of organizations that have made significant progress in integrating agentic AI into their workflows.

CommBank leverages agentic AI to process about 15,000 payment disputes daily. Customers can describe their issues through an AI-assisted channel, which autonomously verifies eligibility criteria and lodges disputes without requiring manual intervention. This automation enhances efficiency, reduces processing time, and improves customer satisfaction.

Telecom company Telenor has deployed conversational agentic AI agents to autonomously handle customer queries, resolve issues, and facilitate sales. This implementation led to a 20% increase in customer satisfaction and up to a 15% growth in revenue within the first year.

Talkdesk introduced agentic AI-powered conversational agents for retail customer service, enabling autonomous management of complex tasks such as order updates, address modifications, and customer routing to in-store specialists. These AI agents provide 24/7 support while delivering hyperpersonalized experiences, enhancing both efficiency and customer engagement.

Levi Strauss has implemented agentic AI for granular demand forecasting across its supply chain. The system autonomously adjusts inventory levels based on real-time demand signals, ensuring that the right products are available in the right locations while minimizing waste and inefficiency.

Redefining AI

The emergence and maturation of agentic AI marks a defining moment in the evolution of intelligence — one that transcends rigid algorithms and human-dependent decision-making. Unlike its predecessors, which followed predefined rules, agentic AI is self-directed, adaptive, and deeply contextual, capable of understanding real-time data, reasoning dynamically, and acting with intent.

This shift is more than an incremental improvement; it redefines how technology engages with the world. No longer bound by static programming, AI is evolving into a thinking entity, a collaborator capable of navigating complexity, optimizing processes, and unlocking vast new possibilities.

Ultimately, the rise of agentic AI represents a transformative leap in the field of AI. By enabling machines to think, learn, and act independently, we are not only enhancing the capabilities of AI but also redefining the relationship between humans and technology. As businesses continue to explore and implement these advanced systems, the future of AI promises to be more dynamic, intelligent, and impactful than ever before.